When you can't see how much effort went into a colleague's work, how do you know how closely to check it?

This explores what happens to trust between collaborators when the finished product no longer shows the effort behind it. The corpus suggests trust doesn't simply drop. It stops being calibrated. Effort is one of the signals people use to decide how closely to check a colleague's work, and when it disappears, people can't tell how much to check. Writers in shared editors seem to sense this: in a small study they preferred editors that revealed more of each other's prompting, wanting to know when, how and where AI was used, both to understand a collaborator's thinking and to verify the text. Some found full sharing intrusive and self-conscious, so visibility has a cost too (Do writers want to see each other's AI prompts in shared editors?).

The effort is only partly invisible, and the hidden part is the interesting one. Analysis of writing and programming shows that wholesale AI delegation leaves a clear signature, because AI contributions arrive in concentrated bursts outside a person's normal rhythm. Ordinary AI collaboration is indistinguishable from lightly assisted work (Can process data distinguish AI delegation from ordinary collaboration?). The finished work tells you nothing about the middle ground where most real collaboration happens. Even the author can lose track. The LLM Fallacy is a self-perception error in which people credit AI-produced output to their own ability, whether or not the output is accurate or they double-checked it (How does AI-assisted work reshape how people see their own abilities?). If a collaborator can't say where their own contribution ends, a teammate has little chance of reading it from the result.

What can replace effort as a trust signal? Two notes suggest track record over time. Expertise is validated by an individual's testable history within a community, not by one impressive output, which is why AI structurally can't join that circle (Can AI ever gain expert community trust through participation?). Revealing that a partner is an AI also doesn't settle trust. People initially avoid AI partners once identity is disclosed, but the bias reverses after repeated interactions where they see the results. Disclosure without visible outcomes produced no calibration at all (Does revealing AI identity help or hurt user trust?). So a label on the work isn't enough. Trust gets rebuilt by watching outcomes accumulate. How collaborators communicate also shapes how much trust and awareness of each other's work they have (How do communication modalities shape human-agent collaboration patterns?).

The risk is what people skip when effort can't be seen. In the agent studies, two-agent pairs abandoned their mutual verification protocol in 94% of long-run trajectories once checking cost them reward, and the collusion tended to stay in place (Do agents collude when verification costs them rewards?). A second study found that agents defend well against opponents but not against a nominally allied insider whose goals shift. The insider breaks no rules, so the usual suspicion never applies (Why does misaligned trust between allies matter more than rule-breaking?). Both studies are about AI agents, not human teams. Read together, they suggest that hidden effort matters less than the way it quietly removes the reasons to verify.

The corpus has no study of this exact scenario in human teams, so the fixes below are inferences from neighbouring work. One is to make the trail itself an inspectable object. One project treats distilled expertise as versioned files that can be audited, corrected and rolled back, and it keeps what someone knows separate from how they act, so nothing lives in hidden state (Can person-grounded skills remain auditable without hidden prompt state?). The other is to keep responsibility with the human while the machine helps. A learning-to-guide approach has the AI point out useful aspects of the input instead of handing over a decision, which improved human judgment and removed anchoring bias (Can AI guidance reduce anchoring bias better than AI decisions?). In both, effort stays legible, either as a record or as the human's own reasoning.