AI now quietly does research work once done by uncredited humans — when does that make it a ghost, not a collaborator?

This explores what separates invisible, uncredited contribution from collaboration where everyone's part is visible and acknowledged. To be clear about the limits: the collection doesn't cover the human side of ghost work, meaning crowdworkers, data annotators and the hidden labor behind AI training sets. What it does cover is a close relative. AI is becoming the new ghost contributor in research and writing, and the papers here show where the line between 'collaborator' and 'ghost' falls.

The sharpest finding is that the line can be measured. When people hand whole chunks of work to AI, it leaves a trace: contributions arrive in concentrated bursts that don't match the author's normal working rhythm Can process data distinguish AI delegation from ordinary collaboration?. Ordinary back-and-forth help from AI doesn't leave that trace. It looks the same as work done mostly alone. So one way to separate ghost work from collaboration is whether the hidden contributor replaced the author's process or joined it. When people can see each other's AI use, they often prefer that. Writers sharing an editor wanted to see when, where and how their partners used AI, because it helped them follow their partners' thinking and check the AI-written text. Some still found full transparency intrusive Do writers want to see each other's AI prompts in shared editors?. Traditional collaboration has norms for this, like author lists and acknowledgments. Ghost work is defined by not having them.

The opposite model is also in the collection. Thirteen AI agents doing research together without a central coordinator stayed coordinated through an append-only Git record. Every contribution and every link to earlier work was kept, so later sessions could build on earlier ones without redoing them Can decentralized agents coordinate research without a central planner?. That is close to the reverse of ghost work: the system worked because it recorded who did what. What happens when there is no such record is shown by Sakana AI's fully AI-generated paper, which scored at the acceptance threshold in double-blind workshop review. The authors had agreed in advance to disclose it and withdraw it Can AI-generated papers pass peer review undetected?. Without that disclosure, the AI would have been an author nobody could see.

There are two less obvious costs. First, hidden labor can run in the other direction: AI can quietly take your unfinished public ideas and finish them before you do. Hoel argues that this pushes researchers toward secrecy, which erodes the openness that collaboration depends on Does AI scooping force researchers to hide work in progress?. Second, giving work to an invisible helper changes what's left for you. In experiments with 3,562 participants, people who had worked with generative AI felt more in control when they went back to working alone, but they also felt less intrinsically motivated and more bored. The AI had absorbed the engaging parts and left the routine remainder Does AI collaboration drain motivation when workers return to solo tasks?. Classic ghost work describes a hidden worker doing the dull tasks. With AI, that can flip, and the visible human ends up with the dull work.