Sometimes, A.I. agents need approval from a user before executing a potentially dangerous operation. With this method, the chance of receiving approval in time should be greatly increased!

, ,

Background:

AI assistants often have “guardrails” associated with their operation, where an AI must get human confirmation for a potentially unsafe action before executing it.

Examples include:

  • AI agent generates an email (but requires human confirmation to send it).
  • AI agent finds something to purchase only (but requires human approval before checkout).
  • AI agent identifies a target for a killer drone to shoot a missile / poison gas / whatever at, but needs a human to say “ah, those might be terrorists or something” before the operation can proceed.

In all of these cases, the human limits the degree to which the AI can accomplish its goal.

The Issue:

The problem here is in obtaining confirmation: the AI can pop up as many dialog boxes and play as many alert sounds as it wants, but if the human is away from the computer or has turned the sound down, then they might not notice the AI request in time.

This is a problem for time-sensitive requests. For example, perhaps the AI has found a vintage shoe rack on an auction site with only 15 minutes remaining, or maybe a killer drone has located the last remnants of humanity fleeing into a fortified bunker. In both cases, time is of the essence!

Proposal:

The solution to this “ignored notification” problem is simple: when the AI needs user confirmation, it will ring a bell (in the real world: see Fig. 1).

This bell would of course be unaffected by the user’s volume settings on their computer, or whether the user was alt-tabbed out into another application.

If the user approves the AI’s request within a sufficiently narrow time frame, the AI may also optionally dispense a treat (such as a small piece of chocolate) to reward the user for their diligence.

Fig. 1: An AI agent may require user approval to run a “dangerous” command. In order to increase the chance of a user noticing the request for approval, the AI can ring a bell in the real world. Additionally, it may choose to dispense a treat to reward the user for clicking the “approve” button.

Conclusion:

Using this technology, the user and AI can work together more effectively, and the AI might even be able to train the user with candy (through classic Pavlovian conditioning) to approve requests more quickly.

PROS: Helps humans and AI agents collaborate effectively with less “waiting for approval” latency.

CONS: The AI must be careful not to dispense too many sugary treats, as excessive consumption may be bad for the health of the human. Uneaten treats may also attract vermin to the user’s desk.

Originally published 2026-02-16.