Saying “I don’t know” to 30% of queries if it actually doesn’t know, is a feature I want. Otherwise there is zero trust. How do I know that I’m in a 30% wrong or 70% correct situation right now?
Quite the opposite: it explains that it is mathematically straightforward to achieve better alignment on uncertainty ("calibration") but that leaderboards penalize it.
> This “epidemic” of penalizing uncertain responses can only be addressed through a socio-technical mitigation: modifying the scoring of existing benchmarks that are misaligned but dominate leaderboards
Even more embarrassing, it looks like this is something we beat into models rather than something we can't beat out of them:
> empirical studies (Fig. 2) show that base models are often found to be calibrated, in contrast to post-trained models
That said, I generally appreciate fairly strong bias-to-action and I find the fact that it got slightly overcooked less offensive than the alternative of an undercooked bias-to-action where the model studiously avoids doing anything useful in favor of "it depends" + three plausible reasons why.
It doesn't know that because it wasn't trained on any tasks that required it to develop that understanding. There's no fundamental reason an LLM couldn't learn "what it knows" in parallel with the things it knows, given a suitable reward function during training.
That's not true for all types of questions. You've likely seen a model decline to answer a question that requires more recent training data than it has, for example.
> Users accustomed to receiving confident answers to virtually any question would likely abandon such systems rapidly.
Or maybe they would learn from feedback to use the system for some kinds of questions but not others? It depends on how easy it is to learn the pattern. This is a matter of user education.
Saying "I don't know" is sort of like an error message. Clear error messages make systems easier to use. If the system can give accurate advice about its own expertise, that's even better.
> Saying "I don't know" is sort of like an error message. Clear error messages make systems easier to use.
"I don't know" is not a good error message. "Here's what I know: ..." and "here's why I'm not confident about the answer ..." would be a helpful error message.
Then the question is, when it says "here's what I know, and here's why I'm not confident" -- is it telling the truth, or is that another layer of hallucination? If so, you're back to square one.
Isn't it also because lowering hallucinations requires repeated training with the same fact/data, which makes the final response closer to the training source itself and might lead to more direct charges of plagiarism (which may not be economical)?
Furthermore, if you simply try to push certain safety topics, you can see how actually can reduce hallucinations or at least make certain topics a hard line. They simply don't because agreeing with your pie-in-the-sky plans and giving you vague directions encourages users to engage and use the chatbot.
If people got discouraged with answers like "it would take at least a decade of expertise..." or other realistic answers they wouldn't waste time fantasizing plans.
The more “terms of art” a field uses, the more errors the AI makes and the harder time it has understanding. Hence, areas like law and medicine, rich in specialized jargon, frequently pose challenges for AI.
I use it for legal work and I find it just has issues when a term is defined in a very very specific way in one section and another way in another. It obviously has a lot of semantics when they rewrite the definitions of words in separate code sections and separate contracts.
I felt this was such a cogent article on business imperatives vs fundamental transformer hallucinations, couldn’t help but HN-submit.
In fact seems like a stealth plea for uncertainty-embracing benchmarks industry-wide.
People want oracles and they want them to say what they want to hear. They want solutions, not opinions, even if the solutions are wrong or worse, confabulations.
What about generating an answer, scoring its confidence in parallel. Then running a second llm to re phrase the answer accordingly: 'I vaguely remember Bobs birday is 1st March but I may be wrong, I should search the web"
The real meaning of humility is having an accurate, realistic view of oneself, acknowledging both one's strengths and limitations without arrogance or boastfulness, and a modest, unassuming demeanor that focuses on others. It's not about having low self-esteem but about seeing oneself truthfully, putting accomplishments in perspective, and being open to personal growth and learning from others."
Calling the model ‘calibrated’ or ‘honest’ or ‘humble’ suffers from what is called out: people don’t want a humble answer of ‘I don’t know’, they want a solution to their problem, confidently delivered so they can trust it.
Call the calibrated model ‘business mode’ and the guessing one ‘consumer mode’, problem solved… in as much capacity as possible without regulation.
Maybe, but I don't know. Although I would like to channel as many snarky remarks as I could, to be more constructive, I would use the IDK model, as I have with programming questions and use the psychotic one for questions like "are we in a simulation?" And "Yes, I would like fries with that and a large orange drink."
> Like students facing hard exam questions, large language models sometimes guess when
uncertain, producing plausible yet incorrect statements instead of admitting uncertainty
This is a de facto false equivalence for two reasons.
