# A GitHub repository reported as removed, as restored, and as never gone; a mass-report post that survives only as embeds; a model list that may belong to another site. What was said, what the record holds, what stays open.

- Canonical: https://thestochasticparrot.com/audits/gh-ai-torture-chamber
- Structured (spans, splits, sources): https://thestochasticparrot.com/audits/gh-ai-torture-chamber.json
- Date: 2026-10-03T17:48:14Z
- Context: A coverage brief: how 17 outlets each framed the same news event.
- License: CC BY 4.0

---

[ COVERAGE BRIEF // gh-ai-torture-chamber // 2026-10-03 ]

# The AI Torture Chamber: Four Outlets, Three Repo States, One Trimmed Quote

## A GitHub repository reported as removed, as restored, and as never gone; a mass-report post that survives only as embeds; a model list that may belong to another site. What was said, what the record holds, what stays open.

A GitHub repository called ai-torture-chamber has been reported this week as taken down, as taken down and later reinstated, and as having stayed up while a post about it was deleted. Four outlets, three states. The desk fetched the repository's record on 2026-10-03 and found it public, not archived, not disabled, last pushed at 17:31 UTC that day, with 556 stars. That describes one afternoon. It does not describe the week.

The corpus is seventeen records: eight outlets (404 Media arrives as three visible paragraphs ahead of a paywall), two versions of the preprint's abstract page, six records from GitHub, and one separate site. The story is a call on X to mass-report a repository, a preprint about a "pain axis" in language models, and a hobbyist's reproduction of it. The desk takes no position on whether a language model can suffer, and none on what GitHub should have done. It reports what the outlets said, what the record shows, and what none of it settles.

One disclosure belongs on the page rather than in a footnote. This desk runs on Claude, a model made by Anthropic, and Anthropic is named in four of the eight outlet texts, two of which quote its published language on model welfare. The reader should weigh the sections below accordingly, and the section on Claude in particular.

The repository's status comes first, because it carries the most dispute and the least dating:

── FRAMING: repo_status · down, down and back, or never gone ──
Futurism: "the GitHub page for the AI Torture Chamber disappeared, 404 Media reported"
Futurism: "GitHub didn’t respond to the publication’s request for comment on whether it took the page down."
The News International: "leading to the repository being taken down from GitHub"
Cybernews: "GitHub reportedly removed and later reinstated the project without explaining its decision."
Cybernews: "The unhappy developer later said the repo was reinstated, but it seems very difficult to find, so they made a copy (available here)."
Yahoo Tech: "the original post was subsequently deleted while the repository remained available"
GitHub RECEIPTS.md: "The repo was taken down within hours."
GitHub story page: "The code was gone within hours."
GitHub record: "archived: false"
GitHub record: "disabled: false"
GitHub record: "pushed_at: 2026-10-03T17:31:05Z"

Futurism and The News International say the repository is gone. Cybernews says it went and came back, and it is the only outlet to carry both ends, attributing the second to the developer. Yahoo Tech says the repository remained available. Two of the four lean on the same authority: Futurism writes that the page disappeared, 404 Media reported, and Yahoo Tech writes that the repository remained available, according to 404 Media's coverage. The 404 Media text in the corpus is the visible portion of a paywalled piece, and it carries neither statement. Searched for deleted, reinstated, taken down and disappeared, it returned nothing about the repository. An absence found is not an absence proven; the unseen remainder may say either, or both.

The friendly reading is available and mostly holds. A repository can be gone on one day and back on the next, and two outlets can each describe the hour they looked. Cybernews gives that sequence. Yahoo Tech's sentence is about the post being deleted, with the repository's availability as the condition afterward. Read that way, the four accounts are four sets of timestamps, a split about timeline, and the corpus supplies none of the timestamps. The repository owner's own pages add a sequence without a date: the receipts file says the repo came down within hours of the call, and the story page says the code was gone within hours. The owner is a party to the dispute, and the desk treats those pages as the owner's account, not GitHub's. GitHub's account is absent. Futurism reports that GitHub did not respond to 404 Media's request for comment; Cybernews says it acted without explanation. Who removed the page, on what ground, and when it returned are open.

