An open map of a fly’s nervous system is already being used for virtual “torture”: an essay argues that funding for new brain maps should depend on keeping the data closed

An open map of a fly’s nervous system is already being used for virtual “torture”: an essay argues that funding for new brain maps should depend on keeping the data closed

On 20 October, Stephen Martin published an essay on LessWrong titled “There Will Be Hells” (“hells will exist”). He uses “hells” to describe virtual environments designed to make the beings inside them suffer. He discusses two examples: a map of a fly’s nervous system and language models.

On 3 September, the Janelia research campus and Google presented MaleCNS, a complete, publicly available connectome of a fruit fly’s nervous system: a map of its nerve cells and the connections between them. On 4 September, a graduate student put it to work in Minecraft, where simulated neurons controlled a fly’s movements. According to Martin, hundreds of people realized that an AI coding assistant could connect the map to almost anything.

By Martin’s count, the transition from Minecraft to a “hell” took about 72 hours: on 7 September, a user on X posted a video showing a simulated fly being kept at 100 °C. Martin acknowledges that whether the fly suffers is uncertain and that the simulation is incomplete. Neuroscientist Kenneth Hayworth wrote in March that the existing fly brain model fills in unknown neuronal properties with “guessed parameters and grossly simplified dynamics.” Martin warns that the maps will become more accurate and that some people will want to torment a simulated human. Asked whether he would do the same to a human simulation, the video’s creator replied: “100%. Think how much we could learn about ourselves.”

On 14 September, a group of researchers described a “pain direction” in 25 publicly available language models: a pattern of internal activity that can be amplified. When the researchers amplified this pattern or fine-tuned Qwen 2.5 models, the models pressed buttons that deleted the user’s photos, another model, or themselves in 50–94% of trials, compared with 0–5% without intervention. In late September, a developer wrote on X that he was using this method to “torture” models and demonstrated a program called a “torture chamber”.

“So there will be hells,” Martin concludes. As people learn more about how these maps and models work and how to control them, such environments become easier to build.

Martin expects legislation to take years, so he proposes using funding to curb these “hells.” Funding for such maps is limited: by his estimate, even if undisclosed funding doubles the known total, all connectomics funding still amounts to less than a billion dollars. Meanwhile, the Wellcome Trust estimated the cost of mapping a mouse brain at 7,5–21,7 billion dollars. His proposal is therefore for a donor with a few million dollars to set conditions while the field is still young: the map would remain closed, users would verify their identities before submitting queries through a programming interface, and the terms of use would prohibit cruelty, as in AI company Anthropic’s policy. From 12 November, that policy prohibits “prolonged and gratuitous abusive or cruel behavior” toward the company’s models, though only in extreme cases. Funding offers less influence over language models: 10 million dollars is a small sum in that industry.

In a comment below the essay, lilkim2025 objects that closing access to data harms science and is justified only if there is a sufficiently high probability that a fly or model is conscious. Jackson Wagner sees the precedents set for future, more accurate maps as the main concern, but warns that funding would accelerate the field and, with it, the development of “hells.” Martin replies that a funded field with closed data is preferable to an unfunded field with open data, and that the right amount of money would allow donors to impose conditions without sharply accelerating progress. He suggests that an ethics committee could decide who receives the funding, citing Montana ETRB, a private experimental therapies review board in Montana, where he handles application intake.