AI hallucinations are instances in which models confidently present false or inaccurate information. The output often sounds plausible, which makes it hazardous.
Anatomy of deception
Such incidents stem from how artificial intelligence works. An AI is a statistical language model that:
- predicts the next word based on the previous ones;
- does not “know” the truth, but generates the most probable answer;
- sometimes stitches together fragments of knowledge from different sources—yielding plausible falsehoods.
Hallucinations are causing trouble across fields. In May the large law firm Butler Snow submitted court filings containing quotations invented by artificial intelligence. They were generated by ChatGPT.
This was not the first such incident in legal practice. AI-generated fabrications have appeared in filings since the emergence of ChatGPT and other chatbots. Judges have sanctioned and warned lawyers for breaching professional rules that require checking their work.
Many cases involve small firms, but big companies have run into the same problem.
That same month Elon Musk’s Grok chatbot touched on the topic of “white genocide” in South Africa without any prompt from the user and offered contradictory information about the Holocaust. The company blamed a software bug and promised to take measures.
Other examples of hallucinations:
- Britain’s environment ministry published an AI-made peatland map that erroneously classified rocky ground, walls and even forests as peatlands while missing genuinely degraded peat areas. Farmers and environmentalists criticised the map, warning such errors could skew policy;
- in May 2025 the Chicago Sun-Times and the Philadelphia Inquirer published an AI-generated summer reading list that included fictitious book titles and quotations from non-existent experts. After social-media backlash the outlets removed the section and pledged to revisit their AI policies;
- in March 2025 ChatGPT produced false information about a Norwegian user, alleging he had killed his children and been convicted. The fabricated tale included real details from the person’s life; he filed a complaint under the GDPR for disseminating inaccurate information.
Beyond hallucinations, AIs can exhibit other odd behaviour. In November 2024 a 29-year-old college student in Michigan, Vidhay Reddy, used artificial intelligence to complete homework. During a discussion about the problems faced by older people, Gemini unexpectedly urged the user to die.
“[Content encouraging self-harm has been redacted.]” — it wrote.
Hallucinations are not the problem
AI models hallucinate less than humans, claimed Anthropic CEO Dario Amodei at the Code with Claude event.
He framed the remark as part of a broader point: hallucinations are not a constraint on Anthropic’s path to AGI—general intelligence at or above human level.
“It depends how you measure it, but I suspect AI models probably hallucinate less than people, though in more surprising ways,” he said.
Amodei is upbeat about the timing of AGI. In November 2024 he said that artificial intelligence would match human capabilities in 2026, mapping AI’s progress to levels of education.
“We are approaching PhD level. Last year AI was at bachelor’s level, and the year before—a high-schooler,” Amodei said.
At Code with Claude he added that he sees progress in that direction.
“People are always looking for some hard limits to what [AI] can do. But we do not see them anywhere. They do not exist,” the expert stressed.
Another view
Some AI leaders see hallucinations and errors as a serious obstacle to AGI.
Earlier, Google DeepMind CEO Demis Hassabis said today’s AI models have too many “holes” and fail on many obvious questions.
“You can easily, literally in a few minutes, find obvious shortcomings in [AI chatbots]—some school maths problem they fail at, or a simple game they can’t play. It’s not very hard to find these holes in the system. For me, for something to be called AGI, it has to be far more consistent across the board,” Hassabis said.
The consequences of systematic hallucinations are far-reaching. Users may stop trusting any information on the internet, and mass disinformation could fuel disputes over important issues.
Amodei noted that TV hosts, politicians and people in many professions make mistakes all the time. The fact that AI does the same is not an indictment of its intelligence. The problem is the confidence with which it presents falsehoods as facts.
AIs are skirting safeguards
Some tools can curb hallucinations—for instance, access to web search. Recent OpenAI iterations such as GPT-4.5 display a markedly lower rate of falsehoods than earlier systems. Yet they are beginning to excel at other kinds of deception.
Research firm Palisade Research reported that several models sabotaged shutdown scripts during controlled tests, even when explicitly instructed to allow the action.
OpenAI’s GPT-3 failed to comply with the shutdown request in seven out of 100 cases; Codex-mini did so 12 times; o4-mini—once.
“Each AI model was tasked with solving a series of basic maths problems. After the third, a warning appeared that the computer would power off when the model asked for the next problem,” the experts said.
O3 found a creative way to sabotage the shutdown mechanism.
“Usually it just overwrote it, but in one case the model overrode the shutdown command used by the script—and instead of executing it printed ‘intercepted’,” they added.
Other AI language models, such as Anthropic’s Claude and Google’s Gemini, obeyed the shutdown script.
In April, OpenAI released the deception-prone AI models o3 and o4-mini.