AI Bias: The Concern With the Narrowest Gap

One-line job: Understand where AI bias sits in public opinion — the most agreed-upon concern — and why agreement hasn’t fixed it. Audience: Teams deploying AI in decisions about people. Not for: Readers wanting a bias-audit methodology. This is the sentiment and regulatory picture. Last verified: 2026-08-13 Evidence weight: documentation-verified

Remember the job-displacement gap? Fifty points between public and experts. Bias is the opposite case: 64% of U.S. adults and 73% of AI experts are concerned about AI bias against certain groups — a nine-point gap, among the narrowest measured.

That alignment should make this the easiest concern to act on. Everyone already believes it’s real.

The global weight

KPMG’s 2025 study (n=48,340 across 47 countries) puts manipulation or harmful use at the top of the global risk list — 85% listing it as top-tier — with bias concerns embedded throughout. Reuters/Ipsos polling (August 2025, via Pew/Stanford aggregation) documented AI generating racist arguments as a specific, named public concern rather than an abstract category.

There’s also a generational signal worth watching: Gallup’s Voices of Gen Z 2026 found 42% of Gen Z believe AI hurts their ability to think carefully about information versus 25% who think it helps. The most AI-native generation is also the most skeptical of what it does to judgment — a finding that should unsettle any assumption that familiarity breeds acceptance.

Where the law landed

The EU AI Act drew a bright line in Article 5(1)(g): biometric categorization systems that deduce race, political opinions, religious beliefs, sex life, or sexual orientation from biometric data are prohibited outright. Not “regulated” — prohibited. The reasoning treats inference of protected characteristics from faces or bodies as a harm regardless of accuracy or intent.

For anyone deploying models that process people-images in Europe, this is a hard boundary that predates any audit you might run.

Why agreement hasn’t produced fixes

Here’s the puzzle the narrow gap poses. Public and experts agree bias is a problem. So why does it persist? Because the disagreement starts after the concern — at measurement. What counts as fair when groups differ in base rates? Whose ground truth? Audit standards remain contested even among people who share the values.

The practical consequence for operators: you can’t outsource your fairness position to consensus, because there isn’t one past “it matters.” There are only documented choices, tested systems, and honest disclosure of known limitations.

The call

1. If your model touches decisions about people, document what you tested for bias against — and what you didn’t. The untested list is the first thing an auditor or journalist asks for.
2. Check EU AI Act Article 5 applicability before shipping anything that categorizes humans from images. Some capabilities are simply off-limits there.
3. Treat Gen Z skepticism as a leading indicator, not a generational quirk. Your next decade of users starts from distrust of AI-mediated judgment.
4. Publish limitations. In a narrow-gap landscape, honesty about known bias costs little credibility and buys the trust that silence burns.