Researchers curb unpublished uploads to AI models after two discovery disputes

An upload can lead to three different uses, and the vendor policies handle each one separately. Inference is the step where the service processes a prompt to produce an answer ^[8]. The provider may also retain the conversation. Training goes a step further, because the data can shape later versions of the model ^[9]. By itself, an upload is no proof that the material was retained or used for training ^[10].

On business and commercial accounts, the training defaults are set correctly. OpenAI's business policy keeps business and API inputs and outputs out of training by default ^[11]. Retention is handled by separate controls for qualifying organizations ^[12]. Anthropic's commercial policy is the same by default for inputs and outputs. A user can still open them to training by granting permission or by sending feedback that carries the conversation ^[13]. The exception is the thumbs-down feedback button. A user who clicks it and submits feedback may send Anthropic the whole conversation connected to that feedback, not just the single prompt ^[14].

Consumer tiers are looser. ChatGPT users can opt out of training on new conversations, though OpenAI may still use information submitted through feedback ^[15]. Claude consumer conversations can be used for model improvement with permission, and flagged conversations may support safety enforcement or safeguards training ^[16]. These policies describe when data may be used. Whether either company reused unpublished research in these disputes is a question they leave unanswered ^[27].

So far, no independent verification shows misuse by either company ^[18]. Similar findings cannot reveal how a model reached an answer. Proving misuse would take evidence of what the model or the researchers could access while the work was underway ^[19]. OpenAI's reported finding covers Buckmaster's prompts from the previous two months ^[4], for work the pair had pursued for a year ^[1]. Its wider statement is that neither its researchers nor its agents saw the work before publication ^[5]. Anthropic's quoted line, that its model was "not trained on any user transcripts," addresses training alone ^[7]. The Neuron argues that vendor assurances should cover retention and access as explicitly as training ^[26]. The DNA result is also unsettled: researchers still do not know the biological function of the reported pattern ^[20].

Priority is a separate concern from data settings. The Neuron frames the second concern as whether researchers can still publish first when AI systems can reach similar findings much faster ^[21]. The workarounds Nature describes swap one cost for another. Samuel Mehr's lab prohibits uploads of protected information to commercial models and discourages their use during research ^[22]. Generic labels of the kind Laurentino uses cut what leaves the lab, but they can deprive an assistant of the context it needs ^[17][23]. Holding preliminary findings until publication gives colleagues fewer chances to review methods early. Labs with approved private systems can keep using AI during research ^[25][24].

In my context, I would put the control at the account. That means a business or API tier, the organization's retention settings confirmed, and a written lab rule for the feedback button. That keeps real variable names available to the assistant, while publication priority still has to be handled on its own ^[21].