If you share a decision with an AI, can you actually tell where your responsibility ends and the system's begins?

This explores whether people working with AI can tell where their responsibility ends and the system's begins. The corpus suggests that mostly they can't. What users sense is shaped by how much agency they feel the AI has, not by who actually holds formal authority. Leonardi's agency-loops framework shows that when people credit an AI with independent judgment, their sense of control and responsibility shrinks, even though on paper they are still in charge Does attributing agency to AI actually reduce human sense of control?. So the accountability people feel can drift away from the accountability they actually hold. The drift isn't permanent, though. It moves in cycles of delegating, attributing, and taking control back.

There's also a party that users can't see at all. Users tend to picture a two-way split between themselves and the AI. In practice, developers sit above both through system prompts. These are confidential instructions that users can't read or change, and small edits to them can suddenly change how the model behaves Who actually controls what AI systems say?. A related distinction matters here. Some human-like qualities are built in by designers (anthropomimesis). Others are projected onto the AI by users (anthropomorphism). Responsibility for each lands on a different party, but from the user's side the two look the same Who bears responsibility when AI seems human-like?.

The accountability split does become visible at certain moments. In one study, people delegating to an AI agent didn't pull back trust on high-stakes tasks that could be undone. They pulled back sharply on tasks that were irreversible and visible to other people, like sending an email, even when the output was fine What makes people distrust AI agents they delegate to?. Users seem to notice accountability when they picture their own name on a consequence they can't take back. Feedback works in a similar way. People's early bias against disclosed AI partners only adjusts once they see repeated outcomes. A disclosure label alone changes nothing Does revealing AI identity help or hurt user trust?.

Designers are starting to treat this as a user-facing problem. One proposal is the "action receipt": a plain-language, bank-notification-style record of what an agent did, written for the person who delegated the task rather than for engineers Should agent accountability be designed for end users instead of engineers?. Another is MemoAnalyzer, which splits memory-privacy work so that users state their preferences and the AI carries out the edits. The aim is to make each role clearer and lighter, though the evidence is still thin Does splitting privacy work between users and AI reduce cognitive burden?. The gap is widest where it matters most. Companion and high-stakes advisory agents score worst on principles like making the agent's intentions visible Do productivity agents actually protect users better than emotional ones?.

A useful lesson comes from multi-agent research. Labeling one AI agent as "accountable" only changed outcomes when the workflow actually passed the work through that agent Does accountability placement actually shape multi-agent outcomes?. The same seems to apply to people. Telling users they're responsible doesn't make them feel responsible. The split becomes real to them only when the interaction routes decisions, receipts, and consequences back through their hands. The corpus has no direct study measuring whether users accurately perceive this split, which is itself a notable gap.