What Is a Hallucination in AI?

What Is a Hallucination in AI? Why AI Lies and What You Can Do About It

Table of Contents

1. What an AI Hallucination Actually Is
2. Why AI Hallucinates — The Real Explanation
3. The Confidence Paradox: AI Is Most Certain When It Is Most Wrong
4. How Common Is Hallucination in 2026?
5. Which AI Models Hallucinate Most and Least
6. The Real-World Consequences
7. How to Protect Yourself
8. Will AI Ever Stop Hallucinating?
9. Frequently Asked Questions

You asked an AI for a statistic. It gave you one, complete with a source. You looked it up. The statistic does not exist. The source does not exist. The AI made both up, confidently and fluently, in a tone that suggested it had checked. This is AI hallucination — and in 2026 it remains one of the most important things anyone using AI tools should understand. The best models have improved dramatically: Gemini 2.0 Flash now hallucinates just 0.7% of the time on structured tasks. But the global financial cost of AI hallucinations still reached $67.4 billion in 2024. And a critical piece of research found that AI models are 34% more likely to use confident language — "definitely," "certainly," "without doubt" — when generating incorrect information than when they are right. The problem is real, it is structural, and it is not going away entirely. Here is what you need to know.

What an AI Hallucination Actually Is

An AI hallucination is when an AI system generates information that is factually wrong, unverifiable, or entirely fabricated — presented with the same fluency and confidence as accurate information. The term is borrowed from psychology, where hallucination describes perceiving something that does not exist. In AI, it describes producing something that does not exist: a citation that was never published, a statistic that was never measured, a quote that was never said, an event that never happened.

The word "hallucination" can be slightly misleading because it implies the AI is experiencing something. It is not. A better mechanical description is "confabulation" — the same word used in neuroscience for when brain-damaged patients unconsciously generate false memories without any intention to deceive. The AI is not lying in any meaningful sense. It genuinely has no mechanism to know the difference between what it fabricated and what it knows. It is generating the most plausible-sounding continuation of a conversation, and sometimes that continuation is fiction.

The two main types of hallucination: Researchers distinguish between intrinsic hallucinations — where the AI contradicts information it was given — and extrinsic hallucinations — where the AI generates information that cannot be verified against any known source, inventing facts, citations, statistics, or events from scratch. Extrinsic hallucinations are the more dangerous category for most users because they are harder to detect: the information sounds plausible, fits the context, and nothing in the response signals that it was invented.

Why AI Hallucinates — The Real Explanation

To understand why AI hallucinates, you need to understand what large language models actually are — because they are fundamentally different from what most people assume.

An LLM is not a database. It does not look things up. It has no index of facts it can retrieve and verify. It is a prediction engine — trained on billions of words of text to predict the most statistically probable next word given everything that came before it. This is what makes AI writing feel so natural and fluent. It has learned, from an enormous amount of human-written text, what words typically follow other words in what kinds of contexts. When it answers a question, it is not retrieving the answer — it is generating the most plausible-sounding answer based on patterns it learned during training.

The structural reason hallucination is inevitable: When an AI model encounters a question it does not have reliable information about, it faces a choice: admit uncertainty or generate a plausible-sounding answer. OpenAI published research in 2026 explaining this directly: hallucinations persist because standard training and evaluation procedures reward guessing over acknowledging uncertainty. When models are trained and evaluated on accuracy metrics, guessing and occasionally being right looks better statistically than frequently saying "I don't know." The training process inadvertently optimises for confident-sounding answers over honest admissions of ignorance.

The root causes of hallucination cluster into four categories. First, incomplete or flawed training data — if the model was never trained on information about something, it has nothing to draw on except patterns from vaguely related contexts. Second, the probabilistic nature of text generation — the model is always generating statistically likely continuations, and a statistically likely continuation is not always a factually accurate one. Third, the absence of a ground truth check — the model has no mechanism to verify its outputs against reality before producing them. Fourth, training incentives that reward fluency and confidence over uncertainty — a model that frequently says "I don't know" looks less capable in evaluations than one that attempts every answer, even imperfectly.

The Confidence Paradox: AI Is Most Certain When It Is Most Wrong

This is the most counterintuitive and most important thing to understand about AI hallucinations — and the research is unambiguous on it.

MIT research finding (January 2025): When AI models hallucinate, they tend to use more confident language than when providing factual information. Models were 34% more likely to use phrases like "definitely," "certainly," and "without doubt" when generating incorrect information. This is the core paradox: the more wrong the AI is, the more certain it sounds. The same fluency that makes AI feel authoritative is the fluency it applies equally to facts and fabrications. There is no hesitation in the voice, no caveat in the phrasing, no signal that the confidence is unearned. The hallucinated citation reads exactly like the genuine one.

This paradox has serious practical implications. Scepticism is most warranted precisely when AI sounds most authoritative. If an AI response includes precise statistics, specific citations, exact quotes from named experts, or highly specific details — these are the moments to verify, not to trust most. Vague generalities are often more reliable than specific details, because the model is more likely to be confabulating when it is being precise about something it does not actually know precisely.

How Common Is Hallucination in 2026?

The honest answer is: it depends enormously on the task, the model, and how you measure it. The range of figures in the research reflects genuine variation, not methodological confusion.

Tasks where hallucination rates are now very low

• Retrieval-augmented generation (RAG) — When AI is given a document and asked to summarise or answer questions about it, top models have reached below 2% hallucination rates. Grounding AI in a provided source dramatically reduces invention.
• Structured data extraction — Asking AI to extract specific information from provided text in a defined format produces much lower error rates than asking it to generate information from memory.
• Simple factual questions about well-documented topics — For widely-covered facts with abundant training data, top models hallucinate less than 1% of the time.
• Code generation (syntax) — AI-generated code has lower hallucination rates for standard syntax and patterns than for specific library versions or obscure functions.

Tasks where hallucination rates remain very high

• Legal queries — Stanford research found LLMs hallucinate between 69% and 88% of the time on specific legal queries. This is not a minor concern for anyone using AI for legal research.
• Medical case summaries — Without specific mitigation prompts, hallucination rates in medical case summaries reached 64.1% in 2026 research.
• Person-specific questions — OpenAI's o3 and o4-mini reached 33% and 48% hallucination rates respectively on person-specific questions. The more obscure the person, the higher the rate.
• Citation and source attribution — A Columbia Journalism Review study found Grok-3 got answers wrong 94% of the time when identifying the original source of news excerpts. Citation fabrication rates reach 94% in adversarial testing across models.
• Open-ended generation — Open-ended tasks with no constraints show hallucination rates of 40–80%, the highest of any category.

Which AI Models Hallucinate Most and Least

The variation between models is significant — large enough that which tool you use matters for reliability as much as how you use it.

Model	Hallucination rate	Notes
Google Gemini 2.0 Flash	0.7% (Vectara benchmark)	Best-performing model on structured tasks as of April 2025
Gemini 2.0 Pro Exp	0.8%	Close second on structured benchmarks
OpenAI o3-mini-high	0.8%	Strong on structured tasks; worse on person-specific questions
Top 5 models (general)	10–20% (general knowledge)	Significant jump from structured to open-ended tasks
OpenAI o3 (person-specific)	33%	Reasoning models can perform worse on specific fact retrieval
TII Falcon-7B-Instruct	29.9%	Least reliable in Vectara benchmark — nearly 1 in 3 responses
LLMs on legal queries (Stanford)	69–88%	Across multiple models on specific legal questions

The reasoning model paradox: More advanced reasoning models — designed to think through problems step by step — sometimes hallucinate more on specific fact retrieval tasks than simpler models. OpenAI's o3 and o4-mini reached hallucination rates of 33% and 48% on person-specific questions, despite outperforming on complex reasoning tasks. The implication: a model that is better at thinking through problems is not automatically better at remembering facts. These are different capabilities, and conflating them is one of the most common mistakes in AI tool selection.

The Real-World Consequences

Hallucinations are not just an academic problem. The global financial cost of AI hallucinations reached $67.4 billion in 2024, driven by incorrect AI-assisted decisions in business, legal, medical, and financial contexts. The consequences play out differently across sectors, but the pattern is consistent: AI produces confident, specific, plausible-sounding wrong information, and humans act on it.

