Readers want AI use disclosed more than writers do, and they discount hard work, focusing on AI text that goes straight into the piece.

This explores where the line between human and AI contribution falls when researchers and writers disclose AI use, and who gets to draw it. The corpus has no settled taxonomy, such as 'AI for editing is fine, AI for ideas must be disclosed.' What it does offer is a sharper finding: the line depends on who's looking. In a 727-person vignette study, readers rated disclosure as more necessary than writers did, in every condition. Two things drove those judgments. One was whether AI text went directly into the final piece. The other was whether that contribution was irreplaceable, meaning the human couldn't easily have produced it. How much effort the writer put in made no difference Do readers and writers differ on AI disclosure necessity?. So the intuition that 'I worked hard on this, so it's mine' doesn't persuade readers. What they care about is whether the AI supplied something essential.

That raises a practical problem: disclosure is close to an honor system. A review of 30 studies found that people can't reliably tell AI-generated text, images, or voice from human work. Accuracy hovers around chance and hasn't kept up as AI output has become more realistic Can people reliably spot content made by AI?. When no one can check, suspicion fills the gap. Comments that get accused of being AI-written turn out to lack any features that actually separate AI text from human text. The accusations act more like gatekeeping than detection, and the people they wrong are human writers who aren't believed Do unfounded AI accusations harm human writers instead?. Disclosure norms therefore protect honest authors as well as readers.

The research-publishing cases show how concealment, not AI use itself, becomes the ethical breach. Researchers found 18 arXiv manuscripts with hidden instructions telling AI reviewers to praise the paper. That counts as questionable research practice because it is both concealed and self-serving, whatever the authors say they intended Are hidden AI prompts in preprints a deceptive research practice?. The opposite case is the AI Scientist-v2 experiment. Three fully AI-generated papers went to an ICLR workshop under an agreement made in advance with the organizers. One scored well enough to be accepted and was then withdrawn as planned Can AI systems generate research papers that pass peer review?. The AI's contribution was total, yet the experiment was considered legitimate because the right people knew about it beforehand. These cases sit inside a broader arms race in which AI speeds up paper production, automates review, enables manipulation, and prompts new defenses, each feeding the others Does AI create a coupled arms race in research production and review?.

There is also a reason to want disclosure beyond individual honesty. Researchers who use AI publish about 3 times as many papers and get 4.8 times as many citations. Across science as a whole, though, topic coverage shrinks and collaboration falls, as work concentrates on data-rich problems Does AI help individual scientists while narrowing scientific focus?. Disclosure is one of the few ways to see that drift happening. Another line of work argues that human-AI collaboration beats fully autonomous AI research on both speed and safety, partly because a human stays in a position to check the work Can human-AI research teams improve faster than autonomous AI systems?. On that view, a disclosure statement is a record of who verified what, as much as a confession of AI use.

The surprise in this material is that 'how much did the AI do?' may be the wrong question. Readers seem to ask 'could the human have done this without it?' Institutions seem to ask 'was it hidden?' Neither is a percentage of authorship. The corpus does not answer the more practical question of how journals and conferences should word their disclosure categories, so treat those policy specifics as an open gap.