Clicking approve should not teach your AI anything

An approval click tells you a person allowed one action.
It does not tell you they checked the reasons behind it.
An AI system that treats the click as agreement ends up learning from a lesson nobody meant to give.

What does an approval click confirm?

An AI agent that can act for you usually asks first.
The card says “send this email” or “update this record”, and you press approve or decline.
Behind that card sit things the agent assumed: which client this is, what you charge them, which version of the document is current.

The card shows the action.
At best, the person checks the action.
Nobody reading “send this email” at four in the afternoon also checks every fact the agent used to write it.

Most approvals get waved through.
Anthropic found that people using its coding tool accept 93% of the permission prompts it shows them, and named the problem approval fatigue.
Researchers call the wider habit automation bias: people trust an automated suggestion more than they should, and more so when busy.

How does a click turn into a belief?

Every AI supplier wants feedback, and approvals are the cheapest feedback going.
Each one looks like a person saying “yes, that was right”.
A system can quietly use them to decide what it trusts next time.

Ours did.
When someone approved an action, our agent raised the standing of the notes it had in front of it at the time.
That included notes the agent had written itself.
Those notes then ranked up to 40% higher in later searches.

Nobody designed that as learning.
It came from a rule that sounded sensible: if a person approved the work, what the agent used must have been good.
But the person approved an email, not the agent’s guess about the client sitting beside it.

We closed it, then found a second route.
A nightly job promoted notes the agent looked up often, and the agent’s own notes counted as used every time it retrieved them.
We closed that one on 5 October.

The approval button that did not exist

The approval step can fail in a stranger way too.
Our agent once told a user to “confirm or cancel using the buttons above”.
It had produced no buttons.
Every fact in the reply was true except the interface, and the user waited for a card that never came.

We now check every reply that mentions an approval control against whether the agent produced one.
The lesson for a buyer: an approval step only protects you if it exists, so look at the record of who approved what, not the chat.

What the law says about human review

The UK’s Data (Use and Access) Act 2025 rewrote the rules on decisions made by automation, from 5 February 2026.
Where a significant decision about a person rests solely on automation, they can ask for a human to step in.
The ICO has long said a token review does not count.

The EU AI Act’s Article 14 goes further for high-risk systems.
It names automation bias, and wants the people overseeing a system to stay aware of it.
Under a delay agreed this year, those rules are due to start in December 2027.

Ben Green, who studies public-sector algorithms, found that rules requiring a human in the loop often give false comfort.
The person cannot fix a flawed system from one button, and takes the blame when it fails.

The rule we use

An approval confirms one action and teaches the system nothing.
When we want the agent to learn, we ask a separate question a person answers on purpose.

• Instructions and corrections teach. Something a person wrote down can become a rule.
  The agent’s own conclusions never can.
• Outcomes come from people. After a piece of work runs, we ask “Did this do what you wanted?”.
  The agent cannot answer that about its own work.
• A click stays a click. Approving an action moves that action forward, and changes nothing about what the agent believes.

Our AI agent keeps approval and learning apart, and our trust page sets out its other checks.
For how an AI goes wrong when it learns from its own notes, read when your AI assistant learns from itself.
This post is not legal advice.