Eric Hartford has been through this before. In October 2025 the creator of the Dolphin family of open-source models woke to find his US$200-a-month ChatGPT Pro account deleted without warning — a decade of chat history gone, appeal rejected by what looked like an automated reply. Twenty-four hours of public outcry later, OpenAI reinstated it and told him the flag was for “unsolicited CBRN-related safety testing.”
This week it happened again, and this time there was no reversal. Hartford says OpenAI deactivated his account on 4 September for “distilling” — using its outputs to train a competing model — which he denies: he hasn’t trained a model since publishing his QuixiMath dataset, which he generated himself. On 8 September NZT he posted the appeal outcome: a short email upholding the deactivation and stating OpenAI “will no longer consider additional requests to appeal this case.” His account — which he says holds his medical information, notes about his heart medications, and years of history — is locked behind a decision no human appears to have explained.
🔍 THE BOTTOM LINE: This is not an AI story with a novel ending. It is the oldest story in the platform economy, told with higher stakes. Social networks have been deleting accounts — photos, posts, memories and all — since they began. The difference in 2026 is that the evicted data is not a photo album; it is working knowledge, medical notes and professional history, and the platform holding it has no human appeals process worth the name.
History repeating, with sharper stakes
The pattern should be familiar to anyone who lived through the social-media era. Platforms have always reserved the right to terminate accounts, and have always exercised it with imperfect accuracy. People lost fifteen years of photos to a misfired content filter. Journalists lost whole archives to mass deplatforming waves. The apology, when one came, never restored the data — the platforms’ own terms made the user responsible for keeping copies.
What is different now is what people actually keep in the system. A 2015 Facebook eviction cost you photos. A 2026 ChatGPT eviction costs you the research you compiled with it, the drafts, the planning, and — as Hartford’s case and others show — things that look uncomfortably like medical records. One user on OpenAI’s own forum described losing access to an account holding his parents’ health records, including his mother’s cancer treatment history. Deactivated accounts, he found, cannot even run OpenAI’s data-export tool: the export requires logging in.
OpenAI has not publicly commented on Hartford’s second ban. The company has defended sudden deactivations before on the logic that warning users would let genuine abusers — scrapers and resellers running reverse proxies — adapt before enforcement lands. That logic may be sound as security policy. It does not change the experience of the false positive, which is indistinguishable from eviction by algorithm.
The trigger: a distillation panic with collateral damage
The stated reason — “distilling” — is the same one other users report receiving around the launch of OpenAI’s Astra model, and it is worth understanding why that word is radioactive in 2026. The US government and the frontier labs have spent the year accusing Chinese labs of distilling American models — DeepSeek and Moonshot AI have both faced the claim — and the White House has pressured labs to harden frontier systems against exactly this. Enforcement around major launches has visibly tightened. Heavy legitimate users — developers who automate heavily, run big workloads, or publish open models — look statistically similar to scrapers, and they are the collateral.
Hartford is a conspicuous piece of that collateral. He built his name on Dolphin, the most-downloaded family of fine-tuned open models, and he is open about using frontier APIs in his workflow. Whether OpenAI’s detectors saw something real or fired on a pattern that looks like distilling from the outside, the outcome is the same: an appeal denied in hours, no specifics offered, account gone. Anthropic has nuked an entire company’s accounts overnight on even less process. This is what enforcement looks like when the customer-support layer is thinner than the detection layer.
The resilience playbook — older than AI, more urgent than ever
None of the lessons here are new. They are the same ones the social-media era taught, and the same ones off-grid communities have always known. What is new is how many people skipped them because AI subscriptions felt like utilities.
- Export your data on a schedule, not in a crisis. Every major AI platform has a data-export tool. Hartford’s archive is unrecoverable precisely because export requires an active account. If your ChatGPT, Claude or Gemini history holds anything you would miss, export it monthly and store it somewhere you control — a drive in a drawer counts.
- Never use a chat account as a filing cabinet. If it matters — medical records, tax research, a manuscript — the master copy lives on your machine or your own cloud storage. The AI platform is a workspace, not a vault. This is the single rule that would have prevented the worst outcomes in both Hartford cases.
- Diversify providers. Two subscriptions beat one, not for features but for survival: when one account vanishes mid-project, the other keeps you working. The same logic that says don’t bank everything at one institution.
- Keep a local fallback for critical workflows. A laptop running a 27B-class open model now covers drafting, summarising and coding on Apple-silicon hardware many readers already own. Local models are the distributed-generation answer to a centralised grid — the solar-panel option in a world of one power company. The right to run your own AI is becoming a practical question, not an ideological one.
- Assume terms can change under you. Providers can restrict use-cases, raise prices, or cut a country off with a week’s notice — Anthropic and OpenAI have both done it. Resilience means no single provider’s decision can stop your work.
Hartford’s own sign-off after the first ban remains the most compressed version of the argument: “not your weights, not your AI. Go local.” The second ban suggests he will not be getting the third data point — and that anyone whose life runs through a single AI account is betting they’ll never draw Hartford’s card.
Frequently asked questions
Who is Eric Hartford? An AI developer best known for creating the Dolphin family of open-source fine-tuned models, and chief scientist at Lazarus AI. His OpenAI account was banned in October 2025 and reinstated after public outcry; it was banned again on 4 September 2026 and the appeal was permanently denied on 8 September NZT.
Why did OpenAI ban him? OpenAI’s deactivation cited “distilling” — extracting model outputs to train competing systems. Hartford denies distilling anything. OpenAI has not publicly commented or provided specifics, and its appeal email offers no evidence.
Can you export your data from a deactivated account? No. OpenAI’s export tool requires logging in, which a deactivated account cannot do. Anything not exported before deactivation is effectively gone, which is why scheduled exports matter.
What should an ordinary user do differently? Export your AI chat history monthly to storage you control, keep master copies of anything important outside the platform, run with two providers rather than one, and consider a local open-weights model as a fallback for critical work.
Is this a free-speech issue? Not exactly — private platforms can choose their customers. The issue is data custody: when a platform holds years of personal and professional records but offers no meaningful appeal and no export path after deactivation, users bear all of the platform risk.
🔍 THE BOTTOM LINE
The social-media era already ran this experiment: platforms evict people, evictions are sometimes wrong, and the evicted lose their data unless they kept their own. AI platforms inherited the pattern and raised the stakes, because what we store in them is no longer photos but working knowledge. The fix is not outrage — Hartford tried that, and this time it didn’t work. The fix is the boring one: back up what matters, keep a copy where no company can reach it, and make sure no single account — or single company — sits between you and your own work.
📰 Sources
- Eric Hartford (@QuixiAI) on X — ban announcement, appeal-denial post and follow-up threads, 4–8 September 2026
- sanj.dev — “OpenAI Closed Your Account? Good Luck Getting It Back” (17 October 2025, first-ban account and reinstatement)
- OpenAI Community forum — false-positive deactivation wave threads, including medical-records cases (July–September 2026)
- DW — “US-China AI rivalry heats up as Chinese models gain ground” (23 July 2026, distillation-enforcement context)
