Opinion

Mathematicians Are Hiding Unpublished Work From OpenAI

A Mastodon post asking whether researchers dare feed unpublished proofs to ChatGPT went viral across platforms, exposing a gap in OpenAI's data policies for independent academics who fall between consumer and enterprise tiers.

Font size

It started small.

On August 22, Andreas Thom, a mathematician at TU Dresden, posted a question on Mathstodon: would you dare feed an unpublished mathematical proof to ChatGPT?

The question itself was simple, but it landed on a gray area no AI company has been willing to address head-on.

Within 48 hours, the Mastodon post was reposted to X by Valerio Capraro, picked up on Bluesky by did:plc:ckaz32jwl6t2cno6fmuw2nhn, and eventually reached the front page of Hacker News. Cross-platform amplification at that speed is itself a signal; a post only travels that far when it hits a nerve.

What Researchers Are Actually Afraid Of

The comment threads surfaced three concrete fears.

First, unpublished results becoming training data.

An assistant professor working in number theory wrote: "I let GPT-5 polish a draft of a proof, and the next day I incorporated that idea into a paper and submitted it. If the model on the other side had already learned that trick six months earlier, what happens when reviewers see something resembling it in the model's output the day my paper goes public?"

This is not paranoia. The handling of input data by AI companies has long been opaque: API calls are retained by default for 30 days for abuse monitoring, enterprise users can request Zero Data Retention (ZDR), but no AI vendor has disclosed the precise coverage of ZDR, the granularity of its audit trails, or the forensic path available when something goes wrong.

Second, reasoning paths leaking through reproduction.

More sensitive than the conclusion is the proof strategy itself. One anonymous commenter put it bluntly: "I can rewrite the conclusion, but the construction of that reduction took two years to develop. The moment I feed it to a model, I'm handing the unpublished keys to a system I don't control."

Third, the impossibility of withdrawal.

GDPR Article 17 grants a "right to be forgotten," but only if you can prove the data is held by a particular US-based company. For a 200-line LaTeX draft embedded in an inference request, how do you prove it still sits in some server log—and how do you prove it has been deleted?

When Valerio Capraro reposted to X, he attached a measured remark: "We're letting researchers' careers serve as free annotation labor for AI companies."

What OpenAI's Privacy Terms Actually Say

To check the facts rather than rely on impressions, I read through OpenAI's current Terms of Service and the API Data Usage Policies.

Core facts:

  • By default, API input data is deleted after 30 days and used only for abuse and safety monitoring.
  • Enterprise users can request Zero Data Retention, but this path is not available as a self-serve option for academic researchers—you have to go through enterprise sales.
  • Nothing in the terms promises that "unpublished academic work" will be handled in any isolated category.
  • The training-data opt-out is on by default, but only for ChatGPT on web and mobile; for API calls, the opt-out also requires an enterprise account to take effect by default.

In other words, OpenAI's commitment framework is built for two tiers: consumers and large enterprises. The middle tier—independent researchers, postdocs, small labs—is a vacuum.

A phrase that kept recurring in the discussion: trustworthy and verifiable are different things. AI companies want you to "believe" they won't misuse your data, but they haven't given researchers an executable audit trail.

Practical Workarounds Researchers Are Using

For researchers who don't want to be held hostage by this uncertainty, several workable paths currently exist.

Option 1: Local open-source models with private deployment.

This was the most frequently mentioned approach in the discussion. Open-source families like Qwen3 (Alibaba) and DeepSeek-V3 (DeepSeek) now approach closed-source flagships on mathematical reasoning tasks, and running a 70B-class model locally is the hardware threshold that came up most often in math circles.

A minimal deployment sketch (using vLLM; adjust model and parameters for your hardware):

pip install vllm

python -m vllm.entrypoints.openai.api_server \
  --model Qwen/Qwen3-235B-A22B-Instruct \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.92 \
  --host 127.0.0.1 \
  --port 8000

All proof drafts stay on your own GPU server, with no third-party contact. The trade-off is either buying your own cards or renting a cloud instance with H100s.

Option 2: Call the API, but with opt-out and an explicit agreement.

If local inference isn't enough, at least do these three things:

  1. Apply for OpenAI's Zero Data Retention (through the enterprise application, not the ChatGPT web interface).
  2. Sanitize proof drafts before sending them to the model—replace variable names with non-semantic symbols, replace the statement being proved with placeholders.
  3. Record in the paper's Methods section that "Model XX was used in ZDR mode," so reviewers and the broader community know how your data was handled.
import re

def sanitize_proof(proof: str) -> str:
    # Replace math variables a, b, c, ..., z
    proof = re.sub(r'\b([a-z])\b', lambda m: f'x{ord(m.group(1))-97}', proof)
    # Replace possible conclusion statements
    proof = re.sub(r'\b(qed|theorem|lemma)\b', '[REDACTED]', proof, flags=re.IGNORECASE)
    return proof

raw = r"By lemma 3, a divides b. Therefore b = a * q. QED."
print(sanitize_proof(raw))

Option 3: Skip cloud LLMs entirely and rely on code-assist tools.

Mathematics has a distinctive advantage here: a large share of "AI assistance" in the field can be handled by interactive proof assistants like Lean 4, Coq, and Isabelle. These tools run entirely in your local environment, so there's no data leaving your machine. Lean 4's mathlib already covers a substantial slice of undergraduate-level mathematics, and pairing it with an LLM for tactic suggestions is currently the most defensible combination.

Editorial Take

This story deserves more attention than the standard "AI accelerates research" narrative, because it punctures an illusion that's been carefully maintained: AI companies want researchers to treat these systems as a "free postdoc," but none of them want to take on the responsibilities that come with that role.

My prediction: within the next 12 months, a clear "high-sensitivity researcher exodus" will become visible—primarily toward vendors that have opened dedicated isolation channels for academic use cases, and secondarily toward local open-source deployments. The leading labs in mathematics, cryptography, and medicinal chemistry are the most likely to break ranks first, because once their results leak, the cost of reproduction is far higher than in other disciplines.

This conversation also differs from the copyright disputes of the past year: copyright fights end in rulings and settlements, while the "trust cost" of unpublished academic work is neither quantifiable nor traceable. If AI companies keep the middle tier silent in their ToS, they won't just lose a handful of mathematicians—they'll erode the legitimacy of the "AI for Science" story itself.

Source: https://mathstodon.xyz/@andreasthom/117240535270608201