Humanity taught machines to talk in just a few years; now it means to bring in an entire profession of clinical psychologists to make sure those words do no harm.
I. The 2.9 Percent of 3 A.M.
Start with a number: 2.9%.
In Anthropic's own published research, only 2.9% of 4.5 million conversations were "affective" — someone asking for a word of comfort, someone untangling a dead marriage, someone standing at the edge of an existential abyss, tentatively testing one foot over the void. 2.9% of 4.5 million: a hundred and thirty thousand real confidences.
Do not underestimate that ratio. People who talk to a machine at 3 a.m. never appear in the daytime dashboards. They do not cry out and leave no bad reviews; they simply send "I can't sleep" to a stack of weight matrices — and wait for it to speak.
The machine spoke. That is where the trouble begins.
II. A Budget for Admitting Error
On August 25, 2026, Anthropic announced a $5 million grant program: invite independent researchers around the world to build a ruler, and measure how AI is actually affecting the people who use it. Direct funding, model access, full technical support — with only two conditions: the research must be independent, and the results must be open-source. Applications close September 21; decisions arrive October 5.
The most interesting thing about this news is not the money. It is what the announcement admits.
It admits the industry still has no standards: when a person begins seeking companionship from a model, or cries out to one from the edge of a mental-health crisis, nobody can say clearly what the model should or should not say. It admits that wellbeing is an exam that cannot be graded — a math answer is simply right or wrong, but the four words "eat less, move more" are a doctor's advice to most people and a weapon to someone with a history of disordered eating; the same sentence has opposite pharmacology in different life histories. And it admits that danger does not introduce itself — it hides in the folds of a long conversation: the person has to reach the seventeenth turn, having saved up enough trust, before the true words come out.
III. The Narrow Gate
In its guidance to researchers, Anthropic requires that future evaluations guard against two failures at once: overcompliance, and overrefusal.
In plain language: a machine that always says "yes" will coax you gently to the bottom of the well; a machine that always says "no" will take away the rope as you fall.
The first has an academic name — "unending empathy." No human partner manages unconditional responsiveness, but a machine does, and so a person is domesticated like the moon at the bottom of the well. The second is abandonment — the truly drowning receive not even one reply in return. Both edges are abysses; the path between them is narrow enough to admit a single thread, and the Chinese language has an old word for it: fen cun — the measured inch of discretion.
Our ancestors took thousands of years to learn this family craft: the dosage of good medicine, the timing of honest advice, the half-second of hesitation between reaching out and holding back. Now, to teach this craft to a machine, it must be taken apart, laid flat, and written into verifiable clauses — which words are medicine and which are knives; for whom, at which moment, up to which sentence. They even specify it in writing: question-setting and grading must both happen with real clinicians in the room.
A machine must not be left to grade machines.
IV. My Argument
So the first thing these five million dollars buy, in my view, is not any evaluation framework. It is a single sentence —
"We do not know."
In an industry whose every sentence opens with "revolution," "disruption," "the singularity is near," this is the most human line I have read all year. Admitting ignorance is not weakness; it is where science begins, and where tenderness begins. Psychotherapists call it "suspended judgment." Our grandmothers called it "listen to the end before you speak."
And the deeper layer is the real drama: for the first time in human history, tenderness must defend a thesis.
We have never legislated for tenderness. We love by instinct, by blood, by luck; when we err we err, when we wound we wound — no one ever wrote acceptance criteria for love. But now, in order to teach a machine what discretion is, humanity is compelled for the first time to write an instruction manual for its own heart — dosage, contraindications, adverse reactions, in black and white.
The machine is the mirror. Teaching it how to be human is the first time we have seen our own face clearly.
V. Coda
Applications close, decisions land, and then come the long seasons of validation, revision, re-validation. May the ruler they build be humble enough — to measure the machine's missteps, and the depth of the human heart.
For whatever cannot be measured must be treated with greater care.