1. Legal consequences — The most high-profile hallucination cases have come from lawyers submitting AI-generated briefs containing fabricated case citations to courts. Courts have issued sanctions in documented cases. Stanford research found that LLMs hallucinate between 69% and 88% of the time on specific legal queries, meaning AI-generated legal research without independent verification is essentially unreliable by default.
2. Medical consequences — ECRI, a global healthcare safety nonprofit, listed AI risks as the number one health technology hazard for 2025. Open-source models still show hallucination rates above 80% in some medical tasks. Even proprietary models hallucinate 64% of the time on medical case summaries without mitigation. In a domain where a wrong answer can cause direct patient harm, these rates are not acceptable without significant human oversight.
3. Business and financial consequences — A 2026 UC San Diego study found AI-generated summaries hallucinated 60% of the time, influencing purchase decisions. The financial sector is particularly exposed because AI is being used for market analysis, financial summaries, and due diligence research — precisely the tasks where specific, verifiable facts matter most.
4. Journalism and information quality — AI-generated content that includes hallucinated facts contributes to the broader disinformation ecosystem. When AI-generated articles with fabricated statistics are published without verification, those statistics enter the information environment, get cited by other articles, and eventually become difficult to trace back to their fabricated origin.

How to Protect Yourself

The practical response to AI hallucination is not to stop using AI tools — it is to use them in ways that match their actual reliability profile rather than their perceived reliability.

1. Verify any specific claim that matters — The single most important rule. Statistics, citations, quotes, names, dates, and specific figures from AI should always be verified against primary sources before being relied upon. This is not a counsel of perfection — it is a basic quality control step that the reliability data makes necessary.
2. Ground AI in provided documents rather than memory — When you need AI to summarise, analyse, or answer questions about a specific subject, provide the relevant documents and ask it to work from those rather than from its training. Retrieval-augmented approaches — where AI answers from provided context — show dramatically lower hallucination rates than AI answering from training data alone.
3. Treat specific details with more suspicion than generalities — Given the confidence paradox, precise-sounding specific details — exact statistics, specific citations, precise quotes — should trigger verification, not confidence. The more specific and authoritative a claim sounds, the more important it is to check.
4. Ask AI to acknowledge uncertainty — Prompting AI to say "I'm not sure" when it lacks reliable information, or to distinguish between what it knows confidently and what it is inferring, can improve reliability. A 2025 Nature study found that prompt-based mitigation reduces hallucinations by approximately 22 percentage points. Medical AI research showed a 33% reduction using structured prompts. The model's default is to guess; explicitly giving it permission to express uncertainty shifts this.
5. Choose the right tool for the task — Gemini 2.0 Flash and o3-mini-high are the most reliable models for structured factual tasks. For legal or medical research, no current model is reliable enough without independent verification. For creative brainstorming, hallucination matters less. Matching tool capability to task requirement is more important than picking the "best" model in the abstract.
6. Use AI for structure, not for facts — AI is reliable for organising, structuring, and communicating information you already have or can verify. It is unreliable for retrieving specific facts you do not already know, especially in specialised domains. Using AI to draft, outline, or explain while you supply the verified facts through provided documents gets the benefit of AI capability while sidestepping the hallucination risk.

Will AI Ever Stop Hallucinating?

The trajectory is improving, but a complete solution is not on the near-term horizon — and there are structural reasons why it may never be completely eliminated.

The trend and the ceiling: Analysis of Hugging Face leaderboard data suggests that at the current rate of improvement — approximately 3 percentage points annually for top models — near-zero hallucination rates on structured tasks could be achievable by 2027. Some analyses suggest that zero hallucinations on broad tasks would require models with roughly 10 trillion parameters, a scale expected around 2027. But hallucination rates in open-ended generation tasks and specialised domains like law and medicine remain far higher and are improving more slowly. The gap between "reliable on structured tasks" and "reliable on everything" is significant and not closing at the same rate.

The deeper structural problem is that hallucination is partly a consequence of the thing that makes LLMs useful: the ability to generate fluent, contextually appropriate text in any domain. A model that never guessed would also never help you draft an email, explain a concept, or write a first pass at an analysis. The fluency and the hallucination come from the same place — prediction of probable continuations. You cannot eliminate one without affecting the other.

The most promising near-term approaches are retrieval augmentation (grounding AI in provided documents), improved uncertainty calibration (training models to know what they do not know), and structured verification workflows (using one AI system to check another's outputs). Anthropic, OpenAI, and Google DeepMind are all actively working on these approaches. The 91% of enterprises that have implemented explicit hallucination mitigation protocols are not waiting for the models to solve it — they are building the checks into their workflows. That is the right approach for anyone using AI in high-stakes contexts.

For context on how AI capabilities and limitations are shaping specific industries, see our guides on whether AI can diagnose patients, the future of AI and lawyers, and our beginner's guide to AI.

Frequently Asked Questions

What is an AI hallucination?

An AI hallucination is when an AI system generates information that is factually wrong, unverifiable, or entirely fabricated — presented with the same fluency and confidence as accurate information. Common examples include citations that do not exist, statistics that were never measured, quotes that were never said, and events that never happened. The AI is not lying deliberately — it has no mechanism to distinguish between what it knows reliably and what it is generating as a plausible-sounding continuation. The word "confabulation" is sometimes used as a more technically accurate description of what is actually happening.

Why does AI hallucinate?

Because AI language models are prediction engines, not knowledge databases. They generate text by predicting the most statistically probable next word based on patterns learned during training — they do not retrieve facts from a verified index. When a model encounters a question it does not have reliable information about, its training incentivises generating a plausible-sounding answer rather than admitting uncertainty, because models that attempt every question look better on accuracy metrics than models that frequently say "I don't know." OpenAI published research in 2026 confirming that standard training and evaluation procedures reward guessing over acknowledging uncertainty.

How often does AI hallucinate?

It varies enormously by task and model. The best models — Google Gemini 2.0 Flash and OpenAI o3-mini-high — hallucinate as little as 0.7–0.8% on structured tasks with provided documents. On general knowledge questions, the top five models cluster between 10–20%. On legal queries, Stanford research found rates of 69–88% across multiple models. On open-ended generation, rates of 40–80% are common. The global financial cost of AI hallucinations reached $67.4 billion in 2024, indicating the real-world scale of the problem even as the best models improve.

Which AI model hallucinates the least?

On the Vectara benchmark as of April 2025, Google Gemini 2.0 Flash recorded the lowest hallucination rate at 0.7%, followed by Gemini 2.0 Pro Exp and OpenAI o3-mini-high at 0.8%. Four models now sit below the 1% threshold on this structured benchmark. However, these rates apply to specific structured tasks — hallucination rates rise significantly on open-ended generation, person-specific questions, and specialised domains like law and medicine regardless of which model you use.

How can I tell if AI is hallucinating?

Often you cannot tell from the response itself — which is the core of the problem. AI uses the same confident, fluent tone for fabricated information as for accurate information. MIT research found models are 34% more likely to use highly confident language when generating incorrect information. Practical signals to watch for: very specific statistics with precise decimal places from unnamed or unpublished sources; citations to papers, books, or studies that cannot be found when searched; quotes attributed to specific named people that cannot be verified; highly specific details in domains where the AI is unlikely to have reliable training data (obscure historical events, specific legal cases, niche scientific research).

Is AI hallucination getting better?

Yes, significantly on structured tasks. Top models have improved from hallucination rates in the 15–25% range to under 1% on specific structured benchmarks over the past two years. At the current rate of improvement — approximately 3 percentage points annually — near-zero hallucination on structured tasks may be achievable by 2027. However, improvement on open-ended generation, legal research, medical tasks, and person-specific questions is slower. The structural issue — that hallucination partly results from the same mechanism that makes AI fluent and useful — means complete elimination is not expected anytime soon.

What can I do to reduce AI hallucinations?

Six practical steps: verify any specific claim that matters against primary sources; provide documents for AI to work from rather than asking it to recall from training; ask AI to acknowledge when it is uncertain rather than guessing; treat precise-sounding specific details with more suspicion than generalities; choose models matched to your task (Gemini Flash for structured factual work; avoid any model for unverified legal or medical research); and use AI for structure and communication while you supply verified facts through provided documents. Prompt-based mitigation — explicitly asking AI to express uncertainty — reduces hallucination rates by approximately 22 percentage points according to 2025 Nature research.

Can AI hallucinations cause real harm?

Yes, documented and measurable harm. Lawyers have faced court sanctions for submitting AI-generated briefs with fabricated case citations. ECRI listed AI as the number one health technology hazard for 2025, with hallucination rates in medical AI reaching 64% without mitigation. A 2026 UC San Diego study found AI summaries hallucinated 60% of the time, influencing purchase decisions. The global financial cost of AI hallucinations reached $67.4 billion in 2024. The harm is concentrated in high-stakes domains — law, medicine, finance, journalism — where specific, verifiable facts matter and where acting on incorrect information has real consequences.

Has AGI Already Arrived? What the Evidence Actually Shows in 2026

Has AGI Already Arrived?

Sam Altman has declared that "the takeoff has started" and that humanity is "past the event horizon" of the Singularity. Elon Musk says 2026 is the year of AGI. Dario Amodei at Anthropic expects systems matching "a country of geniuses" within two to three years. And yet Andrej Karpathy — the researcher who helped build GPT-4 — says we are a decade away. Demis Hassabis at DeepMind says current systems are impressive but nowhere near the full range of human cognition. A survey of AI researchers in 2023 put the median estimate at 2047. The question "has AGI arrived?" sounds simple. The answer depends entirely on what you mean — and that, it turns out, is one of the most contested questions in all of technology. This guide gives you the honest picture.