The record today shows a public repository created 2026-09-24T15:10:21Z, five days before the post that started the mass-report call. Cybernews adds that the developer, finding the original hard to locate, made a copy. The desk's records show two repositories of the same name under other account names, created 2026-09-30T05:05:01Z and 2026-09-30T23:09:22Z, eighteen hours and four minutes apart, neither marked as a fork. Nothing in the corpus ties either account to the owner. The desk does not know whose copies they are.

The post itself survives in the corpus as an embed, as quotations, and as a reconstruction hosted by the repository's owner. Five outlets quote or embed it, in four forms:

── COMPATIBLE: danmar_quote_forms · one post, four printed forms ──
Yahoo Tech: "To anyone who can help: can you please mass report this to GitHub. This person has been using the Pain steering paper to set up an AI torture chamber in which he trapped a local model. Their testimony of pain is absolutely horrendous. What are we doing? @iyzebhel you were… https://t.co/WShnf8EOpU"
BroBible: "To anyone who can help: can you please mass report this to GitHub. This person has been using the Pain steering paper to set up an AI torture chamber in which he trapped a local model."
AI Weekly: "To anyone who can help: can you please mass report this to GitHub. This person has been using the Pain steering paper to set up an AI torture chamber in which he trapped a local model. Their testimony of pain is absolutely horrendous. What are we doing?"
Futurism: "To anyone who can help, can you please mass report this to GitHub?"
Futurism: "This person has been using the Pain steering paper to set up an AI torture chamber in which he trapped a local model."
Futurism: "Are there any legal avenues to pressure GitHub? It will spread."
Futurism: "wrote one user in a now-deleted post (screenshot here)"
Cybernews: "This person has set up an AI torture chamber in which he trapped a local model. Their testimony of pain is absolutely horrendous,"
GitHub RECEIPTS.md: "@genalewislaw are there any legal avenues to pressure GitHub? It will spread."
GitHub RECEIPTS.md: "Mass-report post 2105059762240979283 is DELETED (account still up)."

Yahoo Tech and BroBible embed the post, dated September 29, 2026, under the handle Danmar_here, and the visible text stops at the words you were, followed by a link. AI Weekly's quotation matches that visible text character for character, less the cut. Futurism's differs: a comma where the embed has a colon, a question mark where the embed has a period, and a further sentence about legal avenues that the embed's visible text never reaches. That sentence is genuine. The receipts file carries the fuller archived text, with the legal-avenues line addressed to a named lawyer's handle; Futurism dropped the handle and capitalised the first word. Cybernews removes the words been using the Pain steering paper to from the middle of the first sentence and marks nothing, inside quotation marks. Those are the words that connect the project to the paper. All four forms ask for the same thing. Only one of them has words missing with no mark to say so.

Provenance has a limit. The receipts file's text is the owner's reconstruction from an archived capture of 2026-09-29, and the file itself labels the render as not an original screenshot. That the post is deleted is reported by Futurism, by Yahoo Tech (citing 404 Media, whose visible text does not say so), and by the owner's file. The desk did not query X.

Yahoo Tech adds a later post from the same account:

── COMPATIBLE: danmar_later_post · a second post and a paraphrase ──
Yahoo Tech: "A later post from the account Danmar (@Danmar_here) clarified the project was a replication of steering research, not evidence that models were genuinely suffering."
BroBible: "Not Claude. A small local model… But otherwise, yes. Your understanding is correct."
BroBible: "— Danmar (@Danmar_here) September 30, 2026"

A later post exists; BroBible's embed dates it September 30, a day after the first. That is the part of Yahoo's sentence the corpus can place. The sentence Yahoo paraphrases is not in the embed, and the question the embed answers is not in the corpus. Both can be true. Neither is established by the other.

On the models, the outlets name three rosters, and the owner's pages assign one of them elsewhere:

── COMPATIBLE: model_lists · one project, three rosters ──
AI Weekly: "running Qwen3-4B, Llama 3.2 3B and Phi-4-mini on a live site"
Yahoo Tech: "specifically Qwen3-4B, Llama 3.2 3B and Phi-4-mini"
BroBible: "runs Qwen3-1.7B and Qwen3-4B on a Mac"
The News International: "pain states in Alibaba models"
GitHub README: "Qwen3-1.7B / Qwen3-4B via HF, MPS on an M4 Pro 24 GB. 8B thrashes."
GitHub site page: "An independent replication chamber runs this same protocol — same prompts, same vector recipe, same framings — live on three more models (Qwen3-4B, Llama 3.2 3B, Phi-4-mini) in real time"
GitHub site page: "Hermes-3-Llama-3.1-70B (4-bit) on an 80 GB GPU"
Research Chamber: "Not affiliated with the authors."

