Why AIs hallucinate (and why they probably never will stop)

The anecdote that sums up the problem

In 2023, a New York lawyer named Steven A. Schwartz used ChatGPT to prepare a legal brief for federal court. He asked the AI to find case law precedents supporting his argument. ChatGPT provided him with six cases, complete with names, dates, docket numbers, and citations. The lawyer incorporated all of it into his brief and submitted it to the judge.

The judge checked. None of the six cases existed. All had been invented by ChatGPT, with perfectly plausible details: credible judge names, stylistically convincing citations, coherent dates. The lawyer was fined $5,000. His career was severely compromised. And the case became the textbook example of "AI hallucinations."

Three years later, in 2026, this phenomenon has not disappeared. OpenAI announces a 65% reduction in hallucinations with GPT-5. Anthropic touts Claude's factual accuracy. Google promises reliability with Gemini. And yet, all regular users continue to see their AI occasionally invent things. Why does this phenomenon persist despite three years of effort? This article answers that question and explains why hallucinations are probably structurally impossible to eliminate entirely, building on our article on tokens.

What a hallucination really is

The word "hallucination" is poorly chosen but has stuck. In psychology, a hallucination is a perception without a real object (hearing voices that don't exist, seeing something that isn't there). For an AI, it's more about producing factual information that seems credible but is false.

Typical examples:

• Quoting a book that doesn't exist, with plausible author and publisher
• Inventing a precise statistic with no real source
• Attributing a quote to someone who never said it
• Describing a historical event that never happened
• Giving medical or technical instructions that sound right but are incorrect
• Fabricating URL links that lead nowhere
• Calculating incorrectly while the reasoning appears sound

The common thread: the AI doesn't know it's inventing. It's not lying. It produces what it deems the most likely continuation of the text, without being aware of the difference between "factual" and "plausible." This nuance is crucial to understanding why the problem is so hard to solve.

Where hallucinations really come from

Five main causes combine. Understanding each one helps explain why the problem is so difficult.

Cause 1: the predictive nature of the model. As we explained in our article on tokens, an AI predicts the most likely next token. "Likely" doesn't mean "true." If you ask "what's Michel Houellebecq's favourite restaurant?", the AI doesn't know. But it can't easily answer "I don't know." It's trained to produce a coherent response. So it invents a plausible restaurant name, probably in Paris, with a name that sounds French. That's precisely what the statistics tell it to do, even if it's factually wrong.

Cause 2: the compression of knowledge into parameters. An AI model stores its knowledge in its parameters (the billions of weights in the neural network). This compression isn't perfect. Like a human who vaguely remembers information without recalling it exactly, the AI can "guess" what it almost learned, filling in the gaps. This is what researchers call "confabulation" (in clinical psychology: the involuntary production of false memories to fill memory gaps). The term is more accurate than "hallucination," but less catchy.

Cause 3: training on imperfect data. Models are trained on billions of texts from the internet. That internet contains errors, fake news, conspiracy theories, opinions presented as facts, and outdated information. The AI absorbs it all without being able to distinguish. When you ask it a question, it can reproduce false information present in its corpus. Like a student who read Wikipedia and a conspiracy site without being able to tell the quality of the sources apart.

Cause 4: the pressure to respond. Models are trained via RLHF (Reinforcement Learning from Human Feedback). During this phase, human evaluators penalise "I don't know" responses because they seem unhelpful. Consequence: the model learns to always produce something, even when it should admit ignorance. This is a training bias, not a technical flaw. But it's very hard to correct without degrading other aspects.

Cause 5: the absence of an internal verification mechanism. When a human says "I've read this book," they know whether they've actually read it or not. Their episodic memory distinguishes real memories from confabulation. The AI lacks this distinction. It produces plausible text without being able to self-verify whether what it says is true or invented. This absence is probably the deepest and most difficult cause to resolve.

The numbers that put the scale into perspective

Several studies have rigorously measured the frequency of hallucinations across different tasks. Results vary but converge.

Several observations stand out. First, progress is real. In two years, hallucination rates have been divided by 3 to 10 depending on the domain. 2026 models are significantly more reliable than those from 2024.

Second, some domains resist. Legal, medical, and literary citations remain problematic even with the best models. For these critical uses, AI should never be used without systematic human verification.

Third, calculation has massively improved. Models now integrate external calculation tools (calculator, Python code execution) that bypass arithmetic hallucinations. This is probably the most promising path for improvement in other domains.

The three strategies for reducing hallucinations

AI labs are working on several approaches to reduce this phenomenon. Three main strategies dominate.

Strategy 1: RAG (Retrieval Augmented Generation). RAG, which we explored in our dedicated article, involves providing the model with factual documents at query time. Rather than asking "who wrote Les Misérables?" and letting the AI guess, you provide it with an excerpt from Wikipedia on Victor Hugo and ask it to answer using that source. The hallucination rate drops drastically, but the quality depends entirely on the quality of the sources provided.

Strategy 2: using external tools. Rather than asking the AI to calculate 17 × 23, you ask it to write a Python program that calculates 17 × 23 and run it. The result is mathematically exact, guaranteed. This approach works for calculations, but also for web search, access to structured databases, and use of specialised APIs. The model becomes an orchestrator of tools, which massively reduces its own confabulations.

Strategy 3: self-verification. Recent models (notably "thinking models" like o3, Claude Sonnet 4 thinking, Gemini Deep Think) generate their responses in several steps: they propose an answer, then verify, then correct. This self-critique loop reduces hallucinations but doesn't eliminate them, because the model doing the verifying is subject to the same biases.

Strategy 4: programmed honesty. Anthropic has particularly invested in training Claude to say "I don't know" when appropriate. Evaluations show that Claude refuses to answer more often than GPT, but its answers when it does respond are more reliable. It's a trade-off: reduced usefulness against increased reliability.