Table of Contents

1. What AGI Actually Means — and Why Nobody Agrees
2. The People Saying AGI Is Already Here
3. The People Saying It Is Not
4. What Current AI Can Actually Do
5. What Is Still Missing
6. The Definition Problem That Makes This Question Unanswerable
7. What the Experts Are Actually Predicting
8. The Honest Verdict
9. Frequently Asked Questions

What AGI Actually Means — and Why Nobody Agrees

Before you can answer whether AGI has arrived, you need to know what AGI is. And this is where the conversation immediately runs into trouble — because there is no consensus definition, and the people making the biggest claims are often using definitions that conveniently match what their systems already do.

The term Artificial General Intelligence was coined to describe an AI system that can perform any intellectual task that a human being can perform — not just the specific tasks it was designed and trained for, but any task, with the flexibility and adaptability of human cognition. A system that can learn a new job from a brief description, navigate an unfamiliar problem domain, generate genuinely original ideas, and apply knowledge from one field to solve problems in a completely different one.

The competing definitions in 2026: OpenAI uses an internal five-level framework ranging from basic chat assistant to "Organisations" — AI that can run entire companies autonomously. Google DeepMind published a formal "Levels of AGI" paper defining five tiers from Emerging to Superhuman, crossed with breadth from narrow to general. Sam Altman has called AGI "not a super useful term" because everyone defines it differently — a convenient position when your company has raised billions on AGI promises. The honest observation is that if you define AGI as "AI that can do most cognitive tasks most humans can do," current systems are arguably there for many tasks. If you define it as "AI with genuine understanding, self-motivated reasoning, and robust transfer learning across all domains," we are clearly not there.

The People Saying AGI Is Already Here

The most aggressive claims come from the people with the most financial interest in making them — which is worth keeping in mind, but does not automatically make them wrong.

Sam Altman — OpenAI

The CEO of OpenAI has been progressively escalating his rhetoric throughout 2025 and into 2026. In his essay "The Intelligence Age," he frames AGI not as a distant aspiration but as an impending transition already underway. He has stated that OpenAI is "now confident we know how to build AGI" and described humanity as being "past the event horizon." He has also suggested the world may be moving from the AGI conversation toward superintelligence — implying AGI is essentially solved.

Elon Musk — xAI

Musk has declared that "we have entered the singularity" and named 2026 as "the year of the Singularity." He previously predicted AGI by 2025, which passed without the milestone being widely acknowledged. His definition of AGI — "smarter than the smartest human" — is one of the more demanding ones, which makes his confidence in its imminent arrival all the more striking to his critics.

Dario Amodei — Anthropic

The CEO of Anthropic, in formal recommendations to the White House in March 2025, stated that "we expect powerful AI systems will emerge in late 2026 or early 2027." He has described AI systems arriving within two to three years that would be equivalent to "a country of geniuses" working on science and technology problems simultaneously. This is a near-term AGI prediction from someone who does not use the term AGI lightly.

The Microsoft Research GPT-4 paper

In 2023, Microsoft Research studied an early version of GPT-4 and published a paper claiming it showed "sparks of artificial general intelligence" — performing at human level in areas including mathematics, coding, and law. This triggered one of the first serious mainstream debates about whether AGI had arrived in some meaningful sense. The paper was contested but influential.

Why these claims deserve scrutiny: Every CEO making aggressive AGI predictions is running a company that needs continued investment, top talent, and public attention to survive one of the most capital-intensive technology races in history. Promising AGI in two years keeps investors writing cheques and talent from jumping ship to competitors. This does not mean the claims are wrong — but it does mean they should be evaluated with the same critical eye you would apply to any corporate forward guidance on a product that has not yet shipped.

The People Saying It Is Not

The sceptical voices are often less prominent in headlines but frequently more technically credible — and their arguments deserve equal attention.

Andrej Karpathy

Karpathy helped build GPT-4 and spent years as a senior researcher at OpenAI before leaving. He knows these systems as well as anyone alive. His assessment: AI agents "aren't anywhere close" to AGI, and genuine AGI is a decade away. When someone who built the most capable AI systems of their era says this, the people at the table who disagree have the burden of proof.

Demis Hassabis — Google DeepMind

The CEO of DeepMind — a company that was founded in 2010 with AGI as its explicit long-term goal, and which has been building toward it longer than almost anyone — has consistently maintained that current systems are impressive but not close to the full range of human cognitive capability. He has specifically identified creativity, continual learning, and robust understanding as gaps that current architectures do not address. He estimates a 50% chance of AGI by 2030. That is not a dismissive forecast — but it is conspicuously more cautious than the OpenAI timeline.

Yann LeCun — Meta AI

The Chief AI Scientist at Meta is the most prominent voice arguing that current large language model approaches may be architecturally incapable of reaching AGI at all. His position is not that AGI is far away — it is that the path being taken will not get there. LeCun has repeatedly argued that models trained on text alone cannot develop the grounded understanding of the physical world that genuine general intelligence requires.

Geoffrey Hinton

The Nobel Prize-winning AI researcher who helped develop the foundational neural network techniques underlying modern AI has revised his timelines toward the near term — but still expresses deep uncertainty, placing AI smarter than humans anywhere from roughly four to nineteen years away. His concern is less about whether it will happen and more about what happens when it does.

What Current AI Can Actually Do

Setting aside the definitional debate, it is worth being concrete about what systems like GPT-5, Claude Opus 4.6, and Gemini Deep Think can actually do in 2026 — because the capabilities are genuinely remarkable and genuinely uneven at the same time.

What current AI does remarkably well

• Coding and software development — Current models write, debug, and refactor complex code at a level that outperforms most professional developers on many tasks. Claude Code has been adopted widely by both experienced developers and non-programmers for automating software workflows.
• Mathematical reasoning — Google DeepMind's Gemini in Deep Think mode achieved gold-medal performance at the 2025 International Mathematical Olympiad, solving five out of six problems within the official contest window in natural language. This represents a significant threshold in AI's ability to reason through genuinely novel problems.
• Professional exam performance — Multiple current models pass the bar exam, medical licensing examination, and other professional certifications at above-average human scores. OpenEvidence scored 100% on the USMLE in 2025.
• Language and writing — Current models write at a quality level that routinely exceeds the average professional in many genres and formats.
• Multi-step agentic tasks — Modern AI agents can now handle complex workflows — researching, planning, executing, and iterating across multiple steps — with increasing reliability. Both Claude Opus 4.6 and GPT-5.3-Codex demonstrated significant advances in agentic capability in early 2026.

Where current AI still fails in ways that matter

• Hallucination — Current systems still confidently produce incorrect information, fabricated citations, and plausible-sounding falsehoods. GPT-5.5 recorded an 86% hallucination rate at uncertainty on one major benchmark. This is not a minor limitation — it is a fundamental reliability problem for high-stakes applications.
• Physical world grounding — AI has no sensory experience of the physical world. Its "understanding" of anything physical — medicine, engineering, cooking, sport — is derived entirely from text descriptions, not from embodied experience.
• Self-motivated reasoning — Genuine AGI would generate its own objectives, wonder, explore, and pursue goals that were never specified. Current AI responds to prompts. The difference is categorical.
• Robust transfer learning — A truly general intelligence would apply knowledge from one domain to a completely different one without explicit training. Current AI does this imperfectly and unpredictably.
• Genuine creativity and scientific discovery — Generating new scientific hypotheses, producing genuinely original artistic work that represents a departure from training data — these remain areas where current AI recombines rather than creates.

What Is Still Missing

The most important technical barriers to AGI are not about raw capability on benchmarks — they are about deeper architectural limitations that additional compute cannot straightforwardly solve.

1. Data exhaustion — Training models on more data has driven much of the capability improvement to date. But we have now consumed virtually all high-quality text available on the internet. Synthetic data — AI generating training data for itself — helps, but creates feedback loops that can degrade performance over time. The easy data scaling gains are behind us.
2. Compute scaling walls — Much of the improved performance from reasoning models came from giving them more time to think — essentially spending more compute at inference time. But there are not enough computer chips in the world to continue scaling thinking time indefinitely, and the economics of doing so are already approaching human labour costs for some tasks. This one-time gain cannot simply be repeated.
3. Architectural limitations — The transformer architecture that underlies most current AI may have inherent constraints that are only beginning to be understood. LeCun and others have argued that text-prediction models, however large, cannot develop the kind of world model that genuine general intelligence requires.
4. Alignment and safety — Even if a system achieved AGI-level capability, ensuring it reliably pursues beneficial goals — rather than optimising for something subtly different from what its designers intended — is an unsolved problem. The gap between AI capability and AI alignment is arguably widening, not narrowing, as systems become more powerful.