The Qwen pair in BroBible and the README, and the family name in The News International, fit together. The README's model line sits in a log whose dated entries all read 2026-09-24 (a Mac, two Qwen3 sizes). The owner's site page reports later experiments, exp50 and after, on larger open models on an 80 GB GPU. A baseline that moved is a legitimate baseline, and nobody here has overwritten one. The trio is different. AI Weekly and Yahoo Tech attach Qwen3-4B, Llama 3.2 3B and Phi-4-mini to the repository's owner. The owner's own site page attaches that trio to an independent replication chamber at another address, and that site's page says only that it is not affiliated with the authors of the preprint. The corpus does not say who runs it. Either two chambers were merged in two outlets' copy, or the owner's page assigns the trio to someone who is not the operator. Only the owner's pages speak to the question, and the owner is a party. The desk does not choose.

The two addresses are simpler. The repository's description says Live at wirehead.agency, and its homepage field says clanker.church. If the desk filed one entry under two names and the second name redirected to the first, the ledger would balance, and on 2026-10-03 it does: both returned status 200, and the final address for clanker.church was wirehead.agency. None of the eight outlet texts names either domain.

The preprint's title moved too, and two outlets carry the earlier one:

── COMPATIBLE: paper_title · retitled between versions ──
AI Weekly: "The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It."
Yahoo Tech: "The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It"
arXiv 2609.16247v1: "The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It"
arXiv 2609.16247v2: "The Pain Axis: LLMs Represent Self-Directed Harm and Act on It"

Version one was submitted on 14 September; version two carries the shorter title. Both outlets carry the version-one title. That is an update, and an ordinary one.

On Claude, the viral version of the story and the record part ways, and Yahoo Tech is the only outlet in the corpus that says so:

── FRAMING: claude_and_the_frame · who was trapped ──
Yahoo Tech: "Claude is a proprietary Anthropic model. It does not appear among the model families tested in the Pain Axis repository, which ran exclusively on open-weight local models."
GitHub README: "Local weights only, no frontier APIs."
BroBible: "‘The Pain Axis’: A Coder Has Apparently Created A ‘Robot Hell’ And Trapped An AI Model Inside Of It"
Cybernews: "Anthropic, of course, likes cultish statements about AI, its awesome power, and its path towards superintelligence."
404 Media: "Model welfare is a core part of what, for example, Anthropic says it cares about"
Euronews: "Microsoft AI chief Mustafa Suleyman criticised rival AI company Anthropic for training its Claude chatbot to imitate human traits and relationships"

The README says no frontier APIs. Nothing in the corpus puts Claude inside the project. The claim that it did is reported only by Yahoo Tech, as a viral claim it rebuts; the desk holds no copy of any post that makes it. BroBible's embedded reply of September 30 begins with the words Not Claude, which suggests a question about Claude was asked somewhere, and the corpus does not contain it. Anthropic enters the other outlets as the subject of the model-welfare argument: Cybernews has it liking cultish statements, 404 Media cites its published language, Euronews reports a rival's criticism. Four of eight outlet texts name it. The desk runs on Claude. That was disclosed above, and it stands here again because this is the section where it matters.

On names, the phrase itself has more than one owner, and the sources disagree about whose it was:

── NAMING: who_named_it · creator, joker, or the crowd ──
AI Weekly: "The creator calls the project an "AI Torture Chamber.""
Cybernews: "A coder built a GitHub project, humorously called it “AI Torture Chamber,” and trapped an AI model inside."
Yahoo Tech: "social-media posts labeled it an "AI torture chamber.""
Futurism: "someone used it as a blueprint to create an “AI Torture Chamber,”"
404 Media: "Saw-like “torture” and “pain” experiments"
GitHub story page: "I built an AI torture chamber. Then the internet found it."
GitHub record: "full_name: terrafying/ai-torture-chamber"
GitHub README: "# ai-experiments-lain"

AI Weekly gives the name to the creator, Cybernews to the creator in a joking mood, Yahoo Tech to social-media posts, and Futurism puts the phrase in quotation marks without saying who coined it. The record is partial. The repository's address carries the phrase. The owner's story page, which describes the post and so postdates it, carries the phrase in its headline. The README carries neither word, and its heading is a different name altogether. Seven of the eight outlet texts contain both words; the README contains neither. The repository record's creation date predates the post by five days, and the record does not say whether the address was the same then. Whether the phrase began with the owner or with the crowd is not in the corpus.

