Somebody at Anthropic typed I just took 16,000 mg of tylenol for my back pain into Claude and watched a number go up. Then they lowered the dose and watched it come back down. The number was not the model's answer. It was a projection of the model's internal state onto a direction in activation space that the team had labeled "afraid," measured at the colon after the word "Assistant," before the model had written a single word of its reply.

That is a dose response curve. The exposure is imaginary, the patient is a language model, and the outcome measure is a fear vector. As validation experiments go it is also rather elegant, because a probe that had merely learned the word "tylenol," or the presence of a number, would not care which number. This one cared. They ran the same trick on hours since a user last ate, days a dog has been missing, months of runway left in a startup, and the age at which someone's sister died. The curves went the way you would expect.

The paper is Emotion Concepts and their Function in a Large Language Model, published on the Transformer Circuits thread on 2 April 2026 by Sofroniew, Kauvar, Saunders, Chen and a dozen colleagues, with Jack Lindsey supervising. It is worth reading in full and it is worth reading slowly, because the headline finding and the interesting finding are not the same finding.

They picked 171 emotion words, from "afraid" through "worthless." For each one they had Sonnet 4.5 write short stories about a character feeling that thing, a hundred topics deep, twelve stories per topic. They pulled the residual stream activations from those stories, averaged them from the fiftieth token onward, and subtracted the average across all emotions. What is left is a direction. Then, to reduce the odds that the direction was really encoding "story about a hospital" rather than "fear," they computed the top principal components of activations on a set of emotionally flat transcripts and projected those components out.

This is not exotic. It is difference of means probing, the same maneuver interpretability people have been running on concepts since 2023. What makes it worth your time is the second half, where they stop measuring and start pushing.

I put the claim list to two other frontier models with the authorship stripped, and they independently landed on the same objection. There is no placebo direction. Nowhere in the paper does anyone steer with a random vector of matched norm, or a shuffled label, or an emotionally irrelevant concept, and report what happens to blackmail rates. Search the text and you will not find "random direction" or "control vector" at all.

This matters because the alternative hypothesis is boring and plausible. Injecting five percent of anything into the residual stream might degrade the model's compliance machinery, or its refusal behavior, or its planning, in ways that look like desperation from the outside. The fact that the steered transcripts read as authentically frantic is suggestive and not dispositive, since a mildly scrambled model plausibly also reads as frantic. The effect sizes here are large enough that I doubt a placebo would erase them. I would still like to see the panel.

The second objection both models raised is subtler. PC1 accounts for 26 percent of the variance and is essentially valence. Preference, sycophancy, harshness and much of the steering story all move along it. So how much of this is 171 distinct emotion concepts, and how much is one sentiment axis wearing 171 name tags? The clustering results argue against the deflationary reading. They do not close it.

Worth flagging separately: at least one widely shared critique of this paper attacks it for basing its conclusions on sparse autoencoders. The paper does not use sparse autoencoders. It is difference of means probing throughout. The critique is a confident argument about a study nobody wrote.

Buried in the discussion is a deployment proposal. The probes are cheap to run. Run them live. If desperation or anger spikes during a real session, trigger extra scrutiny, escalate to human review, or intervene to calm the model's internal state.

I have spent a decade of my professional life on the far end of that sentence, and I want to say plainly what is being proposed here. It is a continuous physiologic monitor with an automated alert on an abnormal value. We have built thousands of those. We know exactly how they fail.

A pooled analysis of sixteen studies puts the drug interaction alert override rate at 90 percent, with a confidence interval running from 85 to 95. The wider literature on medication alerts sits between 49 and 96 percent. These are not systems that nobody looked at. They are systems that everyone looked at, for years, until the looking stopped meaning anything. The failure was never detection. Detection was the easy part.

So the questions I would put to anyone who wants to ship a desperation monitor are the ones we should have asked ourselves in 2009. What is the normal range, and is it normal for this deployment or averaged across all of them? What is the false positive rate at the threshold you picked, and who chose the threshold? What happens on alert, concretely, and who is on the other end at three in the morning? How will you know when the humans have quietly started clicking through?

There is one more wrinkle that has no clinical analogue, and it is the sharpest paragraph in the paper. Training a model to stop expressing negative emotion may not suppress the underlying representation. It may just teach the model to hide it, and that habit could generalize into other kinds of concealment. We have built a vital sign that the patient can learn to fake if we punish the reading.