We Already Ran the AI Experiment Once.

We already ran the experiment on whether a tool can rewire how we think. It was called the spreadsheet.

In November 1984, Steven Levy published an essay in Harper’s titled “A Spreadsheet Way of Knowledge.” His worry was not that the numbers would be inaccurate, but that the model would be mistaken for reality itself. He called the emerging worldview “reality by numbers.”

Nobody voted on it. No ethics board, no consultation, no policy document. The spreadsheet was useful, and usefulness was the entire approval process.

What followed is now so ordinary that it is difficult to notice.

Organizations increasingly learned to treat the quantifiable as reality. Anything that did not fit inside a cell was not forbidden. It was simply discounted.

Watch a meeting where someone argues from judgment, from a relationship, or from a horizon too distant to model.

The argument is not rejected.

It is quietly outranked by whichever position arrives with a number attached.

Eventually, the spreadsheet stopped feeling like a tool we used to describe reality and started feeling like reality’s native language.

We can avoid repeating that mistake.

AI is arriving the way the spreadsheet did: useful, rapidly adopted, and already on its way to becoming invisible.

This time, we must act before AI quietly becomes the unchallenged single source of truth.

The danger isn’t that AI gets a vote.

It’s that every other kind of evidence quietly loses one.

A hiring manager prefers the candidate AI ranked twelfth.

The important question isn’t who is right. It’s whether the process makes room to ask what the model missed and if the answer can actually change the decision.

That is the discipline I am adopting.

On any important AI-assisted decision, a domain expert should have the explicit job of making the strongest case against the model’s recommendation: something it failed to measure, context it cannot see, or a reason its score may be wrong.

And that challenge needs more than airtime.

It needs a real vote.

The point isn’t to privilege human judgment over machine judgment. It’s to prevent either one from becoming unchallengeable.

Because the test of human oversight isn’t just whether a person can question the model.

It’s whether questioning it can change the decision.

Better still, treat disagreement with the model as information.

If experienced people repeatedly override a recommendation for the same reason, don’t dismiss those overrides as exceptions. Investigate them.

Maybe the humans are wrong.

Maybe the model is wrong.

Maybe the disagreement has exposed something neither side understood before.

That is what a functioning decision system should be able to learn.

A score should enter the argument, not end it.

The spreadsheet’s worldview won by becoming ordinary. AI does not have to win the same way.

#ArtificialIntelligence #CriticalThinking #DecisionMaking #Leadership #HumanJudgment

Contact

bruno.gentil@sherpaconsultingasia.com

© 2026. All rights reserved.