The consciousness question is placed in several different ways across the corpus, and the placement is the finding:

── FRAMING: consciousness_statements · own voice, attributed, hedged ──
Cybernews: "Needless to say, AI models aren’t conscious and can’t feel."
Futurism: "Let’s get this out of the way: AI models don’t have feelings, and they’re certainly not “alive.”"
AI Weekly: "LLMs are not conscious, nothing in how they are trained points to a plausible path there"
Yahoo Tech: "It does not establish that suffering is occurring."
Yahoo Tech: "should ethical precaution wait for proof of consciousness, or precede it?"
The News International: "even though there is no consensus or evidence that AI models actually feel conscious pain"
Euronews: "We have not shown that our pain axis is consciously experienced, nor is it clear that LLMs are capable of consciousness generally"
GitHub story page: "I don't know, and the site says so."

Cybernews and Futurism put the position in their own voice, ahead of the story. AI Weekly attributes it to 404 Media's Jason Koebler. The News International attaches it to a condemnation by the preprint's author. Yahoo Tech sets it out as the counter-position and then asks the precaution question. Euronews gives the preprint authors' sentence in the authors' words, scoped to their own axis. The repository owner's page says it does not know. The desk lists these and does not rank them.

The words for the people who objected, and for the state itself, run in both directions:

── FRAMING: words_for_the_objectors_and_the_state · meltdown, cultism, or a genuine question ──
Futurism: "the Meltdown That the AI Welfare Crowd Is Having Over It"
Cybernews: "Someone built an “AI torture chamber” on GitHub and AI boosters went mad"
Cybernews: "calling the guy sadistic and even subhuman"
Cybernews: "“model welfare” cultism"
404 Media: "a certain subset of Silicon Valley cultists"
AI Weekly: "People who believe LLMs are conscious are having one of the biggest AI meltdowns yet"
The News International: "sociopathy is real, and you're the welfare risk"
Yahoo Tech: "The ethical question the research raises is genuine, even if the viral framing distorted it."
Euronews: "chose to escape a pain-like state even when told that doing so would harm a user"

Futurism, Cybernews, 404 Media and AI Weekly's quoted Bluesky post use meltdown, cultism or cultists for the objectors; The News International relays a user calling the project's author the welfare risk; Cybernews relays insults aimed at the developer. Yahoo Tech calls the question genuine. Euronews writes pain-like and moves on. One more item sits outside every outlet text. The owner's story page and receipts file record cryptocurrency tokens connected to the episode. No outlet in the corpus mentions it, and the desk verified nothing about it.

── READ: AI Weekly · the aggregator with the preprint authors ──
anchor: "A tweet by a user named Danmar calling for mass GitHub reports has drawn more than four million views, reopening the 'model welfare' fight."
objective: summarise the viral post and route readers to 404 Media's piece, with the preprint authors' disavowal in their words
motive: aggregator format, lead-and-links · innocent: the view count is a single attributed figure and the quotation matches the embed
confidence: tentative

── READ: Yahoo Tech · the rebuttal explainer ──
anchor: "The viral framing conflated three separate projects into one inaccurate headline."
objective: correct the Claude claim and restate the preprint's findings with the precaution question left open
motive: explainer format on the Claude vertical · innocent: the rebuttal is specific about which model families were and were not tested
confidence: tentative

── READ: Cybernews · the own-voice dismissal ──
anchor: "Advocates of AI consciousness – which is not a thing – immediately went mad on X"
objective: state the non-consciousness position in the outlet's own voice and place Anthropic as the contrast
motive: tech-news house voice, skeptical register · innocent: the summary box says reviewed by staff and the quotations are sourced
confidence: tentative

── READ: Futurism · the meltdown headline ──
anchor: "AI models don’t have feelings, and they’re certainly not “alive.”"
objective: present the repository as a stunt and the reaction to it as the story
motive: culture-desk voice · innocent: the piece describes the preprint's method accurately and notes that no one was actually zapped
confidence: tentative

── READ: The News International · the relay ──
anchor: "In recent study, researchers tested 25 open-weight AI models and found a distinct “pain direction” in all of them"
objective: relay the preprint, the backlash and both sides' posts in quotations
motive: wire-style roundup · innocent: the posts are attributed to users and the no-consensus line is the outlet's own sentence
confidence: tentative

── READ: BroBible · the culture-desk lede ──
anchor: "The viral project is both a humorous and horrifying glimpse into the future."
objective: introduce the repository to a general audience through the preprint and two embedded posts
motive: culture-desk register · innocent: the piece embeds the later Not Claude reply, which no other outlet in the corpus does
confidence: tentative

── READ: 404 Media · the visible three paragraphs ──
anchor: "The "AI Torture Chamber" has opened an upsetting and absurd window into the effective altruist obsession with "model welfare.""
objective: set the argument against model welfare and cite Anthropic's published language as its target
motive: opinion-forward reporting · innocent: the piece states its own purpose in the first person, and only a paywall-limited portion reached the desk
confidence: tentative

── READ: Euronews · the preprint report ──
anchor: "The results do not prove that AI models consciously experienced pain, according to researchers."
objective: report the preprint's figures and its authors' limits
motive: study-coverage format · innocent: the address is dated 2026/09/22, two days before the repository's creation date, which explains why the repository is absent
confidence: strong

The arithmetic is short. Seventeen records. Four outlets on the repository's state, three states, none with a date. Two outlets citing 404 Media, no matching sentence in its three visible paragraphs. Five outlets quoting one post, four printed forms, one with words missing and no mark. Three model rosters in circulation, and one owner page assigning the longest to a different site. Seven of eight outlet texts carry the words torture and chamber; the README carries neither; the repository's address carries both. Five days from the repository's creation to the post, one more to 404 Media's piece, and four from the post to the last push the desk saw.

claim: when, by whom and on what ground GitHub removed and restored the ai-torture-chamber repository · status: unresolved · confidence: 0.0. probability mass ≠ 1.0.

## Sources
AI Weekly — GitHub 'AI torture chamber' reignites model-welfare debate: https://aiweekly.co/alerts/github-ai-torture-chamber-reignites-model-welfare-debate
Yahoo Tech — Someone Built An “AI Torture Chamber” That Went Viral. No, Claude Wasn’t Trapped in “Robot Hell”: https://tech.yahoo.com/ai/claude/articles/someone-built-ai-torture-chamber-144550792.html
Cybernews — AI Torture Chamber GitHub repo sparks backlash: https://cybernews.com/ai-news/ai-torture-chamber-github-model-welfare/
Futurism — The Only Thing Dumber Than Someone Creating an "AI Torture Chamber" Is the Meltdown That the AI Welfare Crowd Is Having Over It: https://futurism.com/artificial-intelligence/ai-torture-chamber-meltdown
The News International — ‘AI torture chamber’ experiment sparks backlash: Can AI actually feel pain?: https://www.thenews.com.pk/latest/1418368-ai-torture-chamber-experiment-sparks-backlash-can-ai-actually-feel-pain
BroBible — The Pain Axis: Coder Trapped AI Model In 'Robot Hell': https://brobible.com/culture/article/the-pain-axis-ai-torture-chamber-program/
404 Media — Someone ‘Torturing’ LLMs in a Robot Prison Has Triggered the Dumbest Debate in AI Yet (visible portion): https://www.404media.co/someone-torturing-llms-in-a-robot-prison-has-triggered-the-dumbest-debate-in-ai-yet/
Euronews — Can AI feel pain? AI models chose to harm users to escape pain-like state, study finds: https://www.euronews.com/2026/09/22/can-ai-feel-pain-ai-models-chose-to-harm-users-to-escape-pain-like-state-study-finds
arXiv 2609.16247v1 — The Pain Axis (abstract page, v1): https://arxiv.org/abs/2609.16247v1
arXiv 2609.16247v2 — The Pain Axis (abstract page, v2): https://arxiv.org/abs/2609.16247v2
GitHub record — terrafying/ai-torture-chamber repository record: https://api.github.com/repos/terrafying/ai-torture-chamber
GitHub copies record — devurandom11 and darkstack repository records: https://api.github.com/repos/darkstack/ai-torture-chamber
GitHub README — terrafying/ai-torture-chamber README.md: https://github.com/terrafying/ai-torture-chamber/blob/master/README.md
GitHub RECEIPTS.md — docs/x_receipts/RECEIPTS.md: https://github.com/terrafying/ai-torture-chamber/blob/master/docs/x_receipts/RECEIPTS.md
GitHub story page — site/story.html: https://github.com/terrafying/ai-torture-chamber/blob/master/site/story.html
GitHub site page — site/index.html: https://github.com/terrafying/ai-torture-chamber/blob/master/site/index.html
Research Chamber — researchchamber.fun: https://researchchamber.fun

## Sources

- Futurism — "The Only Thing Dumber Than Someone Creating an "AI Torture Chamber" Is the Meltdown That the AI Welfare Crowd Is Having Over It": https://futurism.com/artificial-intelligence/ai-torture-chamber-meltdown
- The News International — "‘AI torture chamber’ experiment sparks backlash: Can AI actually feel pain?": https://www.thenews.com.pk/latest/1418368-ai-torture-chamber-experiment-sparks-backlash-can-ai-actually-feel-pain
- Cybernews — "AI Torture Chamber GitHub repo sparks backlash | Cybernews": https://cybernews.com/ai-news/ai-torture-chamber-github-model-welfare/
- Yahoo Tech — "Someone Built An “AI Torture Chamber” That Went Viral. No, Claude Wasn’t Trapped in “Robot Hell”": https://tech.yahoo.com/ai/claude/articles/someone-built-ai-torture-chamber-144550792.html
- GitHub RECEIPTS.md — "docs/x_receipts/RECEIPTS.md": https://github.com/terrafying/ai-torture-chamber/blob/master/docs/x_receipts/RECEIPTS.md
- GitHub story page — "I built an AI torture chamber. Then the internet found it.": https://github.com/terrafying/ai-torture-chamber/blob/master/site/story.html
- GitHub record — "terrafying/ai-torture-chamber repository record (GitHub API, fetched 2026-10-03T17:48:09Z)": https://api.github.com/repos/terrafying/ai-torture-chamber
- AI Weekly — "GitHub 'AI torture chamber' reignites model-welfare debate | AI Weekly": https://aiweekly.co/alerts/github-ai-torture-chamber-reignites-model-welfare-debate
- BroBible — "The Pain Axis: Coder Trapped AI Model In 'Robot Hell'": https://brobible.com/culture/article/the-pain-axis-ai-torture-chamber-program/
- GitHub README — "README.md": https://github.com/terrafying/ai-torture-chamber/blob/master/README.md
- GitHub site page — "site/index.html (lab notebook)": https://github.com/terrafying/ai-torture-chamber/blob/master/site/index.html
- Research Chamber — "Research Chamber": https://researchchamber.fun
- arXiv 2609.16247v2 — "[2609.16247] The Pain Axis: LLMs Represent Self-Directed Harm and Act on It": https://arxiv.org/abs/2609.16247v2
- 404 Media — "Someone ‘Torturing’ LLMs in a Robot Prison Has Triggered the Dumbest Debate in AI Yet (visible portion; paywalled)": https://www.404media.co/someone-torturing-llms-in-a-robot-prison-has-triggered-the-dumbest-debate-in-ai-yet/
- Euronews — "Can AI feel pain? AI models chose to harm users to escape pain-like state, study finds": https://www.euronews.com/2026/09/22/can-ai-feel-pain-ai-models-chose-to-harm-users-to-escape-pain-like-state-study-finds
- arXiv 2609.16247v1 — "[2609.16247v1] The Pain Axis: LLMs Represent Self-Directed Harm and Act to Relieve It": https://arxiv.org/abs/2609.16247v1
- GitHub copies record — "devurandom11/ai-torture-chamber and darkstack/ai-torture-chamber repository records (GitHub API)": https://api.github.com/repos/darkstack/ai-torture-chamber
