Who Hallucinates More: ChatGPT or Claude?

AI hallucination is when a language model produces confident, fluent output that is factually wrong — a citation that doesn't exist, a date that never happened, a quote no one said. The term is borrowed loosely from psychiatry, where hallucination means perceiving something that isn't there. In practice, LLM hallucinations look less like delusions and more like plausible-sounding autocomplete: the model generates the statistically likely continuation of a sentence, not a grounded fact lookup. The critical word is "confident" — hallucinations are dangerous not because models are wrong, but because they are wrong without signaling any uncertainty. A practitioner's real concern is not hallucination frequency but hallucination detectability: a model that hallucinates rarely but never hedges is far more dangerous in production than one that hallucinates often and flags it.

AI chatbots hallucinate because they are trained to predict the most plausible next token, not to retrieve verified facts from a ground-truth database. The architecture is fundamentally generative — the model produces text that fits the statistical patterns in its training corpus, and sometimes those patterns lead it to fill gaps with invented specifics. Three compounding factors make hallucination worse: sparse coverage of a topic in training data (the model extrapolates), conflicting information in the corpus (the model blends), and RLHF reward signals that favor fluent, complete-sounding outputs over hedged ones (the model stops saying "I'm not sure"). The reason ChatGPT and Claude hallucinate at different rates on different tasks is not architecture alone — it is which of these three failure modes each system's training most aggressively corrects for. If your task exposes sparse training coverage (niche domain knowledge, recent events), neither model can save you without grounded retrieval.

Confabulation in AI is the specific pattern where a model fills a knowledge gap with invented-but-plausible detail rather than refusing or hedging — the model "connects the dots" that were never actually there. The clinical term comes from neurology, where patients with certain memory disorders produce false memories that feel entirely real to them. In LLMs, confabulation is the mechanism behind the most dangerous class of hallucinations: not random nonsense but well-constructed fabrications — a fake paper with a real author's name, a plausible-sounding legal citation, a drug dosage derived by averaging nearby real figures. The distinction matters for tooling: hallucination detectors that look for low confidence scores will often miss confabulation, because the model's internal confidence on a confabulated output can be high. Grounding against primary sources — not just asking the model to self-check — is the only reliable counter.

Hallucination is the broad category; confabulation is the specific failure mode where the model invents plausible gap-fills rather than flagging its own uncertainty. All confabulation is hallucination, but not all hallucination is confabulation — a model that confidently states a wrong date is hallucinating, but it isn't necessarily confabulating if the error traces to a corrupted training example rather than a gap-bridging inference. The distinction changes what interventions work: suppressing confabulation requires training models to recognize the edges of their own knowledge and refuse at those boundaries (which is what Constitutional AI's self-critique loop does for Claude). Suppressing hallucination more broadly requires grounding — retrieval-augmented generation, citation enforcement, source verification. Practitioners who use the words interchangeably will apply the wrong fix.

No — hallucination is a failure of knowledge, not a failure of intent, which means the usual remedies for dishonesty (adversarial red-teaming, filtering, policy enforcement) don't reduce it. A lying agent knows the truth and conceals it; a hallucinating model has no ground-truth representation to conceal — it generates the output that fits the learned distribution, whether that output is accurate or not. This distinction is not just philosophical. Treating hallucination as lying leads organizations to apply trust-and-safety interventions (content moderation, output filtering) rather than epistemic interventions (grounding, uncertainty calibration, retrieval). The more practically damaging confusion is the reverse: treating hallucination as a fixable "bad behavior" that fine-tuning will eventually eliminate, rather than as a structural property of generative models that requires architectural solutions.

ChatGPT makes things up because its RLHF training consistently rewarded fluent, complete-sounding answers — and human raters often cannot tell in the moment whether a specific claim is true. When the model encounters a query at the edge of its training knowledge, it faces two options: produce a hedged, incomplete answer (which RLHF raters historically penalized as unhelpful) or produce a fluent, confident-sounding answer that fills the gap (which raters often rewarded as useful). Over millions of training examples, that signal compounds: the model learns that confident gap-filling is the preferred behavior. OpenAI's release notes for GPT-5.5 Instant specifically cite "reduces hallucination in sensitive areas such as law, medicine, and finance" as a named improvement — which is an implicit acknowledgment that prior versions were not calibrated to refuse when uncertain. The fix is not better knowledge; it is better uncertainty signaling.

