Tracey Spicer on AI Bias, Three Years On

26 September 2026

Ten years ago I started asking a question that seemed, at the time, a little strange to raise at dinner parties: what happens when we teach machines to think using a history that was never fair to begin with?

That question became Man-Made: How the bias of the past is being built into the future. The answer I found was more unsettling than I expected. AI doesn't invent bias. It inherits ours, at scale, and then repeats it back to us dressed up as objectivity.

Has AI bias improved since Man-Made?

I'm often asked whether things have improved since the book came out. The honest answer is this: in some places, yes; in others, we're still finding new ways for the same old problem to surface.

What does AI bias look like today?

Take a study published just last year, where UK researchers tested how a Google AI model summarised real case notes from social services. They changed nothing but the gender description of the person. The AI described men's health problems using words like "disabled" and "complex" far more often than it did for women with comparable needs. Instead, the women were framed as more capable of managing on their own. The lead researcher warned this kind of bias, left unchecked, could mean councils under-allocate care to women.

That's not a hypothetical harm. That's a machine quietly deciding who gets help and who doesn't, based on a pattern it learned from decades of unequal records.

Why is AI bias so hard to spot?

This is exactly the mechanism I wrote about in Man-Made: AI systems trained on historical data don't just reflect the biases of the past, they can actively widen them, because a biased recommendation looks and sounds like a neutral one. There's no raised eyebrow, no audible prejudice. Just a confident, data-backed answer that happens to be wrong in a very old, very human way.

Now, the bias is being amplified by autonomous AI agents. We’re moving into the era where humans are out-of-the-loop.

What needs to change?

The fix isn't to distrust AI wholesale. It's to insist on the same rigour we'd demand of any other decision-maker with power over people's lives: show your working, test for bias before deployment, and keep a human accountable for the outcome.

Three years on from Man-Made, that's still the conversation I want us to be having.