AI Hallucinations: A Complete Guide to Why They Happen and How to Catch Them

What a Hallucination Actually Is

An AI hallucination is confident, fluent, plausible-sounding output that is factually wrong — a made-up statistic, a court case that doesn’t exist, a book that was never written, a date that’s simply incorrect — delivered with the exact same tone of confidence the model uses for things it gets right. That last part is what makes hallucinations genuinely dangerous rather than just an occasional annoyance: nothing in a model’s phrasing reliably signals when it’s guessing versus when it’s citing something it was actually trained on.

The Technical Reason This Happens

Large language models are trained to predict the next most statistically likely word given everything before it, learned from patterns across enormous amounts of text. Nothing in that core training objective is a fact-checking step — the model isn’t consulting a database and reporting what it finds, it’s generating text that’s statistically consistent with the patterns it learned. Most of the time, “statistically likely” and “factually true” line up closely, because true statements about the world tend to appear more consistently across training data than false ones. But when a model is asked about something obscure, or something at the edge of its training data, or something that requires precise recall of a specific number or name, the statistically likely continuation can drift away from the true one — and the model has no separate mechanism that steps in and says “wait, I’m not actually sure about this.” It generates the next plausible token whether or not the underlying claim is verified.

Why Hallucinations Get Worse in Specific Situations

Hallucination rates aren’t uniform across question types — they climb sharply in a few predictable situations. Obscure or long-tail facts (a minor historical figure, a small company’s founding date, a niche technical spec) are hit hardest, because the model saw less consistent training signal about them and is more likely to blend or invent plausible-sounding detail. Citations and sources are a particularly acute failure mode: a fabricated citation looks exactly like a real one — correct formatting, a plausible author name, a plausible journal or case name — because the model is generating something that pattern-matches “what a citation looks like,” not retrieving an actual reference. Specific numbers, dates, and quotes are riskier than general explanations for the same reason: getting the gist of a concept right requires much less precision than getting a specific figure exactly correct, and models are better at the former than the latter. Multi-step arithmetic and confident wrong math follow the same pattern — a model can walk through a calculation with completely correct-looking reasoning steps and still land on a wrong final number, because the generation process isn’t actually running the computation, it’s generating text that looks like someone doing the computation.

What Current Research Actually Shows

Two credible, current benchmarks are worth knowing by name because a lot of hallucination-rate claims online cite neither. Vectara, an AI infrastructure company, maintains an open hallucination leaderboard built on its HHEM model, which measures how often different models introduce ungrounded claims when summarizing real documents — a task specifically chosen because it gives the model source material and checks whether it stayed faithful to it. Separately, OpenAI’s own GPT-5 system card, published in August 2025, reports the model’s performance on public factuality benchmarks (LongFact and FActScore) and states that GPT-5 with web search enabled was roughly 80 percent less likely to produce a factual error than the prior o3 model on the company’s internal test set, with hallucinations on difficult conversations reduced roughly eightfold. Both are real, methodologically documented improvements — and both are also narrower than they sound in a headline. A benchmark measuring grounded summarization or search-enabled factuality doesn’t tell you the rate for open-ended, non-search questions about obscure topics, which is where hallucination remains most common. Treat any specific hallucination-rate percentage you encounter, including the ones in this guide, as tied to a specific benchmark and task type, not a universal number for “how often AI lies.”

The Legal System Has Already Learned This the Hard Way

The clearest documented case of AI hallucination causing real-world consequences is Mata v. Avianca, a 2023 federal case in which attorneys submitted a legal brief containing multiple citations to court cases that did not exist — ChatGPT had fabricated them, complete with plausible-sounding case names and quotations, and the attorneys did not independently verify them before filing. The presiding judge sanctioned the attorneys. It was not an isolated incident: courts across the U.S. have issued similar sanctions in numerous subsequent cases involving AI-fabricated citations, enough that several state and federal courts have since adopted specific disclosure rules for AI use in filings. The pattern in these cases is consistent and instructive for any user, not just lawyers: the fabricated content was specific, technical, and confidently formatted — exactly the category of output this guide flags as highest-risk.

Model Disagreement as a Useful Signal

One practical technique that doesn’t get discussed as often as it should: asking two different AI models the same factual question and comparing the answers. Different models are trained on different data with different architectures, so a fact one model hallucinates isn’t necessarily one another model hallucinates the same way — when two independently trained models agree closely on a specific, checkable claim, that’s a mildly reassuring signal; when they diverge meaningfully on a specific number, date, or citation, that divergence itself is useful information, a flag that the claim needs independent verification rather than something either model can resolve for you by being asked again. This isn’t a substitute for actually checking a primary source, but it costs almost nothing and catches a meaningful share of hallucinations that would otherwise go unnoticed, particularly because a single model asked the same question twice will often confidently repeat the same fabrication rather than catching its own error.

Practical Detection Techniques for Non-Experts

A few habits catch most hallucinations without requiring any technical background. First, be more skeptical of specific claims than general ones — a model explaining how photosynthesis works in broad strokes is on much safer ground than a model citing a specific study’s exact sample size or a specific court case’s docket number, so scrutinize the latter category harder. Second, ask the model directly for its source, and then independently verify that source exists and actually says what the model claims — don’t stop at “it gave me a citation,” confirm the citation is real. Third, cross-check any number, date, name, or quote that matters against an independent source before relying on it, especially before repeating it publicly or using it to make a decision. Fourth, treat unusual confidence about a very specific, hard-to-verify claim as a mild red flag rather than reassurance — models don’t hedge more when they’re actually less certain, so confident phrasing carries no real signal either way. Fifth, when the stakes are high (medical, legal, financial, or anything you’d be embarrassed to have wrong in public), get independent confirmation regardless of how convincing the AI’s answer sounds.

Why RAG and Better Prompting Help but Don’t Solve the Problem

Retrieval-augmented generation — where the model is given real, retrieved documents and asked to answer based on them — meaningfully reduces hallucination because the model has actual source text to draw from instead of relying purely on generated pattern-matching. This is a large part of why hallucination rates have fallen substantially on document-grounded tasks specifically. But RAG doesn’t eliminate the problem: the model can still misquote the retrieved text, blend it with non-retrieved assumptions without flagging which is which, or misattribute a claim to the wrong part of a long document. Prompting techniques — asking the model to cite sources, to say “I don’t know” when uncertain, or to reason step by step — help at the margins for similar reasons, but none of them give the model a verification mechanism it didn’t have before; they just nudge its existing pattern-generation process toward outputs that happen to be more reliable more often. None of these techniques should be treated as a substitute for actually checking anything that matters.

Bottom Line

Hallucinations happen because language models generate statistically plausible text, not verified fact, and nothing in their core architecture distinguishes a confident guess from a confirmed answer — which is why fabricated citations, wrong numbers, and invented specifics can read exactly as convincingly as accurate ones. Rates have genuinely improved on measured benchmarks, but the honest response to any single AI-generated claim that matters is still to check the specific, hard-to-verify details yourself rather than trust the model’s tone of confidence.