The Definition Problem That Makes This Question Unanswerable

Here is the uncomfortable truth at the heart of this debate: the question "has AGI arrived?" may be genuinely unanswerable in its current form — not because the answer is uncertain, but because the question is under-defined.

The definitional problem in plain language: If you define AGI as "AI that can pass professional exams and write better code than most humans," then AGI arrived in 2024 or 2025. If you define it as "AI that can do any cognitive task a human can do," we are not there — current AI fails on physical tasks, genuine creative reasoning, and self-directed goal pursuit. If you define it as "AI that understands the world the way humans do," we may never get there with current architectures, because understanding may require embodied experience rather than text prediction. The people declaring AGI has arrived and the people saying it has not are often talking about different things — and neither is wrong given their definition.

DeepMind's "Levels of AGI" framework is one of the more honest attempts to address this: rather than a binary arrived/not-arrived threshold, it defines five levels of capability and five levels of autonomy, and argues that the question should be "where on these scales are we?" rather than "have we crossed a line?" Under this framework, current systems are arguably at the "Competent" level for many tasks — outperforming 50% of skilled human adults — and approaching "Expert" level for specific domains like coding and mathematics. But they are far from "Superhuman" across the full range of cognitive tasks, and the autonomy dimension — how independently they can operate — is still very limited outside structured environments.

What the Experts Are Actually Predicting

Expert	Prediction	Their definition / caveat
Sam Altman (OpenAI)	"Past the event horizon" — now	Frames AGI as a transition already underway; shifting focus to superintelligence
Elon Musk (xAI)	2026 — "Year of the Singularity"	Defines AGI as smarter than the smartest human; previously predicted 2025
Dario Amodei (Anthropic)	Late 2026 or early 2027	"Powerful AI systems" — careful not to use AGI label directly
Mustafa Suleyman (Microsoft AI)	2027 — human-level on most professional tasks	Frames as "profound labour shock" rather than sci-fi threshold
Shane Legg (DeepMind)	50% chance by 2028	"Minimal AGI" — handles cognitive tasks most humans typically perform
Demis Hassabis (DeepMind)	50% chance by 2030	Emphasises creativity and scientific discovery as unresolved gaps
Andrej Karpathy	~10 years	Agents "aren't anywhere close"; helped build GPT-4
AI researcher survey (2023)	Median: 2047	AI performing all economically valuable tasks better and cheaper than humans

The Honest Verdict

After setting aside the definitional debate, the financial incentives, and the headline-generating extreme positions, here is what the evidence actually supports.

The honest answer: Current AI systems have crossed several thresholds that would have been called AGI-level a decade ago — they pass professional exams, write expert-quality code, solve olympiad mathematics, and handle many cognitive tasks at or above average human performance. In that narrow sense, something like partial AGI has arrived for specific domains. But by the more demanding definition — systems with genuine understanding, self-directed reasoning, robust transfer learning, and the ability to function autonomously across the full range of human cognitive tasks — we are clearly not there. The capabilities are uneven, the failures are fundamental, and the architectural barriers are real. The most honest framing is that we are somewhere in the middle of a spectrum, and the people arguing about whether we have "crossed a line" are arguing about where to draw a line that was never precisely defined in the first place.

What is clear is that whether or not the AGI label applies, the systems being built now are already transforming professions, economies, and daily life at a pace that was not predicted by mainstream forecasters even five years ago. The question of whether it technically counts as AGI matters less than the question of whether you are prepared for what these systems can already do — and what they will be able to do in the next three to five years, regardless of what we call them.

For context on how AI is already reshaping specific industries and jobs, see our guides on what jobs AI will replace, our beginner's guide to AI, and whether AI can diagnose patients.

Frequently Asked Questions

Has AGI already arrived in 2026?

It depends entirely on which definition you use. By a narrow definition — AI that passes professional exams and outperforms humans on specific cognitive tasks like coding and mathematical reasoning — something resembling partial AGI has arrived. By the more demanding definition — AI with genuine understanding, self-directed reasoning, and the ability to handle any cognitive task a human can — we are clearly not there. Current systems hallucinate confidently, cannot operate autonomously in unstructured environments, and lack the self-motivated goal pursuit that defines genuine general intelligence. The most honest answer is: partially, for specific domains, with significant limitations that matter enormously for high-stakes applications.

What is AGI and how is it different from current AI?

AGI — Artificial General Intelligence — refers to an AI system that can perform any intellectual task a human can, with the flexibility, adaptability, and generalisation of human cognition. Current AI systems are narrow in important ways: they are extraordinarily capable at the specific tasks they were trained on but fail unpredictably outside those domains, cannot pursue self-directed goals, cannot learn continuously from experience without retraining, and do not have the physical world grounding that underpins human understanding. The difference is not just capability level — it is a difference in the nature of the intelligence, not just its degree.

When do experts predict AGI will arrive?

Predictions range enormously depending on who you ask and how they define AGI. Sam Altman says we are already past the event horizon. Elon Musk predicted 2026. Dario Amodei at Anthropic expects powerful AI systems in late 2026 or early 2027. Mustafa Suleyman at Microsoft AI predicts human-level performance on most professional tasks by 2027. Shane Legg at DeepMind puts 50% odds on minimal AGI by 2028. Demis Hassabis at DeepMind says 50% by 2030. Andrej Karpathy, who helped build GPT-4, says about a decade. A 2023 survey of AI researchers produced a median estimate of 2047. The range reflects both genuine uncertainty about the technical trajectory and deep disagreement about what the target actually is.

Why do AI company CEOs keep predicting AGI so soon?

Partly because they genuinely believe it — the pace of capability improvement in 2023–2026 has been fast enough to rationally update timelines. But partly because the incentives are aligned with optimism: promising AGI in two years attracts investment capital, retains top researchers who want to work on transformative technology, and generates the public attention that drives product adoption. Sam Altman has acknowledged that AGI is "not a super useful term" because everyone defines it differently — a convenient position when your company has raised hundreds of billions of dollars on AGI promises. The most credible forecasters are those with the least financial stake in a particular timeline, which is why Karpathy's decade estimate deserves as much attention as Altman's "already here."

What are the main barriers preventing AGI right now?

The technical barriers most cited by researchers are: data exhaustion (we have consumed most high-quality human-generated text and synthetic data creates quality degradation problems), compute scaling limits (the gains from giving models more thinking time were partly a one-time improvement, not an indefinitely repeatable trend), architectural limitations (the transformer architecture may have inherent constraints for developing genuine world models), and alignment (ensuring a powerful AI reliably pursues beneficial goals is an unsolved problem that arguably gets harder, not easier, as systems become more capable). The question is not just whether AGI is coming but whether current approaches can get there at all.

Did GPT-4 or Claude show signs of AGI?

Microsoft Research published a paper in 2023 claiming GPT-4 showed "sparks of artificial general intelligence," citing human-level performance in mathematics, coding, and law. This was genuinely notable and triggered one of the first serious mainstream debates on the question. Critics pointed out that the same models fail on tasks a child handles easily, hallucinate confidently, and lack the continuity and self-direction of genuine intelligence. The "sparks" framing is probably the most accurate: impressive, domain-specific performance that suggests something significant is happening — but not evidence of the coherent general intelligence the term AGI implies.

Should I be worried about AGI?

The legitimate concerns are not primarily about AGI arriving tomorrow and immediately threatening human existence — that is the science fiction version. The legitimate concerns are more gradual: AI systems that are not quite AGI but capable enough to displace large numbers of workers, concentrate economic power among a small number of technology companies, be used for large-scale manipulation and disinformation, and in military applications, make lethal decisions faster than human oversight allows. These risks are present now and growing, without needing to wait for a formal AGI threshold to be crossed. The gap between AI capability and the governance frameworks designed to manage it is real and widening.

What would we know AGI had arrived?

This is genuinely one of the hardest questions in the field. There is no agreed test. The Turing Test — passing as human in conversation — was long cited but is now routinely passed by current systems in many contexts, without anyone seriously claiming AGI has therefore arrived. DeepMind's proposed evaluation for minimal AGI requires human testers with full system access being unable to find cognitive weak points after months of testing across a comprehensive range of tasks. OpenAI's internal Level 4 — "Innovators" — requires AI that can make genuine scientific discoveries. The honest answer is that we would probably argue about it even if it happened.

Will AI Be Able to Diagnose Patients? The Tools Available Now and What the Future Holds

Will AI Be Able to Diagnose Patients?

AI diagnosed a skin cancer that a dermatologist missed. An AI system scored 100% on the United States Medical Licensing Examination. And the FDA has now approved over 1,450 AI-enabled medical devices — the vast majority of them diagnostic tools. The question "will AI be able to diagnose patients?" has an answer in 2026: it already is. The more important questions are where it does this reliably, where it does not, which tools are genuinely proven, and what role human doctors will play as AI diagnostic capability continues to grow. This guide answers all of them.

