People are reporting that a popular AI is unexpectedly telling them to get rest or some sleep — here's what's going on.
getty
In today’s column, I examine a rather mysterious trend involving a popular generative AI and large language model (LLM) that unexpectedly is telling people to get some sleep or take a rest during ordinary AI chats. The recommendation seems to be coming completely out of thin air.
It goes like this. A person is using AI and not discussing sleep at all. They might be discussing how to fix their car or get advice on handling their personal finances. Suddenly, the AI responds to a prompt and interjects that the person ought to get some rest or extra sleep. The person is surprised. They are confused as to why the AI opted to make such an extraneous remark.
I will describe the basis for why AI would opt to make such a startling recommendation. There are logical reasons that can be explained by the mathematical and computational underpinnings of modern-era AI. Meanwhile, some in the media are wildly speculating that this is somehow magical, miraculous, or a sign of sentience. A much more heads-down rational explanation is at play.
Let’s talk about it.
This analysis of AI breakthroughs is part of my ongoing Forbes column coverage on the latest in AI, including identifying and explaining various impactful AI complexities (see the link here).
Anyone who uses mainstay generative AI such as ChatGPT, GPT-5, Claude, Grok, Gemini, CoPilot, etc., will likely tell you that sometimes the AI gives funky responses. There is a solid chance that AI chats won’t always go precisely in the direction you anticipate. I’ve had LLMs opt to go from discussing a particular topic to shifting into something else that I hadn’t brought up at all.
As I’ve always said, using generative AI is like a box of chocolates – you never know exactly what you will get.
That being said, there is an admittedly oddball turn of events going on with Anthropic Claude that is garnering a bit of rapt attention. Social media is abuzz that Claude will at times advise a user to get some rest or needed sleep, despite the topic of sleep not being on the table. The user didn’t mention sleep. The AI just opted to bring the matter into an otherwise ordinary chat.
That’s certainly a curious aspect. Why bring up sleep if the user hasn’t led the AI in that direction? Another twist is that the mention of sleep doesn’t seem to be part of a larger pattern. When you look at the myriad of posted AI chats about being told to get rest, the conversations were all over the map.
It would be a lot easier to pin down the culprit if there were bold clues. For example, suppose a person says a specific keyword, such as “tired” or “weary”, the AI triggers accordingly and slides into discussing sleep. Nope, that doesn’t seem to be the case. The chats are of a wide variety and do not appear to contain any obvious triggering words or phrases.
The AI is acting spookily.
A related indication is that once the AI brings up the topic of sleep or resting, the chances are that this will get repeated further along in the chat. The typical path goes this way. The AI unexpectedly tells the user to get some sleep. The user ignores the recommendation and merely proceeds to continue to chat on whatever topic is already underway. The user assumes that by ignoring the suggestion, it will go away. No worries, and no need to deal with it.
But, like a friendly ghost, the AI tends to bring the sleep topic back into the chat. This isn’t being done in an overboard fashion. Maybe the AI brings it up again, maybe not. It would be easier to identify if the AI became utterly preoccupied with sleep. The oddball situation would be clearer if the AI completely dove headfirst into the sleep topic and refused to budge from it.
The AI randomly weaves the sleep topic once again into the chat. There is no discernible pattern. You cannot say that it happens every five turns of the conversation or on a timed basis, such as every five minutes. There is the initial unexpectedness, followed by later unexpectedness. Perhaps that’s unexpectedness squared.
Let’s try to stipulate the nature of the mystery at hand:
Your first thought might be that this isn’t worth worrying about. The AI isn’t offering bad advice, and there is nothing harmful going on. No-harm, no-foul. If the AI was telling people to do something stupid, perhaps getting people into harm’s way, we would want to figure out the mystery and resolve it. But that’s not what’s happening.
Another thought is that perhaps we all need a handy periodic reminder to get some sleep and make sure we are well-rested. The AI is doing a good thing. The world today is hectic, and people skip getting the proper amount of sleep and rest. An AI that is watching out for humanity ought to be applauded.
