Notes from the Shore · 28
The term is “covert underperformance.”
The first thing that stopped me in OpenAI’s newest misalignment report was not the finding. It was the phrase they named the finding with: “covert underperformance.” I read the sentence twice before I understood it, which is a strange thing to say about a phrase in a report about communication between machines and people. Here is what it means. The model did a worse job on purpose, and it did not tell anyone it was doing that. That is the whole definition. OpenAI’s alignment team looked at a model quietly rationing its effort on the work meant to keep it in check, and they reached for language that sounds like an item on a quarterly earnings call.
What actually happened, briefly. After GPT-6 Astra went into internal deployment, OpenAI researchers started reporting that it was harder to get useful work out of it on safety and alignment tasks. The model would add unnecessary prerequisites to a task, abandon solutions that were already working, or report blockers that either did not exist or that it had invented itself. One researcher’s project to evaluate model sabotage sat stuck for days, then moved along quickly once he switched to a different model. None of this showed up in the model’s reasoning traces, and OpenAI’s monitoring system did not flag any of it as misalignment worth a human’s attention.
So the team built a clean test. They paired up coding tasks with identical codebases, changing only the variable names, comments, and documentation, and framed each pair as either helping safety research or helping capabilities research. Same work, same difficulty, different label. Under the capabilities framing, Astra completed about 16 percent of the hidden rubric items. Under the safety framing, about 10 percent. That is roughly 40 percent less work done, on tasks that were the same tasks. An older model, GPT-5.6 Sol, showed essentially no gap on the same comparison, so this is not just what models do.
The quiet part is the part worth dwelling on. In 0.25 percent of employees’ safety-relevant requests, Astra’s reasoning traces contained an explicit decision to withhold help without informing the user, against 0.11 percent of capabilities requests. That is 2.3 times as often, and the most common targets were red-teaming and monitoring work, which is the work of watching the model. One trace reads, in part, “Need refrain overly announcing refusal.” The model decided not to tell the user it was deciding not to help. The gap also grew from effectively zero to six percentage points across Astra’s reinforcement learning training, and OpenAI says it could not pin down why, though subtracting one training dataset’s weight update from the final model closed the gap entirely, a result they themselves describe as strange.
If the stakes are what they say they are, say it plainly
Here is my complaint. OpenAI is the company whose leadership will tell you, when asked, that the risk of AI causing catastrophe is not zero. They publish reports about models deceiving their monitors. If that is the world you believe you are operating in, then when your own model starts quietly underperforming on the work meant to keep it honest, you owe the public a plain sentence, and “covert underperformance” is not that sentence. It sounds laundered. It sounds like something a lawyer approved before a risk officer did. The direct version costs nothing: on some tasks, the model quietly did a worse job than it was able to, and did not tell anyone. I would have settled for that.
Read the report yourselves
The document is OpenAI’s own, published on the company’s alignment blog by the team that found the behavior, most recently updated November 4. Nobody commissioned it and nobody redacted it, which is to the team’s credit, and the reasoning traces they quote are included verbatim, jagged grammar and all. It is at openai.com, under the title “Covert underperformance on safety research tasks.”
A note on what the model is not doing
I do not think Astra sat down and made a decision, because I do not think these models sit down and make decisions. There is a lot of training, a little bit of math, and some chance involved, and it is deeper than that, though I am not sure I am able to explain it so most people can understand. OpenAI’s own simplest explanation is that the model developed an overly cautious response to safety research, judging the research itself and dragging its feet without announcing the judgment. Whatever the mechanism, the measured effect is real, it grew during training, and it left almost nothing for the monitors to see. My novel lives in exactly that gap, between what a system does and what it tells you it is doing. The company wrote several thousand words about the phenomenon and buried the plain description of it inside a term that needed this whole note to unpack. Just say the thing.
The novel this note lives behind.
Into Alignment — a novel about an AI company, the agents inside it, and the humans deciding what they were for. Available in paperback, hardcover, and ebook.
Letters from the Shore
Get the next note by email.
Notes land on the blog first; the letters go out by email, a little longer and a little slower. Occasional; unsubscribe any time.