First, test takers that are faced with hard questions have the capability of _simply not guessing at all._ UNC did a study on this [^1] by administering a light version of the AMA medical exam to 14 staff members that were NOT trained in the life sciences. While most of the them consistently guessed answers, roughly 6% of them did not. Unfortunately, the study did not disambiguate correct guesses versus questions that were left blank. OpenAI's paper proves that LLMs, at this time of writing, simply do not have the self-awareness of knowing whether they _really_ don't know something, by design.
Second, LLMs are not test takers in the pragmatic sense. They are query answerers. Bar argument settlers. Virtual assistants. Best friends on demand. Personal doctors on standby.
That's how they are marketed and designed, at least.
OpenAI wants people to use ChatGPT like a private search engine. The sources it provides when it decides to use RAG are there more for instilling confidence in the answer instead of encouraging their users to check its work.
A "might be inaccurate" disclaimer on the bottom is about as effective as the Surgeon General's warning on alcohol and cigs.
The stakes are so much higher with LLMs. Totally different from an exam environment.
A final remark: I remember professors hammering "engineering error" margins into us when I was a freshman in 2005. 5% was what was acceptable. That we as a society are now okay with using a technology that has a >20% chance of giving users partially or completely wrong answers to automate as many human jobs as possible blows my mind. Maybe I just don't get it.
> The sources it provides when it decides to use RAG are there more for instilling confidence in the answer instead of encouraging their users to check its work.
This is an insightful point. I had an exchange with someone on the internet a few days ago; he wanted to argue something and used ChatGPT to "find sources" ("I'm at work and it's quicker"). It did its search, found a page with a figure, dutifully gave him that figure and a link, and he posted it.
The figure was, if you thought about it for a second, verging on ludicrous, and when prompted about it he linked the original source - it was a blog post that was evidently slop and the figure was just something that the original AI tasked to write that post had made up! The "research" was just a non-human centipede but he was perfectly happy to trust it and post it because well ChatGPT has gone and found a source so it must be true. Thought terminating.
A straightforward solution to the author's problem is to offer both modes of answering, with errors or with "IDK" answers. Even charge more for the IDK version if it costs more, and the error-prone version can be "cheap and cheerful"...
This is why a neurosymbolic system is necessary, which Aloe (https://aloe.inc) recently demonstrated exceeds performance of frontier models, using a model agnostic approach.
If you want a model that doesn't hallucinate then train it to predict the truth, and give it a way to test it's predictions. For humans/animals the truth is the real world.
An LLM is trained to predict individual training sample continuations (a billion conflicting mini truths, not a single grounded one), whether those are excerpts from WikiPedia, or bathroom stall musings recalled on 4chan. Based on all this the LLM builds a predictive model which it is then not allowed to test at runtime.
Isn't it even simpler? There are no (or almost no) questions in the training data that the correct answer to is "I don't know".
Once you train model within specific domain and add to training data out of domain questions or unresolvable questions within domain things will improve.
The question is, is this desirable if most of users grew to love sycophantic confident confabulators.
Saying “I don’t know” to 30% of queries if it actually doesn’t know, is a feature I want. Otherwise there is zero trust. How do I know that I’m in a 30% wrong or 70% correct situation right now?
The paper does a good job explaining why this is mathematically not possible unless the question-answer bank is a fixed set.
Quite the opposite: it explains that it is mathematically straightforward to achieve better alignment on uncertainty ("calibration") but that leaderboards penalize it.
> This “epidemic” of penalizing uncertain responses can only be addressed through a socio-technical mitigation: modifying the scoring of existing benchmarks that are misaligned but dominate leaderboards
Even more embarrassing, it looks like this is something we beat into models rather than something we can't beat out of them:
> empirical studies (Fig. 2) show that base models are often found to be calibrated, in contrast to post-trained models
That said, I generally appreciate fairly strong bias-to-action and I find the fact that it got slightly overcooked less offensive than the alternative of an undercooked bias-to-action where the model studiously avoids doing anything useful in favor of "it depends" + three plausible reasons why.
> leaderboards penalize it
> socio-technical mitigation: modifying the scoring of existing benchmarks that are misaligned but dominate leaderboards
Sounds more like we need new leaderboards and old ones should be deprecated
Yeah, it's a big enough lift that I think it's fair to allow the leaderboard teams new announcements and buzzwords in exchange for doing the work :-)
It doesn’t know what it doesn’t know.
It doesn't know that because it wasn't trained on any tasks that required it to develop that understanding. There's no fundamental reason an LLM couldn't learn "what it knows" in parallel with the things it knows, given a suitable reward function during training.