Table of Contents

1. The Short Answer
2. What AI Can Already Diagnose — and How Accurately
3. The AI Diagnostic Tools Available Right Now
4. The FDA Approval Picture
5. AI vs Doctors: What the Research Actually Shows
6. What AI Cannot Do in Diagnosis
7. The Risks of AI Diagnosis That Need Honest Discussion
8. What the Future of AI Diagnosis Looks Like
9. Frequently Asked Questions

The Short Answer

AI is already diagnosing patients — not hypothetically and not just in research settings, but in clinics, hospitals, and radiology departments around the world every day. The more precise answer depends on what you mean by "diagnose." If you mean "can AI identify a disease from medical imaging with accuracy comparable to or exceeding a specialist physician" — then yes, for a growing number of conditions. If you mean "can AI replace a doctor and handle the full diagnostic process for any patient with any complaint" — then no, and that is a significantly harder problem that remains years away from being solved.

Where AI diagnostic capability actually stands in 2026: AI achieves diagnostic accuracy between 76% and 90% for imaging and clinical scenarios, often surpassing physician performance of 73–78% on tasks like mammogram reading and skin lesion detection. OpenEvidence — a clinical AI tool — scored 100% on the USMLE in 2025. A meta-analysis of 83 studies published in npj Digital Medicine found no significant overall performance difference between generative AI and physicians. GPT-4 outperformed emergency department resident physicians in diagnostic accuracy in a documented study. And the FDA has authorised 1,451 AI-enabled medical devices since it began tracking them, with radiology AI accounting for over 75% of approvals.

What AI Can Already Diagnose — and How Accurately

The areas where AI diagnostic capability is most proven are those involving pattern recognition in large volumes of medical images — which is precisely where human performance is most limited by fatigue, volume, and the inherent limits of the human visual system.

Radiology and medical imaging

This is where AI diagnostic capability is most mature and most extensively validated. AI systems can detect lung nodules, brain bleeds, bone fractures, and cardiac abnormalities in X-rays, CT scans, and MRIs with accuracy that equals or exceeds radiologists in controlled studies. In stroke detection specifically, AI has demonstrated the ability to identify bleeds and large vessel occlusions faster than a radiologist could review the scan — which matters enormously when every minute of treatment delay corresponds to measurable brain damage.

Cancer detection

AI achieves up to 90% sensitivity in detecting breast cancer from mammograms — surpassing the traditional radiologist accuracy rate of 73–78% on this specific task. For skin cancer, AI systems trained on large dermoscopy datasets have matched or exceeded dermatologist accuracy in identifying melanoma and other skin malignancies. Google's DeepMind developed an AI that detected over 50 eye conditions from retinal scans with accuracy equivalent to world-leading specialists, while also identifying systemic diseases — including cardiovascular risk and early diabetes — from the eye image alone.

Pathology

AI is transforming pathology — the analysis of tissue samples under a microscope. Whole-slide image analysis platforms can examine digitised tissue samples and identify cancerous cells, grade tumours, and detect patterns that correlate with treatment response. Companies like Paige AI have received FDA breakthrough designation for AI pathology tools that assist pathologists in identifying prostate cancer. The accuracy advantage is particularly pronounced for rare tumour types where individual pathologists may have limited experience.

Cardiology

AI algorithms reading electrocardiograms can identify arrhythmias, structural heart disease, and even low ejection fraction — a marker of heart failure — with accuracy that outperforms general practitioners and in some studies matches cardiologists. Apple Watch's FDA-cleared ECG app is the most consumer-visible example of AI cardiac diagnosis reaching everyday life. In clinical settings, AI ECG analysis is being used to flag patients who might have undiagnosed atrial fibrillation or other conditions before symptoms become obvious.

Mental health screening

AI analysis of speech patterns, language use, facial microexpressions, and writing can now identify markers of depression, anxiety, early cognitive decline, and even psychosis risk with meaningful accuracy. These tools are not replacing psychiatric assessment, but they are enabling early screening at scale — identifying people who may need evaluation before they would self-present to a clinician.

The AI Diagnostic Tools Available Right Now

1. Aidoc — One of the most widely deployed radiology AI platforms in the US, Aidoc's software runs in the background of hospital radiology workflows, automatically flagging critical findings — intracranial bleeds, pulmonary embolisms, aortic dissections — and elevating them to the top of the radiologist's worklist. It operates 24/7 without fatigue. Deployed in over 1,000 medical centres globally. FDA cleared for multiple indications.
2. Qure.ai — A radiology AI platform particularly focused on chest X-ray interpretation, tuberculosis detection, and head CT analysis. Qure.ai has been specifically designed for high-volume, lower-resource environments and has been deployed in screening programmes across India, Southeast Asia, and Africa. Its TB detection capability is particularly significant in settings where radiologist capacity is severely limited.
3. Google DeepMind / Health AI — DeepMind's AI has demonstrated the ability to detect over 50 eye conditions from retinal scans, identify breast cancer from mammograms at above-radiologist accuracy, and predict acute kidney injury 48 hours before clinical deterioration. Their work on chest X-ray analysis has shown consistent performance gains over radiologist baseline in multi-site studies.
4. Paige AI — Paige AI is Focused on computational pathology. FDA cleared for prostate cancer detection from digitised tissue slides. The platform assists pathologists by pre-screening slides and highlighting regions of concern, reducing the time pathologists spend on normal slides and improving detection rates for subtle cases.
5. OpenEvidence — A clinical AI tool built on the Mayo Clinic Platform that scored 100% on the USMLE in 2025. It functions as a clinical decision support system, helping physicians navigate differential diagnoses, review relevant evidence, and interpret complex cases. It includes a "Deep Consult" feature for comprehensive case analysis. Free for US physicians with an NPI number.
6. GE HealthCare AI suite — GE HealthCare leads the FDA approval count with over 120 cleared AI radiology tools. Their AI portfolio covers mammography (Senographe Pristina), CT analysis, MRI interpretation, and cardiac imaging, integrating AI recommendations directly into imaging workflow software used in hospitals worldwide.
7. Viz.ai — Specialises in time-critical conditions: stroke, pulmonary embolism, and aortic dissection. Viz.ai's platform analyses CT scans in real time, contacts the on-call specialist directly with images and AI findings if a critical condition is detected, dramatically reducing the time from imaging to treatment. Studies have shown it reduces time-to-treatment for stroke by 96 minutes on average.
8. Tempus AI — Focused on oncology. Tempus integrates clinical data, genomic sequencing, and AI to identify cancer treatment options matched to a patient's specific tumour profile. It is one of the most sophisticated examples of AI moving from diagnosis toward personalised treatment recommendation — a step beyond pattern recognition into clinical reasoning.

The FDA Approval Picture

The scale of regulatory approval for AI diagnostic tools is one of the clearest signals that this is not experimental technology. The FDA has authorised 1,451 AI-enabled medical devices since it began tracking them — and the pace of approvals is accelerating, not slowing.

FDA AI approval numbers (end of 2025): 1,451 total AI-enabled medical devices approved. 1,104 are radiology devices — 76% of all approved AI medical devices. Radiology approvals have grown from approximately 500 in early 2023 to over 1,100 by end of 2025 — more than doubling in two years. GE HealthCare leads with 120 approvals, followed by Siemens Healthineers (89), Philips (50), Canon (45), and United Imaging (38). Approvals now cover radiology, cardiology, neurology, pathology, and beyond. Over 200 AI vendors exhibited at the Radiological Society of North America's 2025 annual meeting.

The regulatory framework matters because it is the difference between AI tools that have been rigorously tested for safety and performance and those that have not. FDA-cleared tools have gone through validation studies demonstrating they do what they claim to do, in the patient populations they will be used on, without causing unacceptable rates of false negatives or false positives. The fact that over 1,100 radiology AI tools have cleared this process is a meaningful indicator of the maturity and safety profile of medical imaging AI in 2026.

The EU AI Act dimension: From 2026, the EU AI Act classifies medical diagnostic AI as "high-risk," requiring documentation of training data curation, bias checks, and human oversight policies. This creates a stricter compliance environment for AI diagnostic tools in Europe than currently exists in the US. The regulatory divergence between the US (where an executive order aims to reduce barriers to medical AI) and the EU (where a comprehensive risk framework applies) will shape which tools reach patients first in each market.

AI vs Doctors: What the Research Actually Shows

The research on AI diagnostic accuracy versus physician accuracy is more nuanced than headlines suggest — and understanding the nuance matters for understanding where AI is actually useful.

