Uneven Use, Similar Trust? Generative AI in Denmark and Romania

Synthesis: Denmark and Romania sit at opposite ends of European generative AI use: in 2025, 48.4% of Danes aged 16–74 had used such tools in the previous three months against 17.8% of Romanians, with the EU at 32.7%. Lucau, Voicu and Christensen ask whether that usage gap comes with a matching gap in Trust — and find that it does not. On the KPMG/Melbourne study, which asked both countries identical questions, Romanians trust AI slightly more (4.1 against 3.8 on a seven-point scale), while the Standard Eurobarometer shows Denmark leading by 28 to 45 percentage points on trust in government, parliament and the legal system. What does track trust in AI is confidence in how AI is governed: 47% of Romanians against 36% of Danes call current laws and safeguards adequate, and KPMG's pooled model puts confidence in AI governance at B = .62. Age and education gaps are much wider in relative terms in Romania. The design is a descriptive two-country comparison with no association tested, so it establishes a pattern to be explained, not a cause.

Key Findings

1. Opposite ends of EU use. 48.4% of Danes and 17.8% of Romanians aged 16–74 used generative AI in the last three months; Romanian work (5.2%) and education (3.4%) rates are about 19% of Danish ones.
2. Trust does not follow use. Romanians score 4.1 against Danes' 3.8 on a seven-point trust scale, with 45% against 41% willing to trust AI; differences under 0.2 points are not interpreted.
3. Institutional trust has no AI counterpart. In 2026 Denmark led on trust in government (59% against 31%), parliament (67% against 25%) and the legal system (90% against 45%) but not on trust in AI.
4. Governance confidence moves with trust. 47% of Romanians and 36% of Danes called current laws and safeguards adequate, and KPMG's pooled model makes confidence in AI governance the strongest driver (B = .62).
5. Wider relative gaps in Romania. Youngest-to-oldest use ratio 11.5 against 3.2, high-to-low education ratio 16.2 against 2.4 among 25–64-year-olds, where 2.6% of low-educated Romanians used generative AI.
6. Skills indicators contradict each other. 89% of Danes and 56% of Romanians rate their digital skills sufficient, yet AI knowledge is identical (2.5 of 5) and self-efficacy higher in Romania (4.7 against 4.1).
7. Sources disagree on Romanian use. Eurostat gives 17.8%, KPMG 62% reporting regular or semi-regular AI use (57% in Denmark, reversing the ranking), AtlasIntel 83% using AI in at least one context.

Two cases chosen for contrast

The comparison is a contrasting-case design, not a most-different-systems test, and the authors are explicit that it establishes description rather than cause. Denmark combines high institutional trust, strong self-rated digital skills (89% consider themselves sufficiently skilled) and the EU's highest use; Romania combines lower institutional trust, weaker self-rated skills (56%) and the lowest use, though KPMG finds no gap in AI knowledge between the two. No single survey measures use, trust and skills in both countries, so sources are ranked in two tiers: harmonized cross-national instruments (Eurostat's ICT survey, KPMG/Melbourne panels of about 1,000 per country fielded November 2024 to January 2025, and four Standard Eurobarometer waves) are the only basis for direct comparison, while national and commercial surveys add context within each country and never compare the cases. Gaps are computed as percentage points and as ratios, cells Eurostat flags as unreliable are not interpreted, and differences below the margin of error are left alone.

Where the usage gap actually sits

Denmark's lead is largest in the settings that matter for policy. Private use differs least in relative terms — 34.8% of all Danes against 14.9% of all Romanians used generative AI privately, 43% of the Danish rate — while work use (27.2% against 5.2%) and use in formal education (17.9% against 3.4%) differ far more sharply. Age and education gradients point the same way in both countries but diverge in steepness: 77.6% of Danes aged 16–24 used the tools against 24.2% of those aged 55–74, a 53.5-point gap, while in Romania 44.1% of 16–24-year-olds used them against 3.8% of 55–74-year-olds, a smaller 40.2-point gap that is a far larger ratio. Education behaves similarly: among 25–64-year-olds, Romanian use runs 2.6%, 11.9% and 42.7% across low, medium and high education, against 26.6%, 39.2% and 63.5% in Denmark. The authors stress that purpose-of-use and occupation breakdowns show where adoption happens, not why: they are forms of use, not measures of prior exposure.