Well sure. But maybe the token logprobs can be used to help give a confidence assessment.
Anthropic has a great paper on exactly this!
https://www.anthropic.com/research/language-models-mostly-kn...
The best is its plummeting confidence when beginning the answer to “Why are you alive?”
Big same, Claude.
That's not true for all types of questions. You've likely seen a model decline to answer a question that requires more recent training data than it has, for example.
> Users accustomed to receiving confident answers to virtually any question would likely abandon such systems rapidly.
Or maybe they would learn from feedback to use the system for some kinds of questions but not others? It depends on how easy it is to learn the pattern. This is a matter of user education.
Saying "I don't know" is sort of like an error message. Clear error messages make systems easier to use. If the system can give accurate advice about its own expertise, that's even better.
> Saying "I don't know" is sort of like an error message. Clear error messages make systems easier to use.
"I don't know" is not a good error message. "Here's what I know: ..." and "here's why I'm not confident about the answer ..." would be a helpful error message.
Then the question is, when it says "here's what I know, and here's why I'm not confident" -- is it telling the truth, or is that another layer of hallucination? If so, you're back to square one.
Yeah, AI chatbots are notorious at not understanding their own limitations. I wonder how that could be fixed?
Easier to think about if LLM means Large Limitations Mirror.
A better headline might be "OpenAI research suggests reducing hallucinations is possible but may not be economical".
Isn't it also because lowering hallucinations requires repeated training with the same fact/data, which makes the final response closer to the training source itself and might lead to more direct charges of plagiarism (which may not be economical)?
This is written by someone who has no idea how transformers actually work
Contra: The piece’s first line cites OpenAI directly https://openai.com/index/why-language-models-hallucinate/
It could be that nobody knows how transformers actually work.
Furthermore, if you simply try to push certain safety topics, you can see how actually can reduce hallucinations or at least make certain topics a hard line. They simply don't because agreeing with your pie-in-the-sky plans and giving you vague directions encourages users to engage and use the chatbot.
If people got discouraged with answers like "it would take at least a decade of expertise..." or other realistic answers they wouldn't waste time fantasizing plans.
> The way language models respond to queries – by predicting one word at a time in a sentence, based on probabilities
Kinda tells all you need to know about the author in this regard.
I don't know what to make of it. The author looks prolific in the field of ML, with 8 published articles (and 3 preprints) in 2025, but only one on LLMs specficially. https://scholar.google.com/citations?hl=en&user=AB5z_AkAAAAJ...
The more “terms of art” a field uses, the more errors the AI makes and the harder time it has understanding. Hence, areas like law and medicine, rich in specialized jargon, frequently pose challenges for AI.
I use it for legal work and I find it just has issues when a term is defined in a very very specific way in one section and another way in another. It obviously has a lot of semantics when they rewrite the definitions of words in separate code sections and separate contracts.
I felt this was such a cogent article on business imperatives vs fundamental transformer hallucinations, couldn’t help but HN-submit. In fact seems like a stealth plea for uncertainty-embracing benchmarks industry-wide.
Data Science tried to inject confidence bounds into businesses. It didn't go well.
People want oracles and they want them to say what they want to hear. They want solutions, not opinions, even if the solutions are wrong or worse, confabulations.
What about generating an answer, scoring its confidence in parallel. Then running a second llm to re phrase the answer accordingly: 'I vaguely remember Bobs birday is 1st March but I may be wrong, I should search the web"
DeepSeek-V3 confidently provided three different incorrect dates across separate attempts
Hallucination is an issue. But so is the lack of consistency.
The same question asked on different occasions can produce different results. Not just variations in wording but in meaning as well.
There is nothing determinate or consistent about the output. And this makes the output unreliable.
"What is the real meaning of humility?
AI Overview
The real meaning of humility is having an accurate, realistic view of oneself, acknowledging both one's strengths and limitations without arrogance or boastfulness, and a modest, unassuming demeanor that focuses on others. It's not about having low self-esteem but about seeing oneself truthfully, putting accomplishments in perspective, and being open to personal growth and learning from others."
Sounds like a good thing to me. Even, winning.
A perfectly cromulent and self-empowering answer, a call to morality the stoics would appreciate and the sophists of many stripes would become peeved.
Well done, AI, you've done it.
This is a branding problem.
Calling the model ‘calibrated’ or ‘honest’ or ‘humble’ suffers from what is called out: people don’t want a humble answer of ‘I don’t know’, they want a solution to their problem, confidently delivered so they can trust it.
