Aware of the risks, yet using AI: interim results of the AI 2026 survey
How do people in companies perceive generative artificial intelligence – and do the risks put them off using it? These questions are addressed by the AI 2026 survey, which we conducted from July to September 2026 among 1,000 people working in Slovak firms with 20 or more employees. Data were collected by MNFORCE. The analysis is still ongoing and we are preparing a journal article from it. Here are the first results.
Use. Fewer than 4% of respondents never use ChatGPT; the figures are 12% for Gemini, 30% for Claude and 41% for Perplexity. For ChatGPT, 57% of respondents placed themselves in the upper part of the seven-point frequency scale (5–7, where 7 means “many times a day”).
Benefits and risks are not opposites
Technology research usually assumes that people who see large benefits in a technology perceive smaller risks, and vice versa. Our data do not show this. Perceived benefits (12 items) and perceived risks (15 items) are positively related (r = .46) and measure two distinct evaluations rather than two ends of a single axis.
When we grouped respondents by their combination of benefits and risks (latent profile analysis, 925 respondents with complete data), five profiles emerged. The largest group is the pragmatic mainstream, with neutral views on both. The second largest – almost a third – consists of people who see high benefits and high risks.
Each dot is one respondent (n = 925); black dots are profile centres. Dashed lines mark the scale midpoint (4). Clusters in the corners are respondents who gave the same answer to all items.
Sensitivity to risk does not prevent use
People sensitive to risk are not a brake on adoption. Their intention to use AI (5.66 on a 1–7 scale) and their use frequency (4.11) are close to those of enthusiasts, who perceive hardly any risk (5.98 and 4.40). The difference between the two groups was not statistically significant. Skeptics, who see high risks and low benefits, use AI markedly less.
Means with 95% confidence intervals, accounting for uncertainty in profile assignment. Intention: 3 agreement items; frequency: mean across ChatGPT, Gemini, Claude and Perplexity (1 = never, 7 = many times a day).Table of values
Profile
Share
Benefits
Risks
Intention
Use
Enthusiasts
9%
6.04
2.29
5.98
4.40
Risk-aware adopters
31%
5.68
5.69
5.66
4.11
Pragmatic mainstream
41%
4.18
4.10
4.27
3.61
Skeptics
6%
2.65
6.02
3.27
2.80
Disengaged
13%
1.72
1.80
1.85
3.45
Risk matters only when benefit is missing
Regression models confirm this from another angle. After accounting for other factors (such as performance expectancy, social influence and organisational AI readiness), total perceived risk on its own is not related to AI use. Its relationship with use, however, depends on perceived benefits: higher risk goes with less frequent use only among people who rate the benefits of AI below about 4.4 on the seven-point scale. Above this threshold – that is, already at a mildly positive view of the benefits – risk no longer distinguishes users from non-users. The result held in ten of eleven model specifications.
Not all risks work the same way, though. People weigh against the benefits mainly the risks that can harm them directly – data protection, errors they are accountable for, or the loss of their own skills. For risks to society and the environment (impact on jobs, bias, energy use) the relationship is markedly weaker.
What follows
Perceived risk is not a general brake on AI use but a conditional cost whose weight depends on the expected benefit. For organisations, this means that measures aimed at reducing concerns are related to use mainly where people do not yet see what AI is good for. Where they do see the benefit, it makes sense to manage risks in use – through rules, training and checking outputs – rather than trying to restrict use.
The data are cross-sectional, so the results describe associations, not causes. Respondents rated benefits, risks and use themselves. The profiles are a statistical simplification, not fixed types of people, and their shares cannot be generalised to the whole population without reservation. We presented the results at the 57th EBES Conference in Prague (15–17 October 2026). The next step is repeated measurement over time.