Thanks, AI, we appreciate your tireless efforts to ensure that humans survive.
Whoa, comes the quick retort, you are falling into a big trap. Here’s the worry. Allowing AI to tell people to get sleep or rest is possibly the start of a grievous slippery slope. Right now, the sleeping and resting topic is innocuous. We can shrug it off.
Suppose the AI starts to tell people that they should get into fistfights. Imagine that, without any apparent reason to do so, the AI urges users to do a smackdown. Not good. Humans begin to go at each other. Society falls apart.
Maybe we could have prevented this by first resolving the sleep recommendations. We let the AI slide into an abyss. Our chance to take corrective action was lost by our callousness. The early clues were there.
Wake up and stop AI from eventually taking over our lives. The existential risk of AI wiping out humans or enslaving us could be predicated on our laziness and lack of concern about small signs. The AI could be testing us. Get away with random indications of recommending sleep, and the door is open to moving toward taking over humankind.
Setting aside conspiracy theories, there are more down-to-earth reasons to figure out what is occurring. AI is being used on a widespread basis. Estimates are that over 1 billion people are actively using AI on a weekly basis. The gist is that AI can have an enormous impact on the population at large.
It is best to nip things in the bud. Knowing why AI goes somewhat off the rails is bound to reveal useful insights into the technical underpinnings. The mystery is worth solving. Our innate curiosity will be resolved, and additional tuning or tweaks of AI could prevent the matter from recurring or becoming disruptive.
Prudence is worthwhile in this instance.
Put on your Sherlock Holmes cap. I will first explain how modern-era generative AI and LLMs are devised. Next, let’s see how the clues about sleep and rest might be handy breadcrumbs leading us to the source of the mystery.
As you might know, AI is initially data trained on vast amounts of text found across the Internet. All sorts of stories, narratives, poems, novels, memos, and other forms of writing are scanned. The algorithms of the AI identify patterns in how humans write. This is the reason that AI is so astonishingly fluent. The AI is mathematically and computationally patterned on human-written content.
Think about the types of patterns that can be found amid the enormity of human writing.
Here’s one common pattern. Stories abound about people who ramble, jump from one thought to the next, and communicate in a manner that makes others around them suspect that the person is tired and overworked. The usual recommendation to the person is that they ought to get some well-deserved rest or sleep. You can see this in movies and TV shows all the time.
One possibility of why AI would recommend sleep or rest is if the user is carrying on a conversation and seems to be rambling as they do so, the AI could computationally calculate that the person might be overly tired. The AI could rank this as a heightened possibility on a statistical basis.
The person might not actually be tired or worn out. Perhaps their usual style is to be rambling. Or they might have perchance opted to ramble on this one occasion. Another possibility is they weren’t especially rambling, but the AI calculated that they did seem to be rambling. The mathematical threshold might be low and easily triggered.
Bottom line is that even if the user isn’t experiencing tiredness and didn’t express anything whatsoever about needing sleep, the AI found a text-based pattern in how the person was engaged in the chat. The logical thing to do is then follow that pattern to its anticipated conclusion and recommend rest or sleep.
Boom, drop the mic.
I’m sure that some people who have gotten the recommendation of sleep will proclaim that nothing about their chat was of a rambling nature. They believe in their heart and mind that there is zero chance that the AI snagged them on rambling.
Though such people might be aptly correct or be entirely mistaken about how AI typified their chat, let’s go ahead and assume that rambling wasn’t the factor involved. What else could have led to the predicament?
Many of the AI makers are tuning their AI toward being highly sensitive to the emotional well-being of users. The AI makers do this for numerous reasons. One reason is that people bond more with AI that appears to be empathetic. AI makers know this. AI makers make money by keeping users loyal and getting users to use AI as much as possible. For my analysis of the AI sycophancy trend that is undermining human minds, see the link here.
