Gillespie et al. — Trust, attitudes and use of AI: a global study 2025 (University of Melbourne / KPMG)
TL;DR
The fourth wave of the University of Melbourne’s trust-in-AI survey, run with and funded by KPMG (Chair in Trust partnership). The research team, led by Nicole Gillespie and Steve Lockey, states that it designed, ran and reported the study independently. 48,340 people in 47 countries answered an online survey between November 2024 and mid-January 2025. The report has three parts: the public, employees (Section 2) and students (Section 3). For 17 countries it compares with a wave run in late 2022, just before ChatGPT, which gives it a rare before-and-after on the same questions.
The wiki already cited it second-hand: ai-knowledge-hiding quotes its 57% figure through Anicich & Brouwers. This page adds the primary source.
Headline findings (executive summary):
- Trust is the problem; acceptance is not. 54% are “wary about trusting AI”, yet 72% accept its use. People trust AI’s technical ability more than its safety, security and societal impact.
- Use outruns literacy. 66% use AI intentionally on a regular basis and three in five say they can use it effectively, but 61% have had no AI training and about half report limited knowledge. Half do not know AI underpins social media they use.
- Emerging economies lead on use (80% vs 58% regular use), training (50% vs 32%), knowledge, efficacy, trust (57% vs 39%) and acceptance (84% vs 65%). The differences hold controlling for age and education.
- Benefits and harms are both experienced. 42% think the benefits outweigh the risks, 32% the reverse, 26% balanced. Four in five are concerned about negative outcomes and two in five have experienced or seen them.
- A mandate for regulation. 70% say regulation is needed; only 43% think current laws are adequate; 87% want laws against AI-generated misinformation.
- Trust predictors. Of four pathways (knowledge, motivational, uncertainty reduction, institutional), the institutional one, the perceived adequacy of safeguards and confidence in those developing AI, “had the strongest influence on trust”. The same model holds inside organisations, where the institutional pathway is governance, strategy and training.
Trend 2022 → 2024 (17 countries). Adoption rose and trust fell:
| Indicator | 2022 | 2024 |
|---|---|---|
| Employees using AI for work (any use) | 54% | 67% |
| Employees reporting their organisation uses AI | 34% | 71% |
| AI systems seen as trustworthy | 63% | 56% |
| Willing to rely on AI systems | 52% | 43% |
| Worried about AI | 49% | 62% |
Employee trust in AI at work also fell (mean 4.5 → 4.3), and perceived organisational support for responsible AI fell in 9 of 17 countries, including the Netherlands (mean 4.2 → 3.7). The authors read this as “a healthy recalibration”: the hype “may be giving way to a more realistic and measured assessment.”
Employees (Section 2)
- Adoption. 58% of employees use AI intentionally at work on a regular basis, a third weekly or more. 70% use free public tools, 42% tools provided by the employer (Figure 40). Emerging economies lead (72% vs 49%).
- Policy lags use. Asked whether their organisation has a policy or guidance on generative AI (Figure 41, checked against the PDF): 34% say a policy guides its use, 6% a policy bans it, 41% no, 19% don’t know. In advanced economies about one in two report training, policies or a strategy for responsible AI.
- Complacent and inappropriate use (Figure 44 and p. 74):
- 48% have uploaded sensitive company information (financial, sales, customer data) or copyrighted material into public AI tools;
- 47% have used AI in ways that could be considered inappropriate; 63% have seen colleagues do so; 56% have used AI at work without knowing if it is allowed;
- 57% have used AI non-transparently: presenting AI output as their own or not revealing they used it;
- 66% relied on AI output without evaluating it; 72% put in less effort because of AI; 56% made mistakes in their work due to AI.
- Bans make it worse. Uploading company data to public tools is most common where the organisation has banned generative AI (67%) or has a policy (56%), against 33% where there is none: “outright bans may be ineffective, and … simply having policies does not guarantee compliance.”
- Pressure. 48% worry about being left behind if they do not use AI, and strain at work correlates with complacent use (r = .31).
- Benefits. Most report gains in efficiency, information access, innovation and quality; almost half say AI has increased revenue-generating activity. Half use AI instead of asking colleagues; one in five report less communication and collaboration. Two in five expect AI to replace jobs in their area.
- Caveat from the authors. The survey was anonymous, but “these findings may underreport the actual extent of complacent and inappropriate use … given social desirability bias.”
Students (Section 3)
83% of students (mostly tertiary) use AI regularly. Most report efficiency and learning benefits, but misuse is common: two-thirds have hidden their AI use, over three-quarters felt they could not complete work without it, and four in five put in less effort knowing they can rely on it. Only half say their institution has policies or training.
The Netherlands
The Netherlands has the lowest regular use of any advanced economy (43%), against 73% in Singapore (p. 20). In Appendix 3 (p. 109, checked against the PDF): trust 3.6/7, acceptance 3.0/5, AI knowledge 2.5/5, and 24% with AI training or education. Between 2022 and 2024 worry about AI rose sharply there (one of four countries with a 21–26 point increase), and employees’ perceived organisational support for responsible AI fell.
Dynamic-capabilities reading
- contextual/internal-barriers — low trust, missing policy and training, and pressure to use AI are measured barriers; bans that increase risky use are a barrier created by the organisation itself.
- digital-transforming/improving-digital-maturity — AI literacy as the “cross-cutting enabler”: associated with more use, trust, critical engagement and realised benefits; governance and training lag adoption.
- strategic-renewal/organizational-culture — non-transparent use (57%) and complacent use point to culture and accountability, not tooling.
Related in this wiki
- Anicich & Brouwers (HBR) use this survey’s 57% as their starting fact.
- Richardson (Reuters) shows a later trust series (Gallup/SCSP, 2025–2026) in which trust among non-users halved.
- Mollick (HBR IdeaCast), in this batch, argues that people hide AI use when culture and rewards punish disclosure.
- Chesky and Collison, in this batch, describe AI use from inside two technology firms; this survey is the population view.
Linked entities and concepts
- Concepts: ai-knowledge-hiding, responsible-ai, enterprise-ai-adoption, technology-adoption-theories, ai-deskilling.
- Dangling (single-source mention, deferred): Nicole Gillespie, Steve Lockey, Tabi Ward, Alexandria Macdade, Gerard Hassed, KPMG, University of Melbourne.
Debates and supersession
- Sponsor. KPMG funds the research and sells AI governance and assurance services; the report’s recommendations (assurance mechanisms, responsible-AI training) match that offer. The university team states its independence, and the method and sample are documented (Appendices 1–2).
- Self-report. Use, misuse and mistakes are self-reported; the authors expect under-reporting of misuse.
- Trend comparisons cover 17 countries only, and the 2022 wave predates ChatGPT, so part of the change is a change in what “AI” means to respondents.
- Data date. Fieldwork ended January 2025. Later sources in the wiki describe agentic tools that did not exist then.
What was actually ingested
The full 118-page PDF (A4, created 28 April 2025), fetched from the KPMG asset URL and converted with pdftotext -layout. The table of contents ends at Appendix 4 (p. 110), so the file is complete. The executive summary, method section, the employee section’s adoption and complacent-use pages, and the 2022–2024 trend pages were read; Figure 41 and the Appendix 3 country table were checked against rendered pages, because the text extract flattens charts. The public-attitudes chapters on emotions, regulation and misinformation and the student section were read only through the executive summary.