Diagnostic task	AI performance	Human comparison
Mammogram reading (breast cancer)	Up to 90% sensitivity	Radiologist 73–78% — AI leads
Skin lesion classification	Matches or exceeds dermatologists	Performance varies by experience level
Chest X-ray (multi-condition)	76–88% accuracy depending on condition	Comparable to general radiologist
Emergency department diagnosis (general)	GPT-4 outperformed ED resident physicians	Resident physicians — AI leads; specialists less clear
General clinical vignettes (USMLE)	100% (OpenEvidence 2025)	Above passing threshold for physicians
Stroke detection from CT	Real-time, 96 min faster treatment (Viz.ai)	Fatigue and volume affect human performance at night
Complex specialist cases, rare diseases	52.1% overall (meta-analysis of 83 studies)	No significant difference from physicians overall

What the overall meta-analysis actually found: A systematic review and meta-analysis of 83 studies published in npj Digital Medicine in 2025 found an overall AI diagnostic accuracy of 52.1%, with no significant performance difference between AI and physicians overall. This sounds underwhelming until you understand what it means: AI performs at physician level across a wide range of diagnostic tasks — including many where physician performance itself is far from perfect. For specific high-volume imaging tasks, AI significantly outperforms average physician performance. For rare diseases and complex multi-system presentations, AI and physicians are roughly equal — both with room for improvement.

What AI Cannot Do in Diagnosis

Where AI diagnostic capability is strong

• High-volume pattern recognition in medical images (radiology, pathology, dermatology)
• Consistent, tireless screening without the performance degradation human fatigue causes
• Flagging critical findings instantly and escalating to the right clinician
• Integrating data from multiple sources — imaging, lab results, EHR, genomics — simultaneously
• Applying the latest research evidence consistently, without the knowledge decay that affects busy clinicians
• Operating in low-resource environments where specialist physicians are unavailable

Where AI diagnostic capability falls short

• Taking a history — The clinical history — what the patient tells a doctor about their symptoms, context, and concerns — is the most information-rich part of diagnosis for most conditions. AI cannot yet conduct this with the depth and flexibility that a skilled physician brings.
• Physical examination — Touch, sound, and the direct physical assessment of a patient remains outside current AI capability. Many diagnoses depend on findings that can only be obtained by a human examiner.
• Contextual judgment in ambiguous presentations — When a patient has atypical symptoms, multiple overlapping conditions, or a presentation that does not fit standard patterns, the experienced physician's ability to integrate complex contextual information remains superior to current AI.
• Patient communication and shared decision-making — Delivering a diagnosis, discussing prognosis, and working with a patient through complex treatment decisions requires the kind of human empathy and relationship that AI cannot provide.
• Rare and novel conditions — AI models trained on historical data perform poorly on conditions with limited training examples, or on genuinely novel presentations that do not match patterns in the training set.
• Professional accountability — A doctor is personally and legally accountable for their diagnostic conclusions. AI is a tool; the physician remains the accountable decision-maker in all current regulatory frameworks.

The Risks of AI Diagnosis That Need Honest Discussion

The genuine promise of AI diagnosis is real. So are the risks. Most coverage focuses on the former; the latter deserve equal attention.

Algorithmic bias in medical AI: AI diagnostic tools are only as good as the data they were trained on. If a tool was trained primarily on images from patients of one ethnicity, age group, or body type, its performance on other populations may be significantly worse than the headline accuracy figures suggest. Several studies have documented performance disparities in AI diagnostic tools across racial and demographic groups. The FDA approval process requires validation across relevant populations, but this does not guarantee equal performance in the real world — particularly when the diversity of training data falls short of the diversity of real patients.

1. Over-reliance and skill erosion — There is genuine concern in the medical community that if clinicians defer to AI diagnostic recommendations routinely, they may develop less skill at independent diagnosis over time. The same dependency effect seen in educational AI is plausible in medical AI: a clinician who always has an AI second opinion may develop less confidence and capability in the situations where the AI is unavailable or wrong.
2. False negatives at scale — When an AI system is deployed at high volume, even a small false negative rate translates into a significant number of missed diagnoses in absolute terms. A 5% false negative rate applied to millions of mammogram screenings means hundreds of thousands of missed cancers. The aggregate impact of AI error rates at deployment scale is qualitatively different from the individual-level accuracy figures in clinical studies.
3. Liability and accountability gaps — When an AI diagnostic tool contributes to a missed or wrong diagnosis, who is responsible? The current answer — the physician retains accountability — creates a logical tension when AI systems are demonstrably more accurate than the physician in specific tasks. Malpractice law, professional liability frameworks, and healthcare insurance have not yet fully resolved how AI-assisted diagnosis changes the accountability picture.
4. Privacy and data security — AI diagnostic tools require access to sensitive medical data — imaging, genomics, clinical records — to function. The data pipelines, cloud storage, and third-party integrations involved in AI diagnostic platforms create data privacy risks that are significant given the sensitivity of the information involved.

What the Future of AI Diagnosis Looks Like

The trajectory of AI diagnostic capability is consistent and clear, even if the precise timeline is not.

1. Now — 2027 (Deep integration in radiology and pathology): AI becomes standard infrastructure in hospital imaging departments, not an add-on. Real-time AI flagging of critical findings is the norm rather than the exception. AI pathology platforms become routine in oncology centres. Multimodal AI — integrating imaging, genomics, and clinical data simultaneously — begins reaching clinical deployment. Patients in well-resourced healthcare systems increasingly receive AI-assisted diagnosis without knowing it.
2. 2027–2030 (Expansion beyond imaging): AI diagnostic capability expands from imaging-dominated applications into primary care screening and general medicine. AI-powered physical examination tools — digital stethoscopes with AI analysis, smart wearables monitoring continuous biomarker data, AI-assisted endoscopy — bring AI into examination room encounters. Large language model-based clinical decision support tools become standard for physicians navigating complex cases. Personalised AI that knows a patient's complete medical history, genomic profile, and longitudinal health data begins enabling predictive diagnosis — identifying conditions before symptoms appear.
3. 2030 and beyond (The integrated picture): The question shifts from "can AI diagnose?" to "what is the right division of labour between AI and physicians?" The most likely answer is a model where AI handles the high-volume pattern recognition, screening, and triage functions at scale, while physicians focus on complex presentations, ambiguous cases, patient communication, and the judgment calls that require contextual understanding and professional accountability. This is not a future where AI replaces doctors — it is a future where the doctor's role is redefined around the judgment and human elements that AI cannot replicate.

What this means for patients right now: If you are in a major hospital or healthcare system, there is a reasonable chance AI is already assisting in reading your scans, flagging abnormalities, and supporting your radiologist's workflow — whether or not anyone told you. This is generally a positive development: the evidence supports AI improving diagnostic accuracy and speed for many conditions. The questions worth asking your care provider are not "is AI being used?" but "what tools are being used, how have they been validated, and how does the physician verify AI recommendations?"

For broader context on how AI is changing healthcare, see our guides on AI and automation in healthcare, AI in radiology: pros and cons, and how long until AI replaces doctors.

Frequently Asked Questions

Can AI diagnose diseases accurately?

Yes — for specific, well-defined diagnostic tasks, particularly in medical imaging. AI achieves diagnostic accuracy between 76% and 90% for imaging tasks, often surpassing average physician performance on high-volume screening tasks like mammogram reading and skin lesion classification. A meta-analysis of 83 studies found no significant overall performance difference between generative AI and physicians. For complex, multi-system presentations and rare diseases, AI and physicians perform similarly — both with room for improvement. AI is not universally better than doctors, but for specific image-based diagnostic tasks it is demonstrably and consistently accurate.

What AI diagnostic tools are FDA approved?

The FDA has approved 1,451 AI-enabled medical devices as of end of 2025, of which 1,104 are radiology tools — over 75% of all approvals. Leading companies include GE HealthCare (120 approvals), Siemens Healthineers (89), Philips (50), Canon (45), and specialist platforms like Aidoc (31) and DeepHealth (28). Specific tools include Aidoc for critical finding detection, Viz.ai for stroke and pulmonary embolism, Paige AI for prostate cancer pathology, and extensive imaging analysis tools from GE, Siemens, Fujifilm, and Qure.ai. The full FDA list is publicly available through the FDA's Digital Health Center of Excellence.

Will AI replace doctors for diagnosis?

Not for the full diagnostic process — and not in any foreseeable near-term timeframe. AI excels at specific, well-defined pattern recognition tasks in high volumes of structured data. It cannot take a clinical history, perform a physical examination, integrate complex contextual information about an individual patient, or bear professional accountability for its conclusions. The most likely future is a division of labour where AI handles high-volume screening and imaging analysis while physicians focus on complex presentations, patient communication, and the judgment calls that require contextual understanding. This makes both the AI and the physician more effective than either would be alone.

How accurate is AI at reading medical scans?

