The research economy has a quiet side hustle, and it's paying $300 a paper to document how you think. A graduate student recently shared that companies are offering cash for detailed walkthroughs of the research process behind first-author papers accepted at ICLR, NeurIPS, and ICML for 2024 through 2026. The stated goal is to build a high-quality dataset. The unstated one, as the submitter notes with weary honesty, is to train the AI that could replace them as a researcher. This is not a conspiracy theory. It is a transaction happening right now, and it forces a question our community needs to answer: Are we selling the blueprint for our own obsolescence at a fair rate?
We have already seen the warning signs. As How AI Is Reshaping the Peer Review Process for Better or Worse documented, AI-written reviews and AI-assisted paper generation are straining the very gatekeepers of academic quality. Meanwhile, ICLR 2027 continues desk reject policy to filter low-quality AI submissions, a direct acknowledgment that the flood of machine-generated content has become unmanageable. These are not separate problems. They are the same loop: AI produces slop, humans reject it, then those same humans are paid to teach the next version of the AI how to pass. The survey described here is not an anomaly. It is the natural next step, a targeted extraction of the cognitive process that separates a strong paper from a desk reject.
Our take is plain: this is a shortsighted bargain, and the math does not favor the researcher. At $300 per paper for an hour of detailed work, the compensation is decent for a grad student on a stipend. But the long-term cost is structural. By training an AI on the specific reasoning, citation strategies, and experimental justifications that earned acceptance at top venues, you are effectively compressing years of learned expertise into a prompt. The dataset being built here is high-fidelity, and high-fidelity data is what makes AI genuinely useful rather than merely novel. The researchers who supply it are not just annotators. They are unwitting curriculum designers for a system that will undercut the value of their own training.
The referral offer makes this even more troubling. Splitting a $200 referral fee to net $400 per submission for the referred researcher and $100 for the original poster is a smart hustle. It is also a recruitment pipeline for the very automation that shrinks the academic job market. Every new participant adds another layer of validated research logic to the dataset. This is not a side effect. It is the business model. The companies behind these surveys do not need to replace every researcher tomorrow. They only need enough examples of how a good researcher thinks to approximate the process convincingly. That threshold gets closer with every submission.
The specific consequence to watch is the erosion of the tacit knowledge that conferences like ICLR and NeurIPS claim to evaluate. If the decision-making process behind accepted papers becomes a commodity dataset, what distinguishes a human submission from a fine-tuned model's output? The desk reject policies at ICLR are a stopgap. They filter obvious slop. They cannot filter a model trained on the actual workflows of accepted authors. The open question is whether the community will treat these surveys as a threat worth regulating, or as just another way to make rent. Either way, the transaction is happening. The only choice left is whether to name it for what it is: a quiet transfer of expertise from the people who earned it to the systems that will learn to do without them.