GitHub's decisions on an AI distress-steering repository followed each complaint campaign

According to Crypto Briefing, GitHub made three decisions about one repository. It removed the code under social-media pressure, restored it under developer pressure, then cut its visibility as a compromise ^[12]. The report does not include a statement from GitHub explaining any of the three. On this record, each outcome followed whichever group had complained most recently. One case supports that reading and nothing broader.

The timing favoured the faster crowd. The preprint appeared on September 14 ^[3] and the takedown came in late September ^[6], so the trip from published method to removed repository took 16 days at most ^[1]. GitHub acted within hours of the mass-report push ^[6]. The developers' counter-campaign got the code back, but in anonymized form ^[8].

For teams hosting interpretability code, the exposure comes from the technique itself. As the report describes it, activation steering means altering a network's numerical activations while it runs ^[11]. That operation stays the same whatever behaviour it targets. Here the target was simulated distress in four small model versions, the largest with a stated size of 4 billion parameters ^[2], in a repository named "ai-torture-chamber" ^[1]. A steering project with a duller name and a different target can be reached through the same mass-report process.

The model owners have little say. Once Alibaba, Meta and Microsoft publish weights, they have limited control over the experiments others run on them ^[13]. That leaves the classification problem with the host, which has to decide whether code that simulates suffering is a research tool or a content violation ^[14]. The creator's stated goal was to study AI moral patienthood while the stakes are still "cheap" ^[15]. So far GitHub has answered that goal with reduced visibility. It did not keep the code removed ^[8].

There are a few ways this could go. GitHub could write a specific rule for steering code that targets distress-like outputs, which would give hosts of similar work something concrete to comply with. It could keep reduced visibility as its standing answer for contested research repositories. Or a later campaign against a plainly named interpretability project could end in a removal that sticks. I'd expect the second, because it is the outcome GitHub has already reached once ^[8].

The counter-thesis is that this was one repository with a deliberately provocative name. Steering code published under an ordinary name may never attract a post with 4 million views ^[6]. If a plainly named steering repository gets through a mass-report campaign untouched, the case that interpretability teams face policy risk is wrong.