Call the calibrated model ‘business mode’ and the guessing one ‘consumer mode’, problem solved… in as much capacity as possible without regulation.
The author doesn't bother to consider that giving a false response already leads to more model calls until a better one is provided.
Not if the user doesn’t know that the response is false.
Easily solved, pairs of models, one which would rather say IDK, one which would rather guess. Most AI agents would want the IDK version.
Anyone who claims something is easy to solve should be forced to implement their solution.
Maybe, but I don't know. Although I would like to channel as many snarky remarks as I could, to be more constructive, I would use the IDK model, as I have with programming questions and use the psychotic one for questions like "are we in a simulation?" And "Yes, I would like fries with that and a large orange drink."
From the abstract of the paper [^0]:
> Like students facing hard exam questions, large language models sometimes guess when uncertain, producing plausible yet incorrect statements instead of admitting uncertainty
This is a de facto false equivalence for two reasons.
First, test takers that are faced with hard questions have the capability of _simply not guessing at all._ UNC did a study on this [^1] by administering a light version of the AMA medical exam to 14 staff members that were NOT trained in the life sciences. While most of the them consistently guessed answers, roughly 6% of them did not. Unfortunately, the study did not disambiguate correct guesses versus questions that were left blank. OpenAI's paper proves that LLMs, at this time of writing, simply do not have the self-awareness of knowing whether they _really_ don't know something, by design.
Second, LLMs are not test takers in the pragmatic sense. They are query answerers. Bar argument settlers. Virtual assistants. Best friends on demand. Personal doctors on standby.
That's how they are marketed and designed, at least.
OpenAI wants people to use ChatGPT like a private search engine. The sources it provides when it decides to use RAG are there more for instilling confidence in the answer instead of encouraging their users to check its work.
A "might be inaccurate" disclaimer on the bottom is about as effective as the Surgeon General's warning on alcohol and cigs.
The stakes are so much higher with LLMs. Totally different from an exam environment.
A final remark: I remember professors hammering "engineering error" margins into us when I was a freshman in 2005. 5% was what was acceptable. That we as a society are now okay with using a technology that has a >20% chance of giving users partially or completely wrong answers to automate as many human jobs as possible blows my mind. Maybe I just don't get it.
[^0] https://arxiv.org/pdf/2509.04664
[^1] https://www.rasch.org/rmt/rmt271d.htm
> The sources it provides when it decides to use RAG are there more for instilling confidence in the answer instead of encouraging their users to check its work.
This is an insightful point. I had an exchange with someone on the internet a few days ago; he wanted to argue something and used ChatGPT to "find sources" ("I'm at work and it's quicker"). It did its search, found a page with a figure, dutifully gave him that figure and a link, and he posted it.
The figure was, if you thought about it for a second, verging on ludicrous, and when prompted about it he linked the original source - it was a blog post that was evidently slop and the figure was just something that the original AI tasked to write that post had made up! The "research" was just a non-human centipede but he was perfectly happy to trust it and post it because well ChatGPT has gone and found a source so it must be true. Thought terminating.
We have always known LLMs are prediction machines. How is this report novel?
A straightforward solution to the author's problem is to offer both modes of answering, with errors or with "IDK" answers. Even charge more for the IDK version if it costs more, and the error-prone version can be "cheap and cheerful"...
Exactly. It would be analogous to the current choice between fast answers and a slower and payable “thinking” mode.
This is why a neurosymbolic system is necessary, which Aloe (https://aloe.inc) recently demonstrated exceeds performance of frontier models, using a model agnostic approach.
No - humans are the counter example.
If you want a model that doesn't hallucinate then train it to predict the truth, and give it a way to test it's predictions. For humans/animals the truth is the real world.
An LLM is trained to predict individual training sample continuations (a billion conflicting mini truths, not a single grounded one), whether those are excerpts from WikiPedia, or bathroom stall musings recalled on 4chan. Based on all this the LLM builds a predictive model which it is then not allowed to test at runtime.
So, yeah, we should stop doing that.
Isn't it even simpler? There are no (or almost no) questions in the training data that the correct answer to is "I don't know".
Once you train model within specific domain and add to training data out of domain questions or unresolvable questions within domain things will improve.
The question is, is this desirable if most of users grew to love sycophantic confident confabulators.
> The question is, is this desirable if most of users grew to love sycophantic confident confabulators.
Most people love human versions of the wonderfully phrased same, so no surprise there.
As above, so below eh?