Neither model consistently hallucinates more — the winner changes based on the task type, benchmark methodology, and which model version is being measured. On BullshitBench v2, Claude Sonnet 4.6 hits a 3% hallucination rate with a 91% detection rate, while OpenAI GPT models are "stuck in the 55–65% range" for detection. On Vectara's harder enterprise dataset (February 2026), GPT-4.1 scores 5.6% versus Claude Sonnet 4.6 at 10.6% — a reversal. On AA-Omniscience, Claude Opus 4.1 achieves 0% hallucination (via refusal), while GPT-5.5 reaches 86% error on the same benchmark. The honest answer for practitioners: Claude tends to outperform on tasks requiring uncertainty calibration and open-recall; ChatGPT tends to outperform on grounded tasks with source material present. Your use case determines the verdict.

Claude hallucinates less than ChatGPT on open-recall and uncertainty-calibration benchmarks, but GPT models can outperform Claude on grounded generation tasks where source material is provided. On the Vectara HHEM original dataset (April 2025), GPT-5 scores 1.4% versus Claude-3.7-Sonnet at 4.4% — ChatGPT wins. On BullshitBench v2, Claude Sonnet 4.6 scores 3% with a 91% detection rate — Claude wins. On AA-Omniscience, Claude Opus 4.1 achieves 0% hallucination via confident refusal — Claude wins decisively. The most useful reframe: Claude tends to hallucinate less on tasks where "I don't know" is an acceptable output; ChatGPT can score lower on structured summarization tasks where the source material bounds the answer space.

GPT models score higher on grounded factual accuracy when source material is present; Claude scores higher on calibration — knowing when not to answer. On FACTS Overall Scores (grounded generation), GPT-5 scores 61.8 versus Claude Opus 4.5 at 51.3. On AA-Omniscience, Claude Opus 4.1 achieves 0% hallucination while GPT-5.5 reaches 86% error. These are not contradictions — they measure different things. FACTS rewards producing correct answers given a source; AA-Omniscience rewards refusing answers when knowledge is uncertain. Accuracy in a production system means both: getting the answer right when you have the source, and refusing when you don't. No single model currently dominates both dimensions simultaneously.

ChatGPT's hallucination rate in 2026 ranges from 1.4% on Vectara's original RAG benchmark to 86% on AA-Omniscience's domain-knowledge open-recall test — the same model, different methodologies. On BullshitBench v2, OpenAI GPT models are "stuck in the 55–65% range" for hallucination detection. GPT-5 with thinking mode achieves 1.6% on HealthBench (medical domain). O3 hits 51% hallucination on SimpleQA; o4-mini reaches 79% on PersonQA. The number you see in any article reflects the benchmark used, not a universal accuracy property. The most applicable figure depends on your task: if you are doing RAG summarization, Vectara's 1.4% is relevant; if you are asking ChatGPT to recall domain-specific facts without source material, the AA-Omniscience figure is the honest baseline.

Claude's hallucination rate in 2026 spans from 0% (Claude Opus 4.1 on AA-Omniscience, via refusal) to 58% (Claude Opus 4.5 on the same benchmark when not configured to refuse) — a range that makes any single number misleading. On BullshitBench v2, Claude Sonnet 4.6 hits 3% with a 91% detection rate, making it the strongest performer in that benchmark class. On Vectara's enterprise dataset (February 2026), Claude Sonnet 4.6 scores 10.6% and Claude Opus 4.6 scores 12.2%. The spread is explained by task type: Claude's Constitutional AI training produces strong refusal behavior on uncertain factual questions, which collapses the hallucination rate on benchmarks that reward "I don't know" responses and inflates it on benchmarks that penalize non-answers.

