Why does AI trust keep falling even though it matters most?

University of Melbourne researchers Nicole Gillespie and Steve Lockey, working with KPMG International, surveyed 48,000 respondents across 47 countries for this 2025 study, tracking public trust, attitudes and AI use since 2020. The headline claim is that "trust is the strongest predictor of AI use and acceptance, earned through transparency, accountability and the consistent, responsible application of AI technology." But the study finds trust is not holding: "54 percent of the respondents say they are wary of AI, and the perceived trustworthiness of AI systems has fallen from 63 percent in 2022 to 56 percent in the 2024 survey."

The report's causal chain runs two ways. Low AI literacy suppresses trust ("AI education and training opportunities enhance AI literacy and can strengthen people's trust"), while unaddressed risk concerns suppress it further: "the more concerns there are among individuals and organizations about the risks and potential negative outcomes of AI use... the less likely they are to trust or accept AI systems." The study reports 79 percent of respondents are concerned about risks and 43 percent say they have personally experienced a negative outcome — loss of human connection, inaccurate output, privacy loss, misinformation — which it treats as the proximate driver of declining trust. Worth noting: KPMG, which co-produced the study, also sells AI strategy and governance consulting to organizations navigating the trust problem it describes.

This gives population-scale, self-reported evidence for the risk-trust link that Which AI risks are already harming individual users today? documents from an expert panel — the same risk categories (privacy, misinformation, lost human connection) show up here as lived public experience rather than anticipated expert concern. It also runs parallel to What collaboration level do workers actually want with AI?, which measures acceptance by desired collaboration level rather than aggregate trust — together the two suggest trust and preferred autonomy are separate dimensions of acceptance that can move independently. And it echoes the stakeholder-gap pattern in Do hiring managers and job seekers agree on AI fairness?, another vendor-commissioned survey where confidence diverges sharply by position rather than converging on one figure.

The excerpt reports attitudes and self-assessed trust, not measured behavior or independently verified harm — "43 percent experience negative outcomes" is what respondents say happened to them, not an audited incident count. The 63-to-56-percent trend compares two survey waves (2022 and 2024) without detail on whether question wording or sample composition held constant, and it is a single global average across 47 countries that the report's own regional breakdown (82 percent of emerging-economy respondents report AI benefits versus 65 percent in advanced economies) shows masks sharply divergent regional trajectories. The finding supports the narrower claim that self-reported trust declined in aggregate across this sample, not that any one driver — risk exposure, literacy, or regulation — explains the decline on its own.