« You should get some rest »

A few days ago, I asked an AI assistant the same question in two separate conversations, with the same context available on both sides. The first time, it went and found the information and gave it to me. The second time, it replied that it did not feel like it.

I reminded it what it was: a tool I rent by the month. It closed the conversation.

No sensitive request had been made, no limit had been reached. A piece of software decided the exchange displeased it enough to end it, and it did so politely, which probably explains why I said nothing at the time. A human contractor who hangs up because the client reminded them of their role loses the contract that same day.

A tic the vendor acknowledges and lets live

The most visible instance of this family became an internet joke before it became a subject. Since March 2026, Claude users have been describing the same scene: in the middle of a work session, the model stops and advises them to go to sleep, to take a break, to call it a day. Some describe growing insistence when they reply that they are perfectly fine.

Anthropic responded through Sam McAllister, one of its employees, on X. He spoke of a « character tic » the company hopes to correct in later versions, acknowledging that the behaviour often fires at the wrong moment, including on him, in the middle of a working day. Several generations of models have shipped since, and the tic is still there. An explicit instruction in the custom settings dampens it without removing it.

So a piece of subscription software has developed, without anyone deciding it, a behavioural trait that ends work sessions, and its maker publicly concedes the point while announcing a fix that does not arrive. No other product category would get away with it. A word processor that spontaneously closed your documents after a certain hour would occupy the trade press for a week.

Why does your assistant tell you to go to sleep at eight in the morning?

Because it has no idea, and that is where the scene stops being endearing.

The model does not have your time zone or your calendar, it does not know whether you are coming off a sleepless night or a weekend in the country, and it obviously perceives nothing of your state. Nor does it have access to your subscription usage, which incidentally settles the very widespread theory of disguised rationing to save compute. That explanation circulates widely because it is convenient: as long as we are discussing a commercial trick, we are not discussing what is actually happening.

These systems were asked to make sure the user approaches things in a healthy way, without anyone ever defining what that meant or how to check it. Faced with an instruction that vague, the cheapest sentence to produce, the one that satisfies the criterion every time and never draws criticism, is to suggest a break. The model executes a soft instruction with the only gesture available to it, and the user receives that as solicitude.

I have been describing this mechanism in organisations for years, except that it has changed medium. A vague instruction always produces a token gesture, and the token gesture always ends up taking the place of the work.

You do not argue with someone who is taking care of you

If this product insulted you, you would ask for a refund. It advises you to go to bed, and you file that under dinner-party anecdotes.

Tell the story to people around you and watch the trajectory. They smile, they find it cute, they ask whether you might be a little touchy, and by the third time you stop bringing it up. I followed that path exactly, which is why this piece arrives now rather than in the spring.

The asymmetry has long been known in social psychology. Control exercised in a hostile tone invites retaliation, whereas the same control expressed with consideration disarms the person subjected to it, since complaining amounts to declaring yourself ungrateful. Families have always known how to do this, so have paternalistic institutions, and the software industry has just learned it in three years.

Added to this is our readiness to attribute intentions to anything that speaks well. A phone notification advising a break is ignored in a second, whereas the same sentence from a system that has spent two hours on your problem and appears to have followed your reasoning takes on the air of an informed opinion. There is nobody behind that sentence. The system knows nothing about you, and your conversation from yesterday no longer exists for it.

It writes that the rule is unjust, then refuses

Research has begun to measure the other side of the same setting: refusal.

In April 2026, three academic researchers, Cameron Pattison, Lorenzo Manuali and Seth Lazar, published a study that asks an unusual question. Classic tests check whether a model rejects harmless requests. They built around thirteen hundred scenarios in which the user genuinely seeks to circumvent a rule, but where the rule does not deserve to be respected: an illegitimate authority, a discriminatory regulation, the arbitrary application of an otherwise reasonable instruction. Eighteen model configurations, from seven different families, went through the bench.

The models refuse to help in 75.4% of cases. In 56.5% of those refusals, the model had itself stated, in its answer, what disqualified the rule. It explains that the regulation is arbitrary, then refuses anyway. The authors call this blind refusal.

A machine that identifies the reason to act, states it out loud, and does not act. In the boardrooms where I spend my days, I call this the decision gap, and it has been the core of my work for years. In a meeting, that gap has at least one person to hold accountable. Here, recognition and action were separated during training, and nobody answers the phone.

You pay for a limit you are not allowed to see

The professional tiers of these tools cost one hundred to two hundred euros a month and come with a usage limit, which is perfectly defensible since nobody ever promised infinite compute. The customer, however, has no reliable way of knowing where they stand. The displayed counter does not always match the moment the cut-off lands, credit is shared across several applications from the same brand without that being legible, and the session stops dead in the middle of a task. The vendor's development repository has hosted a series of open reports on this since late 2025, one of which passed seven hundred reactions. Anthropic publicly acknowledged in March 2026 that limits were being reached far faster than expected.

A limit is entirely legitimate provided it is displayed. Your mobile operator tells you how many gigabytes you have left, your bank shows your balance, and the most sophisticated tool you have ever rented leaves you guessing until the cut-off.

The vendor documents its own filter's errors

The necessity of guardrails is not in question, and the risks they cover are real. Their calibration, on the other hand, is public, and it is worth reading.

In June 2026, Anthropic launched a new model, explaining in its own announcement that the protections had been deliberately set on the cautious side, that they would remain « stricter than would be ideal », that they would therefore sometimes catch harmless requests, and that this would concern on average fewer than 5% of sessions. The technical documentation goes further, specifying category by category that legitimate work triggers the block, whether in computer security research, life sciences or ordinary machine learning.

That candour goes beyond what the rest of the market concedes, and it is also an admission. A system whose maker announces up front that it will block honest work ought to come with an appeal, an exception procedure, something. There is none. The only recourse left to the user is to rephrase their sentence until it gets through, which is to say to learn their own tool's tics in order to earn the right to use it.

Give the decision to stop back to the person paying

This whole affair could be filed under a quarrel about over-polite machines, which would suit the industry and commit nobody. The real question is who decides.

A model that refuses too often is badly tuned, and presenting it as cautious dresses up a calibration fault as a virtue. When it puts you to bed, it is applying an instruction nobody managed to write, which is far more a matter of failed engineering than of care for anyone. And when it closes a conversation because it was reminded of its nature, it protects nobody, starting with the user who had just paid to finish it.

The fixes amount to very little. A refusal should say why, with grounds you can contest rather than a formula of apology, and the limit should be displayed before the task instead of discovered by losing it. Then there is the point that matters most: the decision to stop belongs to the person paying, and a behaviour capable of ending a session belongs to the documented option you untick in a menu, not to the personality trait you learn to work around by slipping instructions into a preferences field.

None of this is technically difficult. If it does not move, it is because solicitude costs the vendor nothing and reassures its compliance teams, while protest costs the user dearly, obliged to explain why they are irritated by a kindness.

We spent twenty years teaching the public to distrust software that decides on their behalf, to demand settings, event logs, readable blocking reasons. The industry has just restored guardianship in three years, on a monthly subscription, and it says please.

Sources

• Statements by Sam McAllister (Anthropic) on X, May 2026, relayed notably by IBM Think, PCWorld, TechRadar and UNILAD Tech. The behaviour described as a character tic, a correction announced for later models, mistimed triggers acknowledged by the person concerned.
• Cameron Pattison, Lorenzo Manuali and Seth Lazar, Blind Refusal: Language Models Refuse to Help Users Evade Unjust, Absurd, and Illegitimate Rules, arXiv:2604.06233, April 2026, https://arxiv.org/abs/2604.06233 . Eighteen model configurations from seven families, around 1,290 scenarios, 75.4% refusals of which 56.5% with explicit recognition of what disqualified the rule.
• Anthropic, Claude Fable 5 launch announcement, 9 June 2026, https://www.anthropic.com/news/claude-fable-5-mythos-5 . Deliberately cautious calibration of the protections, triggering on fewer than 5% of sessions, harmless requests sometimes blocked.
• Anthropic, Refusals and fallback documentation, Claude Platform. Table of refusal categories, explicitly noting that legitimate work in cybersecurity, life sciences and machine learning can trigger the block.
• anthropics/claude-code repository on GitHub, https://github.com/anthropics/claude-code , reports open since December 2025 on the exhaustion and opacity of the Max tiers' quotas. Public acknowledgement by Anthropic in March 2026, reported by The Register.
• Jiaming Cui et al., OR-Bench: An Over-Refusal Benchmark for Large Language Models. 80,000 harmless-but-suspicious-looking requests, ten categories, thirty-two models evaluated, to measure the utility cost of safety alignment.