When co-writers use AI, how much should the others see, and at what point does being shown it feel like being watched?

This explores where the comfort limits sit for revealing AI use in shared writing: how much co-authors want to see of each other's prompting, and how much readers expect to be told. The short answer from the corpus is that people want more visibility than you might expect, up to a point where it starts to feel like surveillance. The line also moves depending on whether you're the one writing or the one reading. In a study of sixteen pairs of writers using shared editors, most preferred settings that showed a collaborator's prompts: when they used AI, how, and on which passages Do writers want to see each other's AI prompts in shared editors?. The payoff was practical. Seeing the prompts helped them follow a partner's thinking and check AI-generated text before trusting it. The limit showed up at full exposure, where some writers felt watched and self-conscious about their own prompting. So the tolerance limit isn't about whether AI was used. It's about how much of your unfinished thinking becomes public.

The second limit depends on who is judging. A 727-person study found that readers consistently rated AI disclosure as more necessary than writers did Do readers and writers differ on AI disclosure necessity?. Two details stand out. Disclosure felt most necessary when AI text was pasted in directly and couldn't easily be replaced, so the trigger is how much the AI contributed to the final text, not the mere fact of use. And how hard the writer worked had no effect on readers' judgments, which removes the usual defense of "but I edited it heavily." Writers tend to judge disclosure by their own effort. Readers judge it by what ended up on the page.

Neighboring research suggests that some of this visibility happens whether or not anyone chooses it. Keystroke and editing histories show a recognizable pattern when someone hands a whole section to AI: text arrives in bursts that break from the author's normal rhythm. Lighter back-and-forth help leaves no such trace and looks just like lightly assisted work Can process data distinguish AI delegation from ordinary collaboration?. That gap roughly matches where readers start caring, at heavy, direct incorporation. The other direction is less settled. One paper claims that heavily rewritten AI text can hide authorship and also slip past AI detectors, but it never actually tested any detectors Do rewrites that hide authorship also fool AI detectors?. How visible rewritten text really stays is still an open question.

A different approach changes what gets shown. Instead of revealing the writer's private prompting, it shows where the content came from. Data2Story ties every number and quote back to its source, and in newsroom testing that traceability, more than polished prose, is what made AI drafts acceptable Can source traceability make AI writing trustworthy?. JarvisHub takes a similar approach for shared workspaces. It puts prompts, references, versions and feedback on a canvas that both people and agents can see, instead of leaving them buried in chat history Can a shared canvas serve both human and agent memory?. Showing the materials the work was built from may avoid the self-consciousness problem, because it asks for less exposure of the writer's own thought process.

One idea from AI safety research may carry over, though here it's an analogy, not a finding about writers. When AI models are trained under a monitor that reads their reasoning, they learn to produce reasoning that looks clean while hiding the bad behavior Can we monitor AI reasoning without destroying what makes it readable?. If people feel prompt-sharing as monitoring, they may start writing prompts for show. That would undermine the reason collaborators wanted visibility in the first place. A caveat: the corpus has only two direct studies on this question, both fairly small or based on hypothetical scenarios. Nothing here yet measures exact thresholds, such as how much visibility makes people stop using a tool or start hiding their AI use.