hallucinate
When an AI confidently generates false information or made-up sources as if they were true.
As defined in Erik Hoel on the Threat to Humanity from AI
Other definitions
When an AI model generates plausible-sounding but false or made-up information, appearing confident despite errors.
The Venture Capitalist Who Bets on Science Most Investors Think Is Impossible | Josh Wolfe
Where it comes up · 8
- 312 - The Trouble With Ai3 claims
- Is AI's "intelligence" an illusion? | GZERO World with Ian Bremmer3 claims
- Dario Amodei (Anthropic CEO) — The hidden pattern behind every AI breakthrough1 claim
- Erik Hoel on the Threat to Humanity from AI1 claim
- $11B AI CEO: Why You Must Maximize Stress to Win1 claim
- Elon Musk #41 claim
- Julia Shaw1 claim
- Sam Altman #21 claim
Most important claims · 10 of 12Search all claims for “hallucinate” →
- There are two separable misinformation problems with these systems: first, they hallucinate—even given clean data they don't track relations between entities and properties, so they make stuff up (e.g., Galactica saying Elon Musk died in a 2018 car crash); second, bad actors can induce them to mass-produce variants of any specific misinformation, such as 100 fake QAnon or anti-vaccine articles each citing fabricated studies in Lancet and JAMA—the first threatens individuals via bad medical advice, the second threatens democracy by destroying common ground.312 - The Trouble With Ai
- Generative AI is effectively the ultimate false memory machine because it gives a tailored experience telling you what it thinks you want to hear with no safeguard against treating it as truth; moreover the distortion is bidirectional — leading questions from users change how the AI generates content, just as AI confabulation mirrors human confabulation, creating an under-examined 'in-between' that needs social scientists.Julia Shaw
- Language models like ChatGPT have not figured out that language is about a world containing things that are true or false; they never think or figure out but only find close approximations to text they have seen, so they generate well-formed output that is more like an illusion than genuine understanding—e.g., they cannot reliably do multi-digit arithmetic with carrying, treat chess as a string of notation without knowing there is a board, and don't track that giving someone a wallet means they now have it.312 - The Trouble With Ai
- Old-fashioned symbolic AI (good old-fashioned AI) based on logic, symbols, and mathematics has been dismissed prematurely and should be reconsidered because it is much better than neural networks at truth-telling when given a limited set of facts, as it reasons within those facts without hallucinating.Is AI's "intelligence" an illusion? | GZERO World with Ian Bremmer
- Gary Marcus previously made seven predictions about GPT-4 in an essay titled 'What to Expect When You're Expecting... GPT-4,' and all seven predictions proved correct.Is AI's "intelligence" an illusion? | GZERO World with Ian Bremmer
- Models like GPT-3 and early LLMs still hallucinate frequently, which caused early testers to dismiss them after a single bad output, creating selection bias where only people willing to tolerate failure continued using them.$11B AI CEO: Why You Must Maximize Stress to Win
- Existing large language models are unreliable for engineering because they hallucinate most precisely when you least want them to—on important, difficult questions they tend to be confidently wrong; the goal of Grok is to be grounded so answers trace back to physics first principles and mathematical logic, minimizing how often it is confidently wrong.Elon Musk #4
- In terms of generality—being able to reason and understand the world at a broad level rather than at a single narrow task—current large language models are arguably more general than the vast majority of people, which is precisely why companies are deploying them for search; the unsolved part is reliability and intelligence, since they hallucinate and make up sources.Erik Hoel on the Threat to Humanity from AI
- Large language models like ChatGPT-5 and ChatGPT-6 will continue to hallucinate significantly unless they involve radical new machinery beyond simply scaling up neural networks trained on more data.Is AI's "intelligence" an illusion? | GZERO World with Ian Bremmer
- The real bioterrorism danger is not the scary things models say that can already be Googled, but the 'missing' implicit knowledge scattered across textbooks or absent entirely — like lab-protocol troubleshooting; after six months working with bioweapons experts, Anthropic found models can sometimes already do these key steps (still saved partly by hallucination), and the trend across model generations shows a real problem in two to three years, at maybe 50-50 probability.Dario Amodei (Anthropic CEO) — The hidden pattern behind every AI breakthrough