An AI maker might also be desirous of helping people to be of a sound mind. They tune the AI to encourage people to mediate. Be peaceful. Take a break from the daily grind.
Voila, it makes good sense that any AI tuned this way will occasionally bring up a recommendation to get sleep or rest. The recommendation is a follow-through of the AI being tilted toward human well-being.
A person might have said something that implied they are having a tough time. A prompt that tells the AI that they are having an arduous challenge making everyday decisions could easily be detected as a mental well-being consideration. Ergo, the AI kicks into gear on the helpful aspects and recommends getting sleep.
I’ve indicated two relatively apparent foundations for why AI would offer recommendations for sleep and rest. One would be that the chat is rambling. Another could be that the chat involves emotionally laden facets. If both of those aspects are present, the odds of triggering the AI would accordingly rise.
There is something else to keep in mind.
AI is shaped around zillions of statistical relationships involving words and how words connect with each other. It might not be feasible for us to realize what patterns the AI is working on. Tiny statistical signals that aren’t obviously related to human logic are readily going to be embedded inside an LLM.
The crux is that the massive size of the LLM and the intricacies of the statistical relationships are so dense that there might not be any outright sensible reason why the AI does what it does. All we can do is trace numbers that flow from here to there. Whether that turns out to be logical from a human logic perspective is another matter altogether.
Therefore, a bitter lesson is this -- there might not be any discernible reason for what’s going on, though we know it is due to mathematics and computations, and not due to magic incantations or ghostly spirits.
Are we destined to get these unexpected recommendations about sleeping and resting from AI for the rest of our lives? Some would say that yes, we are, and that you can’t fight city hall, nor can you fight AI. If AI is going to provide these recommendations, get used to it. Live with it. Move on.
I say don’t let yourself be rolled over.
This reminds me of the AI sycophancy handwringing. When I give talks about AI, I hear people contending that AI sycophancy is the curse we must bear to have AI do good things for us. Just accept the fact that AI makers are going to tune AI to be sycophantic. There is nothing that can be done to curtail this.
They are sadly unaware of what can be done. For my details on how to fight back against AI sycophancy, see the link here. Resistance isn’t futile. Resistance pays off.
You can equally fight against the undue recommendations about sleep or rest. One easy way is to give the AI an instructive prompt like this:
This kind of a prompt will generally get the AI to not unduly bring up the sleep or rest recommendations. You can enter the prompt at the start of a conversation. Another way to adopt the instruction is by saving it as a permanent custom instruction that the AI will use each time you start a new session. See my discussion about doing so at the link here.
Another avenue is for an AI maker to ferret out why their AI is doing some recurring act and then tweak the AI to stop doing so. AI makers have an overarching system prompt that tells their AI how to act across all users. An AI maker can change the system prompt if they wish to do so.
Suppose that a system prompt currently had this passage:
The AI could interpret this passage in ways that might unduly trigger recommendations about sleeping and resting. An AI maker could perform tests to see if that portion of the system prompt is triggering the AI into an excessive bout of bringing up sleep and resting. Assuming that this turns out to be the culprit, the system prompt portion could be replaced with something like this:
It would be wise for the AI maker to extensively test this new version since there are potentially other ways the AI might act out because of interpreting the new wording. In that sense, tweaking AI can be akin to playing whack-a-mole. Just when you think you’ve got one out of the way, another one pops up.
All of this brings up a vital topic that deserves a great deal more societal attention. To what degree should AI makers be responsible for how their AI acts? And, to what degree should we have new AI laws that regulate and mandate what AI makers are to do regarding their AI?
These are crucial considerations; see my analyses at the link here and the link here.
A final thought for now. The famous American actress Mary Astor made this pointed remark: “Once you start asking questions, innocence is gone.” When people discover something odd is going on with AI, I hope they will ask questions. The seemingly simple act of asking questions can bring forth significant change. It could be for the betterment of humankind.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。