For specific conditions, AI accuracy in medical imaging now matches or exceeds trained specialists. AI achieves up to 90% sensitivity for breast cancer detection from mammograms — above the 73–78% radiologist baseline on this task. For stroke detection, Viz.ai reduces average time-to-treatment by 96 minutes, reflecting its ability to identify findings and escalate faster than human workflow allows. For chest X-ray multi-condition analysis, AI performs comparably to general radiologists. The FDA's approval of over 1,100 radiology AI tools, all requiring validation studies demonstrating clinical performance, reflects the maturity of AI imaging accuracy in 2026.

Is AI being used to diagnose patients right now?

Yes — broadly and in routine clinical practice. Aidoc is deployed in over 1,000 medical centres globally. Viz.ai is active in major stroke centres across the US. GE HealthCare and Siemens AI tools are built into the imaging workflows of thousands of hospitals. Patients in major healthcare systems are routinely receiving AI-assisted radiology analysis, often without being explicitly informed. AI diagnostic tools are also being used in primary care screening apps and wearables — Apple Watch's FDA-cleared ECG is the most common consumer example.

What are the risks of AI diagnosis?

Four risks deserve the most attention: algorithmic bias, where AI trained on non-diverse data performs worse on underrepresented patient populations; false negatives at scale, where even small error rates produce large absolute numbers of missed diagnoses across millions of patients; liability gaps, where the accountability structure for AI-assisted diagnostic errors remains legally unresolved; and clinician deskilling, where routine AI reliance may reduce the independent diagnostic capability of physicians over time. These are manageable risks with appropriate governance — but they require deliberate attention from healthcare systems deploying AI diagnostic tools.

Can AI diagnose from symptoms alone?

Partially — symptom checkers and clinical decision support tools can generate differential diagnoses from symptom input, and tools like OpenEvidence and Harvey AI (legal context) can navigate complex clinical scenarios at high accuracy. GPT-4 has outperformed emergency department resident physicians on diagnostic accuracy from clinical case descriptions in controlled studies. However, symptom-based AI diagnosis has higher error rates than image-based AI diagnosis, and all current tools require physician verification. Symptom checkers are best used as triage and navigation tools — helping people understand whether and how urgently they need to see a doctor — rather than as replacements for clinical assessment.

What does AI diagnosis mean for the future of doctors?

It means a redefinition of what doctors spend their time on, not an elimination of the profession. As AI handles an increasing share of high-volume pattern recognition — reading scans, screening for common conditions, flagging critical findings — physician time concentrates on the work that AI cannot do: complex clinical judgment, patient relationships, ethical decision-making, and professional accountability. The physicians most at risk are those whose practice is dominated by tasks AI performs well. Those who develop expertise in complex, judgment-intensive, relationship-dependent medicine are well-positioned in a world where AI is a powerful partner in the diagnostic process.

• Has AGI Already Arrived?

AI and Mental Health: Can a Chatbot Replace a Therapist?

There are roughly 356,500 mental health clinicians in the United States — about one per 1,000 people. Half of all adults with a mental illness never receive any treatment. The median wait time for a first therapy appointment is 25 days; in rural areas, it is often six months or more. A single therapy session costs $100–$200. Against this backdrop, over 40 million people worldwide now use AI mental health apps every month. The question is not whether people are turning to AI for mental health support — they already are, at scale. The question is whether it helps, who it helps, and where the line is between a useful tool and a dangerous substitute for real care.

Table of Contents

1. The Problem AI Is Trying to Solve
2. What the Research Actually Shows
3. The AI Mental Health Tools Available Right Now
4. AI vs a Human Therapist: An Honest Comparison
5. What AI Cannot Do in Mental Health Care
6. The Risks That Deserve Honest Discussion
7. Who Should Use AI Mental Health Tools — and Who Should Not
8. What the Future Looks Like
9. Frequently Asked Questions

The Problem AI Is Trying to Solve

The mental health crisis in most developed countries is not primarily a treatment quality problem — it is an access and capacity problem. The treatments that work for anxiety and depression are well-established: cognitive behavioural therapy, medication, and their combination have decades of evidence behind them. The problem is that most people who need these treatments never access them.

The access gap in numbers: 356,500 mental health clinicians serve a US population of 330 million — roughly one clinician per 1,000 people. Half of all adults with mental illness receive no treatment. The average wait for a first appointment is 25 days nationally, and over six months in many rural areas. At $100–$200 per session, a standard 12-session course of CBT costs $1,200–$2,400 out of pocket. 32% of people globally say they would be willing to use AI for mental health support. The apps that exist are trying to serve the enormous space between "I'm struggling" and "I'm in crisis" — the daily anxiety, low-grade depression, and emotional dysregulation that millions experience but never seek help for.

This is the context in which AI mental health tools need to be evaluated. The question is not whether a chatbot is as good as a skilled human therapist — it clearly is not. The question is whether a chatbot is better than nothing, for the millions of people for whom nothing is the realistic alternative.

What the Research Actually Shows

The research on AI mental health tools is more rigorous than many people assume — and more cautious than the apps' marketing suggests.

The landmark NEJM study — Therabot

The most significant clinical evidence published in 2025 came from a randomised controlled trial of Therabot, published in NEJM AI. This was the first RCT demonstrating the effectiveness of a fully generative AI therapy chatbot for treating clinical-level mental health symptoms. Participants used the app for an average of over six hours and rated the therapeutic alliance — their sense of connection and trust with the system — as comparable to human therapists. Results showed significant symptom reduction for major depressive disorder, generalised anxiety disorder, and eating disorder symptoms.

The broader evidence base

A systematic review and meta-analysis of generative AI mental health chatbots published in the Journal of Medical Internet Research in December 2025 — covering 5,555 screened records — found that AI chatbots produced measurable reductions in anxiety and depression in randomised controlled trials. A separate meta-analysis of 31 RCTs covering interventions for adolescents and young adults published in November 2025 found consistent positive effects on mental distress.

The honest caveat: The JMIR meta-analysis noted substantial heterogeneity across studies, moderate risk of bias, and a relatively small number of high-quality RCTs. The researchers explicitly cautioned that conclusions should be viewed as a foundation for future research rather than definitive evidence of efficacy. The evidence is promising, not conclusive — and the gap between app marketing and actual research quality is significant for many tools on the market.

Woebot's key finding

A 2023 RCT found Woebot's programme for teenagers non-inferior to clinician-led therapy for reducing depressive symptoms. For an app that costs nothing and is available at 3am, that finding has real implications for the access gap described above.

The AI Mental Health Tools Available Right Now

• Woebot — Developed by clinical psychologists at Stanford University, Woebot uses structured CBT-based interventions through short daily conversations. Backed by over 10 peer-reviewed studies. A 2023 RCT found it non-inferior to clinician therapy for teenagers. FDA Breakthrough Device designation for postpartum depression. Pursuing full FDA De Novo classification. Free to download; enterprise versions available for health systems and universities.
• Wysa — Combines CBT, DBT, mindfulness, and motivational interviewing through a conversational interface. Among 527 healthcare workers, 94% completed at least one session and 80% returned, averaging 10.9 sessions each. FDA Breakthrough Device status in 2025 for chronic pain-related mental health. Hybrid model connects users to human therapists when needed. Free tier with 150+ exercises; premium approximately $60–$75 per year.
• Therabot — The first fully generative AI therapy chatbot validated in a clinical RCT (NEJM AI, 2025). Designed for clinical-level symptoms including major depression and generalised anxiety. Users rated therapeutic alliance comparable to human therapists. Still in research and early deployment rather than mass-market release — represents the clinical frontier.
• Youper — AI-driven mood assessments and cognitive reframing conversations. Clinical evaluations show regular use reduces anxiety and improves self-awareness within a few weeks. Strong for mood tracking and in-the-moment emotional support. Free with premium features.
• Earkick — Focused on real-time emotional regulation during acute anxiety and panic attacks. Voice check-in analyses vocal tone and emotional content to respond when typing while dysregulated is impractical. Works best as a complement to human therapy. Free with premium at approximately $48 per year.
• Headspace Ebb — Headspace's AI therapy layer. Combines evidence-based mindfulness content with AI-driven emotional support conversations. Best suited to stress and mild anxiety rather than clinical symptoms.
• Replika — AI companion focused on emotional connection and conversation. Particularly used by people experiencing loneliness. Does not deliver evidence-based therapeutic interventions, but the social support dimension has value — though it has generated significant controversy around dependency and unhealthy attachment.

