Rubber-stamp oversight

Approval is requested after the decision is effectively made, or presented as the action rather than its consequence. The person sees ‘Approve’ beside a summary written by the system that wants the approval. Refusing costs time and effort; accepting costs nothing visible. Over a few dozen prompts, approval becomes reflex.

The reflex is well documented, and older than agents. The human-factors literature calls it automation bias and complacency, and its central finding is uncomfortable: complacency ‘is found in both naive and expert participants and cannot be overcome with simple practice’. In the cockpit studies that review draws on, people given an automated aid followed its recommendation even when it contradicted their training. Surveying 41 policies that require a person to oversee a government algorithm, Ben Green found that ‘people are unable to perform the desired oversight functions’ and that the policies ‘provide a false sense of security’. Legal scholars call the move the MABA-MABA trap: answering a worry about the machine by inserting a person, without designing what the person can see or do.

A gate designed to be passed is worse than no gate, because it transfers responsibility to someone who had no means to exercise it.

Coding agents are the clearest case, because the approval rates are published. Claude Code asks before running a command or editing a file, and Anthropic reports that ‘users accept 93% of them anyway’, describing ‘approval fatigue, where people stop paying close attention to what they’re approving’; its answer was to replace most prompts with an automated check. In a preregistered study of 113 people without software backgrounds supervising a simulated agent, those asked to approve each action approved 40% of the actions that went beyond what they had asked for. Singapore’s 2026 framework for agentic AI tells deployers to watch for exactly this: ‘A low rate may signal rubber-stamping behaviours.’

Browser agents ask for confirmation before a purchase or an email, and the dark-pattern research shows what that is worth: supervising participants avoided more deceptive interfaces than the agent alone, at the cost of ‘attentional tunneling and cognitive load’, and reported noticing fewer dark patterns than when they worked alone (see Agent-directed dark pattern).

Law has been here before, for automated decisions about people. Under the GDPR a human review that is ‘just a token gesture’ does not count; the Article 29 Working Party said in 2018 that a controller ‘cannot avoid the Article 22 provisions by fabricating human involvement’, and the Court of Justice’s SCHUFA ruling in 2023 treated a credit score as an automated decision wherever the bank gave it ‘a determining role’. The ICO’s 2026 draft guidance asks for involvement that is ‘active and not just a token gesture’, by someone with ‘discretion and authority to alter the decision’, and says ‘ad hoc spot-checking isn’t sufficient’. The EU AI Act goes a step further for high-risk systems: the person overseeing one must be able to stay ‘aware of the possible tendency of automatically relying or over-relying on the output’. That duty now applies from December 2027.