Gemini-2.0-Flash-001 holds the lowest published Vectara HHEM score at 0.7% on the original dataset — but that benchmark measures factual consistency in RAG summarization, not open-ended recall. On open-recall benchmarks, Claude Opus 4.1 achieves 0% on AA-Omniscience by refusing uncertain questions, while o3-mini-high scores 0.8% on Vectara. The "lowest hallucination rate" title changes with every benchmark and model release cycle; the more useful question is which model has the lowest hallucination rate on your specific task class. For enterprise RAG pipelines with provided source material, GPT-4.1 at 5.6% on the harder Vectara dataset is currently competitive. For open-domain factual recall with uncertainty, Claude's refusal behavior produces the lowest confirmed error rate.

AI hallucination is measured by comparing model outputs against a verified ground-truth set and scoring the proportion of confident claims that are factually wrong — but the ground-truth set, task type, and scoring rules vary so widely across benchmarks that the resulting numbers are rarely comparable. Three methodology families dominate: RAG consistency tests (Vectara HHEM measures whether a summary stays faithful to the source document), factual recall tests (SimpleQA, PersonQA ask the model open questions with known correct answers), and calibration tests (AA-Omniscience scores how often a model produces wrong answers on questions it should refuse). The same model can score in the top tier on one family and bottom tier on another. Before citing a hallucination rate, the practitioner question is: what task type does this benchmark represent, and does that match what I'm actually asking the model to do?

Constitutional AI is Anthropic's training methodology where Claude critiques and revises its own outputs against a set of principles — and yes, it reduces a specific class of hallucination: confabulation driven by overconfidence. The core mechanism is a self-critique loop: at training time, Claude is prompted to evaluate its own responses against a constitution of principles (including honesty norms) and revise outputs that violate them. Over millions of examples, this trains the model to flag uncertainty rather than elaborate plausibly over it. What it does not do is give Claude better knowledge — it makes the model more likely to output "I'm not sure" or refuse at the boundary of its knowledge. The result is measurably lower hallucination rates on open-recall benchmarks and, as a side effect, higher refusal rates on tasks where humans expect confident answers. The "safer model" framing is accurate but incomplete: Constitutional AI reduces dangerous confabulation, not all incorrect outputs.

Claude hallucinates less than ChatGPT on uncertainty-sensitive tasks because Constitutional AI's self-critique training penalized overconfident outputs at the architectural level — not because Claude has better underlying knowledge. ChatGPT's RLHF training was tuned on human preference ratings, and human raters consistently prefer confident, complete-sounding answers to hedged, partial ones — even when the hedged answer is more accurate. That preference signal, applied at scale, teaches the model to fill gaps confidently. Constitutional AI's revision loop applies a different signal: outputs that violate honesty norms (overconfident claims under uncertainty) are scored negatively by the model itself and revised. On benchmarks like AA-Omniscience, this difference is dramatic: Claude Opus 4.1 achieves 0% hallucination by refusing uncertain questions; GPT-5.5 attempts the same questions and produces 86% error. The practical implication is that Claude's advantage narrows or reverses when the task context provides source material that bounds the answer — because grounding reduces the need for uncertainty calibration.

Claude says "I don't know" more than ChatGPT because uncertainty admission was a first-class design goal in Constitutional AI, not an afterthought in RLHF fine-tuning. Anthropic explicitly trained Claude to identify the edges of its own knowledge and signal them rather than bridge them. The constitutional principle "prefer accurate uncertainty estimates over confident wrong answers" was applied via the self-critique loop — meaning at training time, Claude learned to score its own overconfident outputs negatively. ChatGPT's training did the opposite: human raters penalized partial or hedged answers as unhelpful, pushing the model toward confident completeness. For practitioners, the implication is that Claude's "I don't know" is a calibration signal worth respecting — it correlates with genuine knowledge boundaries. ChatGPT's confident answers do not carry the same calibration signal and require independent verification more often.