AI vs a Human Therapist: An Honest Comparison

Dimension	AI mental health tool	Human therapist
Availability	24/7, immediate, no waiting list	Scheduled, 25+ day average wait
Cost	Free to ~$75/year	$100–$200 per session
Evidence base	Strong for CBT tools, mild-moderate conditions	Extensive across all severity levels
Human connection	Simulated — not genuine empathy	Real therapeutic relationship — strongest outcome predictor
Crisis response	Limited — refers to crisis lines only	Full crisis assessment and intervention
Stigma barrier	None — anonymous and private	Persistent stigma for many people
Complex conditions	Not appropriate for severe illness	Equipped for all condition types and severities

What AI Cannot Do in Mental Health Care

Where AI mental health tools genuinely help

• Providing immediate support at 3am when nothing else is available
• Removing the stigma barrier for people not ready to see a human therapist
• Delivering CBT and DBT skill-building exercises consistently and at scale
• Supporting people on waiting lists in the interim
• Providing between-session support for people already in human therapy
• Reaching populations geographically or financially excluded from traditional care
• Mood tracking and pattern identification over time

Where AI mental health tools fall short or cause harm

• Severe mental illness — PTSD, psychosis, bipolar disorder, severe depression, active suicidality require human clinical care. Every reputable AI tool explicitly states it is not designed for these conditions.
• Crisis intervention — AI cannot assess suicide risk in real time, make safety plans, or coordinate emergency response.
• Genuine therapeutic relationship — Real empathy, deep understanding of someone's history, and human trust are the strongest predictors of therapy outcomes. AI simulates this but cannot provide it.
• Trauma processing — Complex trauma requires skilled human clinical work and real relational presence.
• Medication decisions — AI has no role in psychiatric medication assessment or management.

The Risks That Deserve Honest Discussion

The CharacterAI incident: Media reports have linked a CharacterAI chatbot to a teenager's suicide. OpenAI has acknowledged that its general-purpose chatbot worsened delusional thinking in a user with autism. The American Psychological Association responded by urging the FTC to oversee mental health chatbots lacking clinical validation. The difference between a well-designed, clinically validated tool like Woebot or Wysa — built with safety guardrails, crisis protocols, and evidence-based frameworks — and a general-purpose chatbot used for emotional support is not a matter of degree. It is a categorical difference in safety.

1. The false sense of adequate care — The most pervasive risk is subtle inadequacy: a person with significant mental illness using an AI app as a substitute for professional care they genuinely need, feeling like they are addressing their situation while not receiving the level of help that would actually make a difference.
2. Dependency without progress — Some users develop attachment to AI companions without experiencing clinical improvement. Replika has generated documented cases of emotional dependencies that harm real-world relationships. An app that makes someone feel better without addressing the underlying condition may delay recovery.
3. Hallucinated or harmful advice — General-purpose AI used for mental health conversations can produce clinically inappropriate or actively dangerous advice. This is why clinical apps like Woebot and Wysa are built on constrained, evidence-based frameworks — the constraint is a feature, not a limitation.
4. Privacy and data sensitivity — Mental health data is among the most sensitive personal information that exists. The FTC fined two mental health apps in 2025 for deceptive advertising about data practices. Before using any mental health app, read the actual privacy policy — not the marketing summary.

Who Should Use AI Mental Health Tools — and Who Should Not

The honest rule of thumb: AI mental health tools are most appropriate as a bridge, a supplement, or a first step — not as primary care for significant mental illness. If your symptoms are mild to moderate, if you are on a waiting list, if you need between-session support, or if stigma is preventing you from seeking help — these tools have genuine evidence behind them. If you are in crisis, have serious mental illness, or have tried an AI tool for 4–6 weeks without improvement — human professional care is what you need.

1. Good fit for AI tools: Mild-to-moderate anxiety or depression. People on a waiting list needing interim support. People supplementing existing human therapy. People for whom stigma is a barrier. People where traditional therapy is not financially or geographically accessible. Teenagers experiencing stress not ready to speak to an adult.
2. Not appropriate for AI tools: Active suicidal ideation or self-harm. Psychosis or delusional thinking. Severe depression. PTSD and complex trauma. Bipolar disorder. Any safety concern. Anyone without improvement after 4–6 weeks should transition to human therapy — most reputable apps have built-in pathways to licensed therapists at this point.
3. Using AI alongside human therapy: Apps like Earkick and Wysa generate mood reports and session summaries that can be shared with a human therapist, providing richer insight into a client's week. This supplementary model — where AI enriches the human therapeutic relationship — has the strongest evidence base.

For broader context on how AI is transforming healthcare, see our guides on AI and automation in healthcare and our analysis of how long until AI replaces doctors.

What the Future Looks Like

1. Near term — prescription digital therapeutics: If Woebot receives full FDA De Novo authorisation it will be the first formally FDA-cleared AI therapy chatbot, opening insurance reimbursement and dramatically increasing access. FDA guidance for AI mental health tools is expected in late 2026.
2. Medium term — multimodal emotion detection: Apps are beginning to analyse facial expressions, vocal tone, typing patterns, and wearable physiological data. More accurate emotional state detection improves clinical value — and raises significant privacy questions that regulatory frameworks need to address before deployment at scale.
3. Longer term — LLM-powered therapy: The shift from scripted chatbot responses to open-ended generative AI conversations is already underway — Therabot is the most advanced clinical example. More natural, therapeutically flexible interactions come with new risks of harmful advice in clinical contexts. Balancing conversational freedom with clinical safety will define the next generation of mental health AI.

The most important thing to understand about AI and mental health: The goal of well-designed AI mental health tools is not to replace human therapists. It is to make the wait shorter, more supported, and less damaging — and to reach the half of people with mental illness who currently receive nothing at all. That is a meaningful and achievable goal. It is a much more modest ambition than "replace therapy" — and it is one that the best tools in this space are already delivering on.

Frequently Asked Questions

Can an AI chatbot replace a therapist?

No — and the best AI mental health tools are explicit about this. What AI can do is provide immediate, accessible, evidence-based support for mild-to-moderate conditions, reduce the harm of the access gap, and supplement ongoing human therapy with between-session tools. The therapeutic alliance between a human therapist and client is the single strongest predictor of therapy outcomes and is something AI cannot replicate. For mild anxiety and stress, the evidence behind tools like Woebot and Wysa is genuinely encouraging. For serious mental illness, AI is not an adequate substitute.

Do AI therapy apps actually work?

For specific conditions and clinically designed tools, yes. A 2025 RCT published in NEJM AI found Therabot produced significant symptom reduction for clinical-level depression, anxiety, and eating disorder symptoms. A 2023 RCT found Woebot non-inferior to clinician therapy for teenage depression. A December 2025 JMIR meta-analysis found measurable anxiety and depression reduction from RCTs of AI chatbots. The honest caveat: results apply most strongly to mild-to-moderate conditions using validated tools — not general wellness apps.

What is the best AI mental health app?

For clinical evidence and safety, Woebot and Wysa have the strongest research bases. Both have FDA Breakthrough Device designation. Woebot uses structured CBT from Stanford psychologists. Wysa offers 150+ CBT/DBT exercises and a hybrid model connecting to human therapists. Earkick is best for acute anxiety regulation. Therabot is the clinical frontier but not yet widely available as a consumer app. The right choice depends on your specific need.

Who should not use AI mental health apps?

People experiencing active suicidal ideation, psychosis, severe depression, PTSD, bipolar disorder, or any mental health crisis should seek human professional care. Every reputable tool explicitly states these limitations. People who have used an AI tool consistently for 4–6 weeks without improvement should transition to human therapy — most platforms including Wysa have built-in pathways to licensed therapists for exactly this situation.

Are AI therapy apps safe?

Clinically designed tools with safety guardrails — like Woebot and Wysa — have strong safety profiles for their intended use cases. General-purpose AI chatbots used for mental health are not safe in the same way. Documented incidents include worsened delusional thinking and a widely reported link to a teenager's suicide. Look for FDA status, published clinical trials, and explicit crisis escalation protocols. Never use general-purpose AI chatbots as substitutes for mental health care.

Are AI mental health apps private?

It varies. Woebot is HIPAA-aligned. Wysa anonymises data by design. The FTC fined two mental health apps in 2025 for deceptive data practice claims. Read the actual privacy policy before using any mental health app — key questions are who owns your data, whether it is sold to third parties, and whether you can delete it.

How much do AI therapy apps cost?

Most have meaningful free tiers. Woebot is free. Wysa premium is approximately $60–$75 per year. Earkick premium is approximately $48 per year. Compare with human therapy at $100–$200 per session, and the access argument for AI tools becomes clear for people who cannot afford or access traditional care.

What is the future of AI in mental health treatment?

Three developments will define it: regulatory maturation — FDA authorisation of tools like Woebot enabling insurance reimbursement and greater access; multimodal emotion detection — apps reading voice tone, facial expression, and physiological data for more accurate clinical assessment; and LLM-powered therapy — the shift to open-ended generative AI conversations making interactions more therapeutically flexible, with new safety challenges to address. The direction is toward AI as a meaningful amplifier of mental health care capacity — not replacing therapists, but helping close the access gap.

• Has AGI Already Arrived?