Is there a good benchmark tracking hallucinations? The models are all incredibly good now, even the open ones, and my hope is that the rate of hallucinations is something that's falling off in concert with larger and larger context lengths.
Is there a good benchmark tracking hallucinations? The models are all incredibly good now, even the open ones, and my hope is that the rate of hallucinations is something that's falling off in concert with larger and larger context lengths.
> While OpenAI originally pioneered Codex (which went on to power GitHub Copilot), Google’s direct answer for dedicated, native code completion and natural-language-to-code generation is CodeGemma.
https://g.co/gemini/share/33e7a589a161
People complain about them incessantly, but I can almost never get people to actually post receipts. Every provider allows sharing chats, and anyone can share a prompt that reliably produces hallucinations.
More often than not, people are using images in responses that go awry. Which is fair, the models are sold as multi-modal, but image analyses is still at gpt-4.0 text-analyses levels.
Also knowledge cutoff issues, where people forget the models exist months to a year or more in the past.
I see constant hallucination in claude code when using specific tooling: It thinks it knows aws cli, for instance, but there's some flags that don't exist, it attempts to use all the time in 4.6 and 4.7. When asked about it, it says that yes , the flag doesn't exist in that command, but it exists in a different command (which it does), and yet, it attempts to use it without extra info.
Claude also believes it knows how AWS' KMS works, quite confidently, while getting things wrong. I have a separate "this is how KMS replication actually works" file just to deal with its misconceptions.
For gemini, I typically use it to query information from large corpuses, but it often web searches and hallucinates instead of reading the actual corpus. On a book series, it will hallucinate chapters and events which, while reasonable and plausible, do not exist. "Go look at the files and see if your reference is correct" shows that it's not correct, and it's a mandatory step. But that doesn't prevent hallucination, but makes sure you catch it after the fact, just like a method in a class that doesn't exist gets found out by the compiler. The LLM still hallucinated it.
https://gemini.google.com/share/9cd8ca68025a
I was trying to understand a game I've been playing, The Last Spell. I asked it for a tier list of omens -- which ones the community considers most important. At least a few of the names it posts are hallucinated ("omen of the sun" does not exist, and the omens that give extra gold are "savings," "fortune," and "great wealth").
Obviously not a critical use case but issues like this do keep me on my toes regarding whether the thing is working at all. I should ask 3.5 flash to do the same job. (I did try and it once again hallucinated the omen names and some of the effects.)
I can reliably produce hallucinations with this genre of prompt: "write a script that does <simple task> with <well known but not too-well-known API>." Even the frontier models will hallucinate the perfect API endpoint that does exactly what I want, regardless of if it exists.
The fix is easy enough though, a line in my global AGENTS.md instructing agents to search/ask for documentation before working on API integrations.
Yeah. Better to have more details in your prompt than fewer. For example, I use this:
```
Build a Nango sync that stores Figma projects.
Integration ID: figma
Connection ID for dry run: my-figma-connection
Frequency: every hour
Metadata: team_id
Records: Project with id, name, last_modified
API reference: https://www.figma.com/developers/api#projects-endpoints
```
Note: You do need a Nango account and the Nango Skill installed before it could work.
https://gemini.google.com/share/3717c8505d6b
Two of the three strip titles are hallucinated and two of the three strips are bad examples. Haley is mute in strip 403 and does nothing. Strip 578 is the start of the arc that shows the behavior Gemini is talking about, but has things going wrong so it's not a good example either.
Claude picks a good strip but also hallucinates the strip title: https://claude.ai/share/56be379d-c3da-443e-b60f-2d33c374eba8
I asked gemini 3.1 Pro to search for the linkedin URLs for a list of peers. It generated a plausible list of links -- but they were all hallucinated. On a follow up it confirmed it couldn't actually search, but didn't tell me that without prompting.
"People complain about them incessantly, but I can almost never get people to actually post receipts."
...my chats are all pretty long and involve personal conversations, or I've deleted them. It's a lot to ask for someone to post receipts. The number of complaints is enough data.