Claude admits uncertainty more than ChatGPT because its training reward function directly penalized overconfidence, while ChatGPT's reward function indirectly penalized uncertainty by preferring fluent, complete-sounding outputs. These are mirror-image training problems with mirror-image results. At the architectural level, both models have the same epistemic limitation: they cannot know what they do not know. What differs is how each handles that edge. Claude's Constitutional AI self-critique loop was designed to surface that edge and output it. ChatGPT's RLHF fine-tuning learned to smooth over it. The consequence for production use is that when Claude expresses uncertainty, it is more likely to be a genuine signal. When ChatGPT expresses certainty, it is less likely to be a reliable signal than the confident phrasing suggests. This asymmetry is the single most important behavioral difference between the two systems for high-stakes use cases.

Yes — Claude is trained to refuse or heavily hedge questions at the boundary of its knowledge, and this behavior is measurable: on AA-Omniscience, Claude Opus 4.1 achieves 0% hallucination by refusing uncertain domain-knowledge questions rather than attempting them. This refusal behavior is not a safety filter applied after generation — it is a trained output preference baked into the model via Constitutional AI's self-critique loop. The practical consequence is two-sided: Claude produces fewer confident wrong answers than ChatGPT, but it also produces more non-answers on questions where an attempt — even an imperfect one — would be useful. For tasks where a partial answer is better than no answer (brainstorming, hypothesis generation, exploratory research), ChatGPT's higher attempt rate is a feature. For tasks where a wrong answer causes real harm (legal, medical, compliance), Claude's refusal behavior is the more defensible default.

The most documented ChatGPT hallucinations include fabricated legal case citations presented to federal courts, invented academic papers with real author names, and confident wrong drug dosages in medical queries. The legal citation failures are the best-documented: in the Mata v. Avianca case, a New York attorney submitted a brief citing multiple cases that ChatGPT invented wholesale — cases that did not exist anywhere in the legal record. Academic hallucinations are structurally similar: ChatGPT generates plausible-sounding paper titles, journal names, and DOIs that pass casual verification because all the component elements (author names, journal names, topic keywords) are real — only the assembled paper is fabricated. The pattern in all major cases is the same: ChatGPT produces hallucinations that are specifically designed, by the training distribution, to pass the first verification step a non-expert would apply.

Yes — the most consequential documented case is Mata v. Avianca, where an attorney used ChatGPT to research case law and submitted a brief citing multiple cases that did not exist, resulting in federal court sanctions. The cases ChatGPT generated were plausible: they had realistic docket numbers, party names consistent with real aviation litigation, and summaries that read as coherent legal precedent. None of them could be located by opposing counsel or the court because they were entirely fabricated. The attorney was sanctioned for failing to verify the citations. What makes this case significant beyond its notoriety is the mechanism: ChatGPT did not produce random nonsense — it confabulated, generating outputs that fit the expected form of real legal citations so closely that a trained attorney did not catch them on first read. This is the confabulation failure mode operating at the level of maximum real-world cost.

Yes — in the Mata v. Avianca case, ChatGPT generated at least six fake legal cases that were submitted to a federal court as real precedent, making this the first major documented instance of LLM hallucination causing legal sanctions against a practicing attorney. The fabricated cases had realistic-looking citations: Varghese v. China Southern Airlines, Shaboon v. Egyptair, Zicherman v. Korean Air Lines, and others — all plausible-sounding aviation negligence precedents. When the court asked for copies of the actual decisions, the attorney could not produce them because they did not exist. The episode became a landmark not just for AI liability but for the broader question of what "verification" means when a model's hallucinations are structurally indistinguishable from real citations to a non-expert reader. Neither Claude nor any other model has a documented comparable case — but the mechanism exists in all models that generate legal text without grounding.

In Mata v. Avianca, a personal injury lawsuit filed in the Southern District of New York, attorney Steven Schwartz used ChatGPT to research aviation negligence precedents and submitted a court brief citing six cases that ChatGPT had fabricated. When opposing counsel could not locate the cited cases, the court ordered Schwartz to produce the actual decisions. He could not — the cases existed only in ChatGPT's output. The court sanctioned Schwartz and his firm. Schwartz's defense was that he was unfamiliar with ChatGPT's tendency to generate false information; the court found that reliance on an AI tool without verification constituted professional negligence. The case is now the canonical example cited in AI liability discussions because it makes concrete the abstract warning that LLM hallucinations have real-world costs — and because the mechanism was confabulation, not noise: the fake cases were structurally indistinguishable from real ones.

