𝐘𝐨𝐮𝐫 𝐀𝐈 𝐠𝐨𝐭 𝐛𝐞𝐭𝐭𝐞𝐫 𝐨𝐧𝐞 𝐚𝐧𝐬𝐰𝐞𝐫 𝐚𝐭 𝐚 𝐭𝐢𝐦𝐞. 𝐘𝐨𝐮𝐫 𝐜𝐨𝐦𝐩𝐚𝐧𝐲 𝐠𝐨𝐭 𝐦𝐨𝐫𝐞 𝐰𝐫𝐨𝐧𝐠.

Both can be measured. Both are happening. And the reason is arithmetic, not ideology.

Scott Page, at the University of Michigan, popularized a result called the diversity prediction theorem.

In prediction problems:

Collective error = average individual error − prediction diversity

Read that twice.

A group’s error is not just the average error of its members. You also subtract the degree to which their predictions differ.

That means groups get smarter in two ways:

1. Better individual judgment.
2. Different individual errors.

AI is improving the first and erasing the second.

Set diversity to zero. The equation collapses. The group is exactly as wrong as its average member, permanently, however many members you add.

We test AI one answer at a time.

Is this output accurate?
Is this output fair?
Is this output better than last month’s model?

Measured that way, models really are improving.

But companies do not deploy AI one answer at a time.

They deploy it across thousands or millions of decisions that used to come from people, teams, vendors, analysts, reviewers, managers, and customers who disagreed with each other.

And a million answers from one default model are not a million opinions.
They are one opinion, repeated.

The evidence is already showing up.

In a controlled experiment, some writers were given story ideas from an LLM. Their stories were judged more creative, better written, and more enjoyable. The weakest writers benefited most.

But the AI-assisted stories were also more similar to each other.

The authors called it a social dilemma:

Each writer got better.
The collection got narrower.

That is the AI risk we may be missing: a model can be less biased per answer and still make the organization more fragile in aggregate.

Because the danger is not only bad outputs.

It is correlated outputs.

Disagreement used to be expensive.

You got it by hiring people with different training, incentives, experiences, and instincts, then forcing them to argue.

AI makes disagreement cheaper, but not automatic.

Do not ask once.

Ask three different ways.

Force one answer to argue against another.

Compare the spread before you choose.

For important decisions, add more independence:

a second model,
a dissenting human reviewer,
a different data slice,
or a red-team pass.

The goal is not more process.

The goal is to stop mistaking one polished answer for a portfolio of judgment.

Because a team, human or digital, whose answers are converging is not necessarily getting sharper.

It may just be getting quieter.

#CognitiveDiversity #DecisionMaking #AIStrategy #Innovation #Leadership

Contact

bruno.gentil@sherpaconsultingasia.com

© 2026. All rights reserved.