Matteo
Vincenti
Menu

Does everyone use AI? A possible bias among people working in tech

Can a team’s AI habits distort its assumptions about users? Adoption data, perception studies and a method for deciding how much guidance to offer.

A small group of green figures gathers around a computer inside a transparent enclosure, while other wooden figures stand outside.

People who use AI every day to write code, investigate an error or read documentation can come to see it as a routine part of work. It is a short step from there to assuming that others have the same habits. That assumption becomes a product problem when we skip an explanation, leave an empty box for a request and expect the user to know how to begin.

What interests me is how much our professional environment shapes what we take for granted about users. If a feature feels straightforward because we have learned to use it every day, what makes us expect the same of someone encountering it for the first time?

Answering that question takes different kinds of evidence. Adoption data describes how widely a tool is used. Our estimates of other people’s use describe our perceptions. The ability to complete a task needs to be observed on that task. Confusing these questions can lead us to design for a user who resembles us more closely than the product’s audience.

Adoption, how many use AI, calls for a usage survey. Perception, what we estimate, calls for an estimate and a matched measure. Ability, what users can do, calls for observation on the task.

What are we counting when we say “using AI”?

Someone might have tried generating an image, ask a chatbot for an explanation each week or rely on a coding assistant throughout much of the working day. All might say they use AI, while needing different explanations and tools.

The person’s role in the interaction matters too. Receiving an automatically generated suggestion within an application involves a different interaction from writing a request, supplying context and checking the response. Generative AI refers to tools that produce content, such as text, images or code. We need to be clear about whether we are counting deliberate use of those tools or any service that incorporates AI techniques.

ISTAT’s Citizens and ICT survey for 2025 explicitly measures conscious, intentional use of generative tools. It reports that 19.9% of people aged 16–74 in Italy had used them in the three months before their interview, compared with 32.7% in the European Union for the same age group. Fieldwork took place in the first quarter of 2025, so the figure does not describe the situation in November 2026. People outside that percentage may still encounter AI systems in the services they use. Source: ISTAT, p. 6.

Intentional generative AI use in the three months before the interview among people aged 16–74 in 2025. Italy 19.9%, European Union 32.7%. Axis from 0 to 100%.

Italy and the EU use the same age range and usage window. Fieldwork took place in the first quarter of 2025. Source ISTAT, Citizens and ICT 2025, p. 6. Original source.

A 2025 report by Ital Communications, IISFA and Istituto Piepoli instead finds that 77% of its thousand adult respondents use AI technologies. Its question includes virtual assistants, home automation and other uses, without the same three-month window. That figure and ISTAT’s describe different things. The broader definition is a documented difference, but it cannot explain the entire gap on its own: the sample and fieldwork differ too. Source: Piepoli report, pp. 3 and 15–17.

Frequency changes the picture further. In a Bank of Italy survey conducted in August–September 2024, 26% of respondents reported using generative tools in the preceding twelve months; 10% used them at least weekly. The sample came from a household panel and was weighted. The useful comparison is between the two frequencies within that survey. Its different design and period prevent us from reading it as an increase or decline relative to ISTAT. Source: Bank of Italy, QEF 929, §§2–3 and Appendix 2.

Bank of Italy survey in August–September 2024. 26% of respondents used generative AI in the previous twelve months and 10% used it at least weekly. Same denominator, weekly group included in 26%. Axis 0–100%.

Weighted shares from the same Bank of Italy survey. Weekly users are a subset of those who used GenAI in the previous twelve months. Source QEF 929, section 4 and Appendix 2. Original source.

Even an accurate percentage answers only the question it was collected to address. Having used a tool does not tell us whether someone can spot an incorrect response or identify information missing from a request. To decide how much guidance an interface needs, we need to understand that ability.

The tech community offers a particular perspective

In Stack Overflow’s 2026 Developer Survey, roughly 83% of respondents to the question about current use of AI tools at work report using at least one. That question received 17,464 responses. Source: Stack Overflow, AI tool usage.

The survey was promoted through Stack Overflow’s channels and community. It is not a representative sample of everyone working in IT in Italy. The year, geography, question and context also differ from ISTAT’s. Subtracting the two percentages would not measure the gap between IT professionals and the rest of the population. Source: survey methodology.

The design implication remains useful without that subtraction. Before transferring an expectation from our own circle to users, we need to know whom it applies to. A product for experienced developers and one for people who occasionally submit an online request may need very different explanations. The team’s familiarity is not enough to make that choice.

When can we call it a bias?

If we mostly observe people whose habits resemble our own, the group informing our impressions is selected. To identify an overestimate, though, