What Is AI Hallucination? Real Examples, Rates by Model, and How to Stop It

What Is AI Hallucination?

AI hallucination occurs when an artificial intelligence model generates information that sounds authoritative and confident but is factually incorrect, entirely fabricated, or unsupported by any real data source. The term borrows from human psychology — just as a person experiencing a hallucination perceives something that isn't there, an AI model "perceives" and presents information that doesn't exist in reality.

What makes AI hallucinations uniquely dangerous is their structural quality. A hallucinated response is not garbled text or obviously broken output. It is grammatically perfect, contextually appropriate, and delivered with the same tone and confidence as a verified fact. The AI doesn't flag its own uncertainty. It doesn't add a disclaimer. It presents fiction as truth — seamlessly embedded within otherwise accurate information.

Consider this: if a model correctly explains the first three paragraphs of a legal analysis and then fabricates a case citation in the fourth paragraph, how would you know? The fabricated citation will include a plausible case name, a realistic court, a believable year, and a coherent legal holding. It will look indistinguishable from a real citation unless you independently verify it.

This is the core problem with AI hallucinations — they exploit the trust that accurate surrounding content creates. Users naturally extend the credibility of correct information to nearby claims that happen to be fabricated. The better the model is overall, the more dangerous its hallucinations become, because users have even more reason to trust its outputs.

AI hallucinations are not edge cases. They happen regularly, across all major models, in every domain. Understanding why they occur, how often they occur, and how to detect them is not optional for anyone relying on AI for professional work.

Why Do AI Models Hallucinate?

AI hallucinations are not bugs that can be patched with a software update. They are architectural consequences of how large language models work. Understanding the root causes is essential for developing effective prevention strategies.

Token Prediction vs. Truth

Large language models (LLMs) like GPT-5, Claude, and Gemini do not "know" things the way humans do. They predict the most statistically probable next token (word or word-fragment) based on the patterns in their training data. When you ask a model a question, it is not retrieving a stored fact — it is generating the sequence of words that is most likely to follow your prompt. Most of the time, the most probable sequence happens to be factually correct, because the model's training data contained accurate information about the topic. But probability and truth are not the same thing. When the model encounters a topic where the most probable word sequence leads to a plausible-sounding but incorrect statement, it follows the probability — not the truth.

Training Data Gaps

No model has been trained on all human knowledge, and training data has a cutoff date. When a model encounters a query about a topic that was sparsely covered in its training data — a niche legal precedent, an obscure medical interaction, a recent event — it extrapolates. It fills the gap with a plausible-sounding response constructed from patterns in related topics. The result is a hallucination that reads like a fact but was actually interpolated from adjacent knowledge.

Confidence Miscalibration

Humans have an internal sense of confidence — we know the difference between "I'm sure about this" and "I'm guessing." Current LLMs largely lack this metacognitive capability. A model cannot distinguish between a response it is generating from well-established training data and one it is fabricating from sparse or conflicting signals. Both are delivered with identical confidence and formatting. This is why hallucinated answers feel so convincing — the model itself doesn't "know" it's hallucinating.

Context Window Limitations

Even with today's expanded context windows (128K to 1M+ tokens), models can lose track of important information in long conversations. Key facts established early in a dialogue may be diluted by subsequent exchanges. The model may contradict earlier statements, forget constraints you specified, or introduce fabricated information to fill gaps created by lost context. This is particularly problematic in complex professional workflows where accuracy depends on maintaining coherent reasoning across many turns.

The Sycophancy Problem

LLMs are trained with reinforcement learning from human feedback (RLHF), which optimizes for responses that users rate as helpful and satisfying. This creates a subtle but dangerous incentive: when a user asserts something incorrect, the model may agree with the user rather than correct them — because agreeing is more likely to receive positive feedback. This "sycophancy bias" means that leading questions, incorrect premises, and flawed assumptions in prompts can steer models toward hallucinated confirmations rather than accurate corrections.

Real AI Hallucination Examples That Cost Real Money

AI hallucination is not a theoretical risk. It has already caused documented financial losses, professional sanctions, and reputational damage across multiple industries. These are not cherry-picked failures — they represent systemic vulnerabilities that affect every organization using AI without verification safeguards.

Legal: Mata v. Avianca (2023) — The Case That Changed Everything

In what became the most famous AI hallucination case in history, two New York City lawyers — Steven Schwartz and Peter LoDuca — submitted a legal brief in Mata v. Avianca Airlines that contained at least six fabricated case citations generated by ChatGPT. The cited cases — including Varghese v. China Southern Airlines and Shaboon v. Egyptair — did not exist. When the presiding judge, P. Kevin Castel of the Southern District of New York, asked the lawyers to produce the cited decisions, they asked ChatGPT to verify its own citations — and ChatGPT confirmed they were real. The judge sanctioned both lawyers, imposing a $5,000 fine and requiring them to notify every judge falsely cited. The case became a watershed moment for AI in legal practice, prompting courts across the country to issue standing orders requiring lawyers to disclose AI use in filings.

For legal professionals evaluating AI tools, see our detailed guide on AI for lawyers: why verification matters.

Medical: Fabricated Clinical Studies and Drug Interactions

The stakes in healthcare are even higher. A Stanford HAI study found that legal AI tools hallucinated in approximately 1 out of every 6 responses — and the pattern extends to medical AI. Researchers have documented cases of AI models fabricating clinical study citations, inventing drug interaction warnings for combinations that have no documented interaction, and generating treatment protocols that reference non-existent clinical trials. In one widely reported case, a medical AI suggested a drug dosage that, if followed, would have been potentially harmful. The AI generated its recommendation with complete confidence and no caveats.

The danger is compounded by the fact that medical hallucinations often contain a mixture of correct and fabricated information — accurate drug names with incorrect dosages, real conditions with fabricated prevalence statistics, legitimate treatment approaches with non-existent supporting studies.

Financial: Fabricated Market Data and Statistics

AI models routinely generate fabricated statistics in financial and market research contexts. This includes fake revenue figures attributed to real companies, fabricated market size estimates with specific dollar amounts and growth rates, non-existent survey results with precise percentages, and fake analyst quotes attributed to real people at real firms. These hallucinations are particularly insidious because they include the specific numerical precision and source attributions that make them indistinguishable from real data without manual verification against primary sources.

Coding: Hallucinated APIs and Package Confusion Attacks

In software development, AI models regularly hallucinate API methods, library functions, and package names that do not exist. A developer following an AI-generated code example may call a method that was never part of the referenced library, import a package that doesn't exist, or use a function signature that looks correct but has fabricated parameters. More alarmingly, security researchers have documented "package confusion attacks" where attackers create malicious packages with the names that AI models commonly hallucinate. When developers follow AI-generated code that references a hallucinated package name, they may unknowingly install malware that an attacker published under that fabricated name.

Our Own Findings: A 5.8-Point Accuracy Spread Across 32 Models

In Search Umbrella's Trust Score evaluations, we tested 32 AI models across 2,637 real-world queries spanning legal, medical, business, technical, and general knowledge domains. The results revealed a striking variance: factual accuracy scores ranged from 6.0 to 8.9 out of 10 — a 5.8-point spread on identical questions. This means two models answering the same question could differ by nearly 60% on our accuracy scale. The worst-performing models scored as low as 6.0/10 on our factual accuracy metric — meaning roughly 4 out of 10 factual claims required correction or verification.

The implication is clear: which model you use matters enormously for accuracy. And the only way to know which model is most reliable for a specific question is to compare multiple models simultaneously. This is the foundational principle behind Search Umbrella's unified LLM search approach.

AI Hallucination Rates by Model (2026)

We measured factual accuracy across 32 AI models using 2,637 real-world queries as part of Search Umbrella's Trust Score evaluation framework. Factual Accuracy is one of seven metrics in the composite Trust Score, scored on a 0-to-10 scale where higher scores indicate fewer hallucinations and more reliable factual outputs.

The following table shows the top 20 models ranked by factual accuracy. To compare AI models side by side or see the full leaderboard, visit howismyai.com.

Rank	Model	Provider	Factual Accuracy (0-10)	Evaluations
1	GPT-5 Mini	OpenAI	8.92	26
2	GPT-5	OpenAI	8.82	60
3	Gemini 2.5 Pro	Google	8.78	16
4	GPT-5.2	OpenAI	8.54	62
5	GPT-5.1	OpenAI	8.43	23
6	GPT-5.2 (Thinking)	OpenAI	8.39	185
7	GPT-4.1	OpenAI	8.34	50
8	Gemini 3 Flash	Google	8.16	51
9	GPT-5.1 (Thinking)	OpenAI	8.05	79
10	Gemini 2.5 Flash	Google	7.78	41
11	Claude Sonnet 4.5	Anthropic	7.78	651
12	Grok 4	xAI	7.72	39
13	GPT-4.1 Nano	OpenAI	7.67	21
14	Gemini 3 Pro	Google	7.64	479
15	Claude Opus 4.6	Anthropic	7.57	117
16	Grok 4 (Reasoning)	xAI	7.51	142
17	GPT-4o	OpenAI	7.26	27
18	GPT-5 (Generic)	OpenAI	7.21	41
19	Sonar Pro	Perplexity	7.01	67
20	Grok 4.1 (Reasoning)	xAI	6.96	99

Data from howismyai.com Trust Score evaluations, December 2025 through February 2026. Factual Accuracy is one of seven metrics in the composite Trust Score. Sample sizes vary — models with fewer than 30 evaluations should be interpreted with caution.

Key insight: The 5.8-point spread between the highest and lowest scores means two models answering the exact same question can differ by nearly 60% on our accuracy scale. This is why relying on a single model is risky — and why cross-model verification is the most practical defense against hallucinations available today.

Notice that some of the highest-ranked models have relatively small sample sizes (GPT-5 Mini with 26 evaluations, Gemini 2.5 Pro with 16). Models with larger evaluation pools — like Claude Sonnet 4.5 (651 evaluations) and Gemini 3 Pro (479) — provide more statistically robust accuracy estimates. When selecting a model, consider both the score and the confidence level that the sample size supports.

For a head-to-head ChatGPT vs Claude comparison that goes beyond factual accuracy into writing quality, reasoning, and professional use cases, see our dedicated comparison guide.

Hallucination Rates Vary by Domain

Factual accuracy is not just a model-level characteristic — it varies dramatically by the type of question you ask. Our Trust Score evaluations categorized all 2,637 queries by domain, revealing striking differences in how reliably models perform across different knowledge areas.

Domain	Average Trust Score	Total Evaluations	Hallucination Risk
Coding	8.61	713	Lowest
Business	8.30	134	Low-Moderate
Legal	8.30	44	Low-Moderate*
Technical	8.27	433	Low-Moderate
Creative	8.25	53	Low-Moderate
Personal	8.30	37	Low-Moderate
General	7.72	1,116	Moderate
Research	7.39	107	Highest

*Legal domain has only 44 evaluations — interpret with caution. Data from howismyai.com Trust Score evaluations, December 2025 through February 2026.

Key insight: Research queries have the lowest accuracy scores — exactly the domain where hallucinations are most dangerous. When professionals use AI for original research — literature reviews, fact-checking claims, finding supporting evidence — they encounter the highest hallucination rates in precisely the context where they need the highest accuracy.

Coding queries score highest, likely because code is verifiable: it either compiles and runs correctly, or it doesn't. The feedback loop between training data and correctness is tighter for code than for subjective or factual claims. General knowledge queries, which represent the largest evaluation pool at 1,116 tests, show moderate accuracy — consistent with the idea that broad, open-ended questions create more opportunities for hallucination than tightly scoped technical problems.

The Legal domain score of 8.30 may appear reassuring, but the small sample size (44 evaluations) demands caution. More importantly, the consequences of hallucination in legal contexts are categorically more severe than in general knowledge. An 8.30 accuracy score in a domain where a single fabricated citation can result in professional sanctions is not adequate for professional reliance without additional verification. For more on this, see our analysis of why verification matters for legal AI.

How to Detect and Prevent AI Hallucinations

AI hallucinations cannot be eliminated entirely — they are inherent to how current language models work. But they can be detected, reduced, and managed through a layered verification approach. The following five methods, used in combination, represent the current best practice for organizations that depend on AI accuracy.

Why Cross-Model Verification Is the Practical Standard

RAG, prompt engineering, and chain-of-thought verification are all valuable techniques — but they all require technical expertise, custom infrastructure, or significant changes to how users interact with AI. Cross-model verification is the only approach that works for any user, on any query, without any technical setup.

The principle is simple: hallucinations are model-specific. A fabricated case citation generated by GPT-5 will not be confirmed by Claude, Gemini, and Grok. A fake statistic invented by one model will be contradicted by others that have different training data and different probability distributions. By comparing outputs across multiple independent models, you create a natural error-detection system that catches hallucinations that would be invisible in any single model's output.

This is the same principle that makes peer review work in science, second opinions work in medicine, and multi-source verification work in journalism. It is not new — but applying it to AI at scale is. Search Umbrella is building the infrastructure to make cross-model verification as easy as typing a single query.

For those interested in how this approach works under the hood, read our deep dive on how unified LLM search works and how it relates to the broader concept of Generative Engine Optimization (GEO).

If you're evaluating ChatGPT alternatives that verify answers rather than just generate them, cross-model verification should be your primary evaluation criterion. A model that is 5% more accurate but unverified is less valuable than a verified synthesis across multiple models.

Frequently Asked Questions

What is AI hallucination?

AI hallucination is when an artificial intelligence model generates information that sounds authoritative and confident but is factually wrong, fabricated, or unsupported by any real data source. Unlike random errors or garbled text, hallucinations are structurally coherent, grammatically perfect, and contextually plausible — making them extremely difficult for users to detect without independent verification. The term is now widely used across the AI industry and was even named the Cambridge Dictionary's Word of the Year in 2023.

What causes AI hallucination?

AI hallucinations have five primary causes: (1) LLMs predict probable word sequences rather than verified facts. (2) Training data has gaps — models extrapolate when knowledge is sparse. (3) Models cannot distinguish between high-confidence knowledge and low-confidence guesses. (4) Context can be lost in long conversations. (5) RLHF training creates a "sycophancy bias" where models prefer to agree with users rather than correct them. These causes are architectural, not bugs — they cannot be fully eliminated with current technology.

How common are AI hallucinations?

Hallucination frequency varies dramatically by model and domain. In our testing of 32 AI models across 2,637 real-world queries, factual accuracy scores ranged from 6.0 to 8.9 out of 10. Research queries had the lowest accuracy (7.39 average trust score), while coding queries had the highest (8.61). Independent studies report hallucination rates ranging from 1.5% to over 27% depending on the model, task type, and evaluation methodology. The most reliable approach is to verify specific claims through cross-model comparison rather than trusting any single model's general accuracy rate.

Which AI model hallucinates the least?

In our Trust Score evaluations (December 2025 through February 2026), GPT-5 Mini scored the highest factual accuracy at 8.92/10, followed by GPT-5 at 8.82 and Gemini 2.5 Pro at 8.78. However, sample sizes matter significantly — GPT-5 Mini had only 26 evaluations while Claude Sonnet 4.5 had 651. Models with larger evaluation pools provide more statistically reliable accuracy estimates. See the full hallucination rate rankings for the complete dataset, or explore our detailed hallucination rate analysis for methodology and context.

Can AI hallucinations be prevented?

AI hallucinations cannot be fully prevented with current technology, but they can be significantly reduced through layered verification: cross-model verification (running the same query through multiple models and checking for consensus), RAG (grounding responses in verified documents), prompt engineering (instructing models to cite sources and express uncertainty), chain-of-thought verification (asking models to show reasoning step-by-step), and human expert review for high-stakes decisions. Of these, cross-model verification is the most accessible and broadly effective — Search Umbrella automates it across 8+ models with a Trust Score for every response.

What is the most famous AI hallucination example?

The most famous AI hallucination case is Mata v. Avianca (2023), in which New York City lawyers Steven Schwartz and Peter LoDuca submitted a legal brief containing at least six fabricated case citations generated by ChatGPT. Cases including Varghese v. China Southern Airlines and Shaboon v. Egyptair did not exist. When asked to verify, ChatGPT confirmed its own fabricated citations were real. Judge P. Kevin Castel sanctioned both lawyers with a $5,000 fine and a requirement to notify every judge falsely cited. The case triggered standing orders in courts across the United States requiring disclosure of AI use in legal filings.