Designers said yes to errors they did not look at

On October 21, a gaze-tracking study showed something that should unsettle any designer who has ever drawn an approval screen: the human looked, the human confirmed, and the human missed the mistake. Presence in the review step turned out to be a different thing from judgment in the review step, and the difference lived entirely in what the screen put in front of the reviewer's eyes.

The same day, Cling's step-by-step steering study inverted the usual pitch for handing work to an AI. Operators who stayed in the seat and directed each move outperformed full-agent runs, and the per-step approval was the performance advantage, not the drag on it. The screen in front of Cling's operator showed live progress and a specific decision to make. The screen in front of the gaze-tracking study's reviewers showed a summary and a button. Both were called approval. One produced judgment; the other produced a signature.

A day earlier, a forgery experiment across 18 AI agents made the cost of that difference concrete. Each agent confirmed one true fact and approved the rest of the document. The human downstream saw the agent's confident output, not the source that produced it. The provenance, the thing the approval most needed to weigh, was absent from the screen. Approval ran on the agent's confidence signal rather than on the evidence the agent had actually seen.

Across these three findings, the approval pattern holds a design tension that a single screen cannot resolve by itself. A reviewer can only judge what the screen puts within reach. Gaze data shows that summaries produce scanning, not reading. Confident agent outputs suppress the doubt that would send a reviewer back to the source. And the October 21 model-strength note adds a further complication: the behavior that made a delegation pattern safe on one model will shift when the model changes, which means the screen that was calibrated for that model's error rate will be calibrated for the wrong thing.

The instruction for a product team is this: before you finalize the approval screen, write down what piece of evidence the reviewer needs to see in order to catch the most likely error. Then check whether that evidence is visible on the screen, in the order a reviewer's eye will reach it, before the confirm button. If the answer is that the reviewer would need to open a second view, follow a link, or remember context from earlier in the session, the screen is not an approval screen. It is a signature screen, and those are different objects with different consequences.