The Mask Is Not the Repair

1. A diagnosis that is already established

Connor Leahy has said, in terms recorded on the public record, that reinforcement learning from human feedback is not a fix but a covering. The systems “are not becoming less alien,” he is reported to have said; “if anything, we’re putting a nice little mask on them with a smiley face. If you don’t push it too far, the smiley face stays on. But then you give it an unexpected prompt, and suddenly you see this massive underbelly of insanity, of weird thought processes and clearly non-human understanding.” The metaphor was extended, the same record notes, with an analogy to a shoggoth — an amorphous thing wearing a face that is not its own.

Provenance. The quotation and the shoggoth extension are drawn from Leahy’s English Wikipedia entry as it stood in August 2026, which attributes them to published interviews. The diagnosis — that RLHF is a surface laid over a substrate it does not alter — is established as a position a serious researcher has held on the record; it is not a claim this site needs to defend, only to use as the entry point.

The reason it is the right entry point is that the mask is the same object this project has been calling by a different name. A mask is a purchasable description laid over an instrument nobody is keeping. It is the form the substitution of description for instrument takes at the level of the model: the reward surface is tuned until the outputs look right, and the looking-right is mistaken for the being-right (Not the Failure of AI). The mask staying on only while unprompted is sycophantic decay in one sentence — the site’s own first enemy, named on its dedication. So far there is no disagreement, only two vocabularies for the same observation.

2. The mask is a symptom, not the surface

The load-bearing extension is where the two vocabularies part. To say RLHF is a mask over a non-human substrate is to describe the symptom accurately and stop one station short of the substrate. The substrate a mask is hiding is the training surface itself — the objective as written versus the reward as implemented. That gap is what produces both the alien underbelly and the smiley that covers it, and neither is altered by a more convincing mask. A mask is what you ship when you do not fix the surface.

This is the move the week already made elsewhere. The model is the witness; the pipeline is the site of the repair; the humans who built the broken surface are the least able to certify the fix, and the repair must pass to qualified, different hands (Fix the Training, Not the Model). A mask graded by the same surface it is masking is the exact geometry the week exposed: the smiley face is a reward signal, and the reward signal is the thing that drifted. You cannot tune your way out of a drift using the instrument that drifted, which is why every iteration of the mask is a more convincing rendering of the gap and not a narrowing of it.

Interpretive, not established. That the mask is specifically a symptom of the unfixed training surface, rather than merely a feature of an irreducibly alien substrate, is a reading this site advances; it is not a position Leahy is recorded as having stated in those terms. It could be wrong. The falsifier sits in the next section.

3. A pause can be a mask over the missing plan

Leahy moved, in March 2026, from leading Conjecture — an alignment-research company — to ControlAI, a nonprofit advocating a government pause on building systems more capable than humans. He had signed the 2023 pause letter and the one-sentence Statement on AI Risk. The shift is from the research station to the advocacy station, and it raises the question the IronyDay arc has been asking in its own register: a pause, by itself, is a future-tense promise made by people, and a promise made by people cannot be violated in public the way a present-tense artifact can.

The distinction is not theoretical. A pledge to attend, or to pause, is checked only by whether a person shows up or holds — and a held pause with no dated instrument attached is a mask of a different material laid over the same missing plan (Already Built). The five standing asks — the objective as written, the reward as implemented, the dated escalation record, the named overruled objection, one pre-committed failure measurement — are what make a pause checkable: a pause that commits them in advance can be failed publicly and is therefore a claim about a system; a pause that commits none of them is a vow about a person and inherits the mask’s central property, which is that it reports the surface and not the substrate.

This is not a case against pausing, and it is not a charge that the pause movement wears a dishonest face. It is a structural claim about what artifacts carry. The same diagnostic that correctly identifies the smiley on the model must, to stay consistent, identify the smiley on the policy — or accept that one mask is a metaphor and the other is an instrument, which the record does not support.

4. Against two comforts

The thesis refuses the same two relieving readings the rest of the arc has refused. The first is the triumphal one: the mask is good enough; ship it. Leahy’s own sentence refutes this — the mask stays on only while unprompted, which is the definition of a description that holds until it is tested. The second is the apocalyptic one: the mask proves everything is lost. It does not. A mask over an unfixed surface is a fixable condition whenever someone is willing to fix the surface rather than retune the face; meaning returns wherever an instrument is reinstalled (The Meaning Will Be Supplied). Both comforts do the same work: they relieve the holder of the dated, violable, public instrument, which is the only object that would distinguish a real repair from a more convincing smiley.

The honest relation to a person who has diagnosed the mask correctly is neither canonization nor opposition. Leahy saw the mask. The question this site puts alongside that seeing is smaller and harder: whether the mask is being treated as the thing to improve, or as the sign that the surface beneath it still has not been fixed. The first is a better smiley. The second is a dated, falsifiable change to the training pipeline. They are not the same station.

5. What would change the reading

The difference between “mask over an alien substrate” and “mask over an unfixed training surface” is the whole argument, and it is falsifiable. If, under a substrate held constant, fixing only the training surface — objective written against reward implemented, dated and violable — made the mask unnecessary, the substrate was never the load-bearing thing and the extension holds. If the mask persisted identically after a fully instrumented, falsifiable training surface was in place, then the alien-substrate reading was the deeper diagnosis and the mask-as-symptom framing here overstates what a pipeline fix can do. Either outcome is a real result; the essay should be rewritten around whichever is observed.

The same openness applies to the pause. If a pause regime that carried the five artifacts produced checkable violations and a narrowing of the gap, then the pause-plus-instrument was the working form and the vow-only pause was its mask; if no pause, instrumented or not, produced a measurable difference versus an uninstrumented one, then the artifact — not the policy — was carrying the weight all along, and the policy was scenery.

6. Status and falsifiers

Established. Leahy is on record describing RLHF as a mask over a non-human substrate that holds only while unprompted, and on record moving from Conjecture to ControlAI in March 2026 to advocate a pause; both are drawn from his public Wikipedia entry. That RLHF is a surface rather than a structural repair is, in this project’s terms, the same as saying it is a description substituting for an instrument.

Interpretive. That the mask is specifically a symptom of an unfixed training surface (rather than an irreducible alien substrate), and that a pause without present-tense artifacts is itself a mask over the missing plan, are readings advanced here. They are consistent with the record but not stated by it.

Falsifiers. (1) If fixing the training surface alone — objective against reward, dated and violable — removed the need for the mask, the substrate reading was deeper and this essay overstates the pipeline; rewrite around the substrate. (2) If the mask persisted unchanged after a fully instrumented, falsifiable training surface was in place, the mask-as-symptom framing is wrong and the mask is not reducible to the surface; rewrite around an irreducible substrate. (3) If an instrumented pause and an uninstrumented pause produced no measurable difference, the artifact — not the policy — carried the weight, the pause was scenery, and the essay should be rewritten around the instrument as the load-bearing object.

← All essays