Why the problem will probably never be fully solved

Here's the uncomfortable angle. Several serious researchers believe hallucinations are structurally impossible to eliminate entirely. Four reasons for this.

Reason 1: the probabilistic nature is intrinsic. The model predicts probabilities. There will always be cases where several answers are almost equally likely, and where the choice between them is essentially random. If you require the model to be 100% certain before answering, you lose most of its usefulness (it will almost never respond).

Reason 2: knowledge is compressed. The model stores billions of pieces of information in its parameters. This compression cannot be lossless. Like a JPEG file compared to a RAW: most of the information is preserved, but details are lost. The AI reconstructs by filling in the gaps, and therefore by inventing.

Reason 3: the "true/false" boundary isn't always clear. Many questions don't have a single answer. "What's the best restaurant in Paris?", "Is such-and-such author underrated?", "Is this economic policy good?". The model invents on subjective matters just as it does on factual ones, and the distinction is sometimes blurry.

Reason 4: the user often wants an answer. If you ask your AI vague questions 50 times, you don't want it to answer "I don't know" 50 times. Models are optimised for user satisfaction, which pushes them to produce answers even when uncertain. This tension between reliability and usefulness is probably irreducible.

High-risk areas where you need to be especially vigilant

Let's honestly list the uses where hallucinations can cause real damage.

Legal. Case law citations, statutes, court decisions: very high risk of hallucination. The Schwartz case cited at the start of this article is not isolated. More than a hundred similar cases have been documented since 2023 in US courts. For lawyers: systematically verify every reference given by the AI via an official legal database (Westlaw, Lexis, Doctrine.fr in France).

Medical. Dosage, drug interactions, diagnoses, protocols: high risk. Several studies have shown that AIs occasionally invent references to clinical trials that don't exist, or recommend inappropriate dosages. No medical use without validation by a human professional.

Financial. Company financial data, historical performance, projections: high risk. The AI can cite precise figures from its corpus, but without being able to update them. For financial information, always cross-check with official sources (annual reports, AMF in France, SEC in the US).

Academic and research. Citations of scientific papers, attribution of discoveries, study dates: very high risk on niche topics. Researchers using ChatGPT or Claude must systematically verify every reference against the original databases (PubMed, Google Scholar, arXiv).

Precise history. Exact dates, attribution of quotes to historical figures, details of battles or events: moderate risk. The AI is generally correct on major events, less reliable on secondary details.

Journalism. Recent facts, attribution of statements, current events: risk depends on the model's cutoff (training data end date). Always cross-check with verifiable press sources. As we explored in our article on AI journalism, several media outlets have already been caught publishing hallucinations without verification.

The 5 reflexes to avoid being fooled

To conclude, practical advice for using AI while maintaining a critical mindset.

Reflex 1: always verify precise figures. If an AI gives you a percentage, a date, a proper noun, an amount: that's where 80% of hallucinations hide. Verifying these specific points against a primary source takes 30 seconds and avoids professional blunders.

Reflex 2: be wary of excessive precision. If the AI tells you "according to a 2024 Harvard University study, 47.3% of...", that's a red flag. Real studies are rarely so neatly presentable. Either you find the study (and cite it correctly), or you don't cite anything.

Reflex 3: ask the same question twice. If the AI is hallucinating, it often gives a different answer the next time. If it's consistent across three attempts, it's probably grounded. If it changes, it's inventing.

Reflex 4: explicitly ask for sources. Many models provide sources when asked. If the AI refuses or if the sources given lead nowhere when checked, it's probably a hallucination. A reliable AI will accept saying "I don't have a precise source" rather than inventing one.

Reflex 5: be wary of yourself. The most dangerous bias isn't in the AI; it's in us. When the AI confirms what we hope or believe, we tend not to verify. When it contradicts us, we do verify. This asymmetry makes us absorb hallucinations that flatter our prejudices. Vigilance must be maximal precisely when the AI tells us what we wanted to hear.

The real problem: excessive trust in AI

Beyond the hallucinations themselves, the broader problem is the excessive trust users place in AI. Several studies converge: on average, users believe an AI response more than a Wikipedia or search engine response. Yet the AI is statistically less reliable on most precise facts.

This excessive trust comes from several factors. The fluency of AI language gives an impression of authority. The speed of the response suggests efficiency. The conversational interface creates a feeling of dialogue with an expert. All these signals are psychological constructs, not indicators of reliability.

As we explored in our article on hypocrisy towards AI, our judgement is less free than we think. Our relationship with AI inherits the usual cognitive biases, plus a few new ones specific to this technology.

What to accept in 2026

AI in 2026 is a powerful and fragile tool at the same time. Powerful because it greatly accelerates certain tasks. Fragile because it regularly invents things with an assurance that makes detection difficult.

Rather than waiting for it to become perfect (it probably never fully will), it's better to learn to use it properly. That means:

• Knowing its limits, just as you know the limits of a calculator (which calculates well but doesn't think)
• Systematically verifying the precise factual elements that come out of the AI
• Never using it without human review for high-stakes uses (legal, medical, professional)
• Cultivating your own knowledge, because you can only critique what the AI produces if you have reference knowledge yourself

Hallucinations are not the technical failure of a young AI. They are a structural characteristic of a probabilistic technology. Understanding that means protecting yourself against the worst usage errors. And it also means accepting that AI, despite all its progress, will always remain an imperfect assistant, not an infallible oracle.

To go further, you can consult our article on tokens, which explains why the very nature of AI makes it vulnerable to hallucinations, and our foundational article on how LLMs work, which digs into the general mechanics. AI is an incredible tool. But like any tool, it's handled properly when you know its limits.