William Gibson is reported to have said that “the future is already here – it’s just not very evenly distributed.” With survey results showing widespread use of LLMs in both scholarly writings and peer review, the point appears apt to academic contexts.
Just two years ago, a talk on the ethics of LLM use in academia would have touched upon important yet well-known issues: biases, hallucinations, the impact on publish-or-perish dynamics, and similar concerns. We might have even debated whether these models should be used at all. Today, however, a different set of questions appears more relevant. With sakana.ai claiming to have published the first fully AI-generated paper that passed peer review, and evidence indicating that LLM performance on complex cognitive tasks doubles approximately every seven months, we must now seriously consider what happens as LLMs begin fully automating first specific parts of academic processes, and eventually academia as a whole. Traditional notions of merit, evaluation, production, peer review, and grant applications may soon lose their meaning or ecological validity. Those who master these changes stand to benefit immensely, while those who do not risk being left behind. We face the prospect of a digital divide so significant that it evokes Nick Agar’s concerns regarding enhancement potentially leading to two separate human species. How should we navigate this transition?
Drawing on our work with AUTOGEN, ensemble workflow methods, proposed guidelines for the ethical use of LLMs, the complexities of assigning credit and blame in AI-assisted work, and related efforts, this talk outlines the current technological capacities and the state of the ongoing ethical debate. However, its primary focus is on exploring deeper questions about the future of scientific progress and academic labor.