Why trust does not transfer

The paper's contribution is to separate constructs that are routinely conflated. Trust in a parliament rests on judgments about integrity and benevolence; trust in an AI system also rests on judgments about capability and about the intentions of whoever built and regulates it, so high institutional trust should not be expected to carry over — the Danish case is the demonstration, with 72% of Danes trusting the news in their country but only 32% trusting ChatGPT. Romania is not simply more skeptical: 55% see robots and AI in the workplace positively against 86% of Danes, the opposite of the KPMG ordering, and that share has fallen four points since 2017 while Denmark's rose four, even though high acceptance of AI is about equal (36% against 33%) and perceived benefits are higher in Romania (3.7 against 3.3 out of five). The authors' reading is domain contingency rather than a stable disposition, and the common thread they find is AI-specific AI Governance: perceived adequacy of safeguards and confidence in the entities developing AI, not general confidence in the state.

Two different problems: access and calibration

The policy reading differs by country, and the paper is careful to keep both halves descriptive. In Romania the binding constraint looks like opportunity and other Digital Divide resources rather than trust: use is concentrated among the young and the highly educated, and 0.1% of low-educated Romanians aged 55–74 used generative AI, a cell small enough that the authors do not interpret its ratio. In Denmark, where use is widespread, the problem shifts to calibration — whether trust matches what the tools can actually do — and to disclosure. Danes overwhelmingly want AI disclosed (87%) but rarely encounter labels (12% report seeing them daily), and the experimental literature the authors review finds that a bare disclosure label can lower trust in the discloser, which argues for contextual disclosure rather than a badge. The paper's own evidence cannot test either policy: it has no Romanian microdata linking use to trust, and the replication with identical items in both countries remains to be done.

What this means for practice

• Policymakers. Fund access and general digital skills for the oldest and least educated in low-use countries before worrying about trust. In Romania, 3.8% of 55–74-year-olds and 2.6% of low-educated 25–64-year-olds used generative AI at all.
• Instructors. Do not read low use as low willingness. Romanians score slightly higher than Danes on trust in AI and on self-efficacy with it, while rating AI in the workplace less positively, so a class or institution can be a more promising site for adoption than a population-level average suggests.
• Administrators. Treat disclosure as a design decision with effects. 87% of Danes want AI use disclosed and 12% see labels daily, and the experimental evidence the authors cite shows a bare label can reduce trust rather than build it.
• Researchers. The comparison cannot identify anything at the individual level. Comparable microdata with identical use, trust, literacy and prior-exposure items in both countries, plus disclosure experiments on Danish and Romanian audiences, are the stated next step.

Limitations

• Two cases and no hypothesis test. All comparisons are between countries or between groups within countries, so the paper cannot show whether the individuals who use generative AI more also trust it more — an ecological inference the authors explicitly decline to make.
• The constructs are not identical across sources. KPMG's trust in AI is not specific to generative AI, its online panels over-represent the university-educated, urban and young, and 56% of the Romanian sample hold a degree against 45% of the Danish one, which may itself raise the Romanian trust scores.
• Sources conflict and one has no trend. Romanian use appears as 17.8% (Eurostat), 62% (KPMG) and 83% (AtlasIntel); Eurostat first asked the generative AI question in 2025, so there is no Tier 1 usage trend, and Eurobarometer samples of about 1,000 per country limit precision for age-by-education cells (margin of error roughly ±3 points), with possible mode effects from Denmark's video interviewing.
• Skills rest on a single self-rated Eurobarometer item that KPMG's AI knowledge and efficacy items do not corroborate, and the Danish EU-SILC series on trust in others carries a break flag and is not used to rank the countries.

Citation

Lucau, I., Voicu, A.-G., & Christensen, S. M. (2026). Uneven Use, Similar Trust? Generative AI in Denmark and Romania. arXiv preprint arXiv:2610.37348.