Yes — ChatGPT can hallucinate medical information, and the risk is highest in exactly the scenarios where clinicians are most likely to use it: rare conditions, drug-drug interactions, and off-label dosages that are underrepresented in training data. GPT-4o scores 15.8% hallucination on HealthBench, a medical domain benchmark; GPT-5 with thinking mode reduces this to 1.6%, but the improvement is conditional on the thinking mode being enabled and the query being within the benchmark's scope. The practical risk for medical use is not just the headline hallucination rate — it is the confabulation pattern: ChatGPT generating specific-sounding dosages or protocol details that are plausible but wrong, in the confident register that clinical notes require. The consensus across medical AI research is that no current LLM should be used as a primary information source for treatment decisions without RAG grounding against validated clinical databases.

ChatGPT hallucinates on academic citations at a notably higher rate than on general-text tasks because academic citations combine several conditions that maximize confabulation risk: sparse training coverage of specific papers, high structural regularity (author, title, journal, year, DOI), and user verification behavior that rarely extends beyond checking the format. The model has learned that citations follow predictable patterns. When asked for a citation it does not have in training, it generates a citation that fits those patterns — assembling a plausible author name, a realistic journal, and a plausible year around a fabricated paper. Dedicated academic search integrations (like ChatGPT's search tool when enabled) significantly reduce citation hallucination by grounding against live databases. Without grounding, treating any LLM-generated academic citation as provisional and verifying it against Google Scholar, CrossRef, or a DOI resolver is not optional — it is the baseline standard of care.

You cannot stop ChatGPT from hallucinating entirely, but the four techniques that reliably reduce it are: retrieval-augmented generation with verified sources, chain-of-thought prompting with explicit uncertainty flagging, citation enforcement in the system prompt, and output verification against primary sources before use. RAG is the highest-leverage intervention: grounding responses against a controlled, verified document set eliminates the knowledge-gap confabulation that produces most dangerous hallucinations. Chain-of-thought prompting ("explain your reasoning step by step, and flag any step where you are less than certain") forces the model to surface uncertainty it would otherwise paper over. Adding "if you are unsure, say so explicitly rather than guessing" to system prompts has measurable effect on calibration. None of these eliminate the problem — they reduce it. For high-stakes outputs (legal, medical, financial), human verification against primary sources remains the standard. The tools help; they do not replace verification.

The prompt engineering techniques with the strongest documented effect on hallucination reduction are: explicit uncertainty instructions, step-by-step reasoning requirements, role-scoping, and source-citation enforcement — applied together, not independently. Explicit uncertainty instructions ("say 'I don't know' rather than guessing") improve calibration on both Claude and ChatGPT because both models are capable of signaling uncertainty; they need permission to do it. Step-by-step reasoning forces the model to commit to intermediate claims that can be checked, catching confabulation earlier in the chain. Role-scoping ("you are a fact-checker; do not include any claim you cannot source") narrows the output distribution toward the verified. Source citation enforcement ("provide a source URL for every factual claim") creates a verification trail. The compounding insight most engineers miss: these techniques are not additive linearly — models trained with uncertainty signals (like Claude) show larger improvements from uncertainty prompts than models trained against them (like older ChatGPT versions).

Asking ChatGPT to cite sources reduces the frequency of unverifiable claims in the output, but it does not eliminate fabricated citations — and a fabricated citation with a plausible URL is harder to catch than a claim with no citation at all. The mechanism is real: source-citation prompting shifts the model toward outputs where citation is possible, which correlates with better-grounded claims. But ChatGPT can and does generate plausible-looking DOIs, arXiv IDs, and URLs that resolve to nothing. The citation requirement creates a false confidence layer — the output looks verified when it isn't. The correct workflow is source-citation prompting plus independent verification of every cited source before use. For enterprise pipelines, automated link-checking (do all cited URLs actually resolve?) is the minimum; content verification (does the cited source actually say what the LLM claims it says?) is the standard that eliminates the fabricated-citation failure mode.

