Can AI be biased against the people it affects?

Short answer

Yes — concern about AI bias against certain groups is consistently high across multiple major surveys, and the gap between AI experts and the general public on this question is unusually narrow. In April 2025, ~64% of U.S. adults and 73% of AI experts said they were highly concerned about AI bias against certain groups — a 9-point gap, much smaller than the 50-point gap those same groups show on AI’s effect on jobs. The worry is rising, not falling, and at least one major regulator has started to write bias-style harms into binding law.

Why people are concerned

Three things drive public concern.

First, the public has watched specific incidents. The August 2025 Reuters/Ipsos poll, cited through Pew’s aggregation, recorded widespread worry that AI can be used to generate racist arguments and to stir up political chaos — concrete, recognisable harms rather than abstract ones.

Second, the worry is global. KPMG’s 2025 study, conducted with the University of Melbourne across 47 countries (n=48,340), found that 85% of respondents globally listed manipulation or harmful use as a top-tier AI risk. Bias and representational harm sit inside that “harmful use” cluster.

Third, the workplace is where many people expect to feel it directly. In February 2025, 52% of U.S. workers said they felt worried about AI use in the workplace, compared with 28% who felt hopeful — and the most exposed roles are the ones where bias-intersection concerns (race, class, disability) already exist.

What is true

• The public-expert gap on bias is narrow. Pew’s April 2025 survey found 64% of U.S. adults and 73% of AI experts were highly concerned about AI bias against certain groups — a 9-point gap.
• Concern is rising, not declining. KPMG’s 2025 study reports that the share of people saying they were worried about AI rose from 49% in 2022 to 62% in 2024 across 13 of 17 longitudinal countries. The view that AI’s benefits outweigh its risks fell from 50% to 41% over the same period.
• Generational skepticism is climbing. Gallup’s 2026 Voices of Gen Z survey found that 42% of Gen Z respondents believe AI hurts their ability to think carefully about information, compared with 25% who believe it helps.
• One regulator has already acted on a specific bias-style harm. The EU AI Act’s Article 5 prohibits biometric categorization systems used to deduce race, political opinions, religious beliefs, sex life, or sexual orientation from a person’s biometric data. The Act’s penalty schedule for prohibited practices is up to €35 million or 7% of worldwide annual turnover, whichever is greater (Baker Donelson summary).
• Cross-jurisdictional regulator trust varies widely. Across 25 countries surveyed in October 2025, a median of 53% of adults trust the EU to regulate AI, 37% trust the United States, and 27% trust China.
• Concern about too little regulation outweighs concern about too much. Stanford HAI’s 2026 AI Index, restating the same Pew numbers, reports that 41% of Americans say federal AI regulation will not go far enough, compared with 27% who say it will go too far.

What is exaggerated, misleading, or unsupported

• “Everyone wants AI banned.” Not established. KPMG’s 2025 study reports that 58% of respondents on average still view AI systems as trustworthy, and Stanford HAI 2026 restates a global figure of 59% saying AI offers more benefits than drawbacks. Concern about bias and approval of AI can rise at the same time.
• “AI is racially biased across all use cases at the same rate.” Unsupported in this evidence base. The surveys cited above measure concern, not demonstrated bias rates. Community-specific bias claims require per-product audit data, not survey aggregates.
• “The public and experts disagree on bias.” Misleading. They disagree strongly on jobs (a 50-point gap), but on bias they converge within 9 points.
• “AI bias is a U.S.-only worry.” Not established. Pew’s U.S. data, Pew’s 25-country data, KPMG’s 47-country data, and Eurobarometer’s EU data all show similar levels of concern.
• “The Reuters/Ipsos August 2025 numbers prove what people think about every AI product.” Too broad. The Reuters-hosted poll page is access-restricted on this network; figures are cited here through Pew’s aggregation of related AI-risk concerns (Pew aggregation), not as a standalone primary statistic.

What remains uncertain

• Specific per-community bias rates. Concern is well-documented; the actual measured bias rate for any given AI product against any given group is not in the survey data and would need product-level audits or regulator findings.
• Enforcement of EU AI Act Article 5(1)(g). The prohibition exists, but as of mid-2026 there has been no major published enforcement action. What “deduce” means in practice is still being tested.
• Whether new generative-AI products have moved the needle. Edelman 2026 lists misinformation (50%) and the growing use of generative AI (37%) among the top events shaping trust over the past five years, but the connection between those shifts and bias-specific worry is correlational, not causal.
• Whether the public-expert gap on bias stays narrow as capabilities change. Today’s 9-point gap is a snapshot. Both groups could shift, and in different directions.

Where we are likely headed (editorial judgment, 2–5 years)

This section is editorial judgment, not a research result. Looking two to five years out from mid-2026:

The direction of travel on bias-specific regulation looks like expansion, not retreat. The EU’s Article 5 model is likely to spread, in pieces, into other jurisdictions — at minimum the high-profile parts (biometric categorization, untargeted facial scraping, social-scoring-style systems). In the United States, the 41%-vs-27% lean toward “more regulation, not less” suggests bipartisan appetite for narrowly scoped bias rules even when broader AI bills stall. Workplace AI bias, where the most people actually feel exposed, is the most likely site of the next concrete rules.

Public concern looks unlikely to fall on its own. KPMG’s two-year rise from 49% to 62% worried, plus Gen Z’s flat or hardening skepticism, suggests baseline concern is now structurally higher than the 2022 level. Products that ship visibly fairer outcomes may move the number; products that ship visibly biased ones will accelerate the trend.

The gap between expert and public views is the part most likely to wobble. If generative-AI capabilities produce widely-publicised failures, public concern may run ahead of expert concern; if the public sees credible audits and enforcement, the gap could close further. Watch product-level incidents and first EU AI Act fines as the leading indicators.

What this means for people and small businesses

• Treat AI decisions that affect you as contestable. Hiring, lending, housing, and insurance AI systems increasingly have a documented paper trail. Ask which system was used, what data trained it, and how to request a human review.
• Pay attention to the regulator you trust most. Pew’s 25-country data shows trust in AI regulators ranges from 27% (China) to 53% (EU) across the median adult. Where you live changes which rules apply and which regulator will hear a complaint.
• For small businesses building or buying AI tools: bias exposure is now a regulatory risk, not just a brand risk. Under the EU AI Act, prohibited-practices fines reach up to €35 million or 7% of worldwide annual turnover. Even outside the EU, the U.S. public leans toward more regulation (Stanford HAI 2026).

Bottom line

AI can be biased against the people it affects — that is the established direction of both the evidence and the law. The worry is real, broad, and rising; the gap between what experts and ordinary people think is unusually small; and the EU has already turned one specific bias-style harm into a hard prohibition. What is not established is that every AI product is biased at the same rate, that bias concern equals rejection of AI, or that the current numbers will hold as generative-AI products change. The practical stance is to take bias seriously without panicking about AI as a whole, and to watch first EU AI Act fines and product-level incidents as the next round of evidence.