No matter how big the model is there will be edge cases where it has no data or is out of date. In these cases it just makes stuff up. You can detect it yourself by looking for words like usually or often when it states facts, e.g. "the mall often has a Starbucks." I asked it about a Genshin Impact character released in June 2025 and it consistently interpreted the name (Aino) as my player character because Aino wasn't in its data.
Honestly I'm surprised your haven't encountered it if you're using it more than casually. Pro is much better but not perfect.
Claude has gotten good in the past month or two at recognizing when it might need to search the web for updated info rather than saying that it has no idea what I'm talking about or making stuff up.
Are the knowledge cut off issues well known? I don't remember seeing them prominently displayed.
Also, prompts that reliably produce hallucinations is kind of a hard ask. It's inconsistent. One day the LLM I work with quotes verbatim from the PCIe spec and it's super helpful. The next day it gives me wrong information and when I ask it what section of the spec that information comes from it just makes up a section number
I see hallucinations ALL the time. It's only obvious when you're prompting about a subject you know well.
And when I say all the time, I mean it, and this is for Opus 4.7 Adaptive.
I often have to say, please do searches and cite sources, as if it doesn't it will confidently give me wrong or outdated information.
If you're often asking questions about a topic that's not in your specialist knowledge you won't notice them.
Hallucination is also much better controlled in the context of agentic coding because outputs can be validated by running the code (or linters/LSP). I almost never notice hallucinations when I’m coding with AI, but when using AI for legal work (my real job) it hallucinates constantly and perniciously because the hallucinations are subtle—e.g., making up a crucial fact about a real case.
Yes, you can catch many mistakes that LLMs make whike coding, but I wouldn't necessarily call it "controlled." Every now and then the LLM will run into dead ends where it makes a certain mistake, the compiler or unit tests find the mistake, so it tries a different approach that also fails, and then it goes back to the first approach, then tries the second approach again, and gets stuck in an endless loop trying small variations on those two approaches over and over.
If you aren't paying attention it can spend a long time (and a lot of tokens) spinning in that loop. Sometimes there might be more than two approaches in the loop, which makes it even harder to see that it's repeating itself in a loop. It's pretty frustrating to see it working away productively (so you think) for 20 minutes or so only to finally notice what's going on
Just ask any real question about stuff. LLM is not about code only...
well there is https://artificialanalysis.ai/evaluations/omniscience
It's a gibberish input detection benchmark, and does not measure output hallucinations.
I haven't been bothered by hallucinations in premier models since early last year. Still see it in smaller local models though.
I'm really running into this deep at the edges of content creation. Take, for example, a need to general some kind of legal work. The cost of painstakingly checking and rechecking each case cited is reducing the value of these frontier models immensely.
Coding, however, is solved like magic. Easier to add tests, to be fair.
It really depends what you are asking it. If the answer is in the training data, then the odds of it lying to you are much lower than if you are asking it for something it has never seen before.
maybe something like this? https://petergpt.github.io/bullshit-benchmark/viewer/index.v...
As long as the model uses web search, they almost never hallucinate anymore. The fast models (haiku, gpt-instant, flash) still sometimes have the problem where they don't search before answering so they can hallucinate
I've seen chatGPT and Gemini hallucinate even from web search, it's better is not sufficient
if last year's models were the ones people got familiar with in late 2022, hallucinations would be an underrepresented rumor, there would be no articles about it because its so rare. overconfident lawyers wouldn't have messed up dockets in court with fake case law, in other domains that move faster, sources would be only partially outdated with agentic search and mcp servers filling in the gaps
AI psychosis would be the problem people talk about more, not just outright agreement but subtle ways of making you feel confident in your ideas. "yes, buy that domain name buy these other ones for defensibility"
(the domain name is dumb and completely unmarketable)
The models still hallucinate bad when called via APIs, especially if web search is not enabled. Gemini hallucinates quite frequently even with the app and search enabled. More recent (e.g. ChatGPT 5.x and Deepseek v4) prompts/harnesses search very aggressively, which does greatly mitigate hallucinations.