𝐀𝐧 𝐮𝐧𝐛𝐢𝐚𝐬𝐞𝐝 𝐀𝐈 𝐰𝐨𝐮𝐥𝐝 𝐛𝐞 𝐰𝐨𝐫𝐭𝐡𝐥𝐞𝐬𝐬.

𝐀𝐧 𝐮𝐧𝐛𝐢𝐚𝐬𝐞𝐝 𝐀𝐈 𝐰𝐨𝐮𝐥𝐝 𝐛𝐞 𝐰𝐨𝐫𝐭𝐡𝐥𝐞𝐬𝐬.
Not dangerous. Not naive. Worthless.

One distinction first, because it matters. Bias in the legal sense, a system that quietly treats women or minorities worse, is a defect and it has to be corrected. I am describing something different: a stable disposition to prefer certain answers, certain framings, certain ways of saying a thing.

Strip that disposition out and the model has nothing left to rank one draft above another with. It cannot tell you which option is better, because better is a preference.

So "removing bias from AI" is not an engineering objective that additional funding will eventually deliver, it is a category error.

What the labs actually do is choose a taste and encode it, which is editorial policy applied at planetary scale. I find that description considerably more accurate than "bias mitigation," and considerably more alarming.

Two things are genuinely achievable here, and neither of them is neutrality.

Disclosed bias. Publish what the system prefers, and name the person accountable for deciding it.

Plural bias. Never run a consequential decision through a single model. Use two from different lineages, then treat their disagreement as your signal rather than your noise.

The second one is the good news, and it is nearly free. Paste the same brief into a competing model and read only the passages where the two answers genuinely diverge. Those divergence points are precisely where your own judgment is worth spending, and until recently there was no inexpensive method of locating them.

Almost nobody does this.

Anyone selling you unbiased AI is selling one of two things: ignorance, or undisclosed editorial policy. Neither is what you are paying for.

So ask your next vendor what their model prefers, and who disagrees with it.

Then watch what happens to the room.

#AIGovernance #Leadership #AIStrategy #DecisionMaking #AIBias

Contact

bruno.gentil@sherpaconsultingasia.com

© 2026. All rights reserved.