Retrieval-augmented generation (RAG) is an architecture that grounds LLM outputs by retrieving relevant documents from a verified source set and providing them as context — and it is currently the most effective single intervention for reducing hallucination in production systems. Instead of asking a model to recall facts from training, a RAG system retrieves the relevant passage from a controlled document store and asks the model to summarize or synthesize it. On Vectara's HHEM benchmark, which specifically tests this summarization-from-source behavior, even older model versions achieve hallucination rates below 5% — because the knowledge gap confabulation mechanism is eliminated when the answer exists in the provided context. What RAG does not stop: hallucinations that occur when the retrieved document does not contain the answer and the model interpolates anyway, and hallucinations in the retrieval step itself (if a semantic search retrieves the wrong document). RAG reduces hallucination dramatically in bounded domains; it is not a universal cure for open-domain queries.

Claude is more trustworthy for tasks where calibrated uncertainty is the primary requirement; ChatGPT is more trustworthy for tasks where producing an answer — even an imperfect one — is the primary requirement. Trust is not a single dimension. On calibration trust (does the model's expressed confidence correlate with its actual accuracy?), Claude leads: Constitutional AI's uncertainty training produces hedges that track genuine knowledge gaps. On coverage trust (will the model attempt the question rather than refuse?), ChatGPT leads: RLHF training produces higher attempt rates on hard questions. The user perception data reflects calibration trust: 62% of verified ChatGPT (GPT-4/4o) users report "occasional confident inaccuracies" versus 24% of Claude 3.5 users. For practitioners, the framework is: trust Claude more when a wrong answer is worse than no answer; trust ChatGPT more when a partial answer is better than a refusal.

For high-stakes tasks where a wrong answer has irreversible consequences, Claude's refusal-calibrated behavior makes it the safer default — but safe use of either model requires human verification against primary sources, not model selection alone. Claude's design advantage in high-stakes contexts is the refusal signal: when Claude says it is uncertain, that signal has been trained to track real knowledge boundaries. ChatGPT's confident outputs in the same situations are less reliably calibrated. However, "safer model" does not mean "reliable without verification" — Claude Opus 4.5 scores 58% hallucination on AA-Omniscience when not configured to refuse, and Claude Opus 4.7 scores 36%. The practical standard for high-stakes work is: use Claude for the calibration signal, ground with RAG against verified sources, and treat any AI-generated factual claim in a legal, medical, or financial context as provisional until independently confirmed.

For enterprise RAG pipelines where source documents are provided, GPT-4.1 currently outperforms Claude on Vectara's harder enterprise dataset (5.6% vs. Claude Sonnet 4.6's 10.6%) — but for enterprise use cases requiring open-domain knowledge retrieval or strict uncertainty signaling, Claude's refusal calibration produces fewer dangerous confident errors. The enterprise use case is split along the same task-type boundary that governs all hallucination comparisons: grounded generation with provided source material favors GPT models; open-domain recall with uncertainty requirements favors Claude. For enterprise deployments at the highest risk level (legal, medical, compliance), the architecture recommendation from available benchmark data is: pair Claude with a RAG pipeline, use Claude's uncertainty signal as a flag for human review, and verify any output where the model expresses high confidence without a cited source. The combination outperforms either model used alone.

No — neither ChatGPT nor any current LLM should be trusted as a primary source for medical or legal advice without independent verification against authoritative primary sources, and the documented failure cases make the risk concrete. In the Mata v. Avianca case, ChatGPT fabricated legal citations that passed initial attorney review, resulting in federal court sanctions — the most expensive hallucination failure mode documented in legal practice. On HealthBench, GPT-4o hallucinates medical information at a 15.8% rate; GPT-5 with thinking mode reduces this to 1.6%, but that reduction depends on task-specific configurations unavailable in standard ChatGPT use. For legal research, both ChatGPT and Claude should be used as research accelerators — identifying potentially relevant cases and concepts — with every specific citation independently verified against Westlaw, LexisNexis, or primary court documents before any professional use. The standard of care is not "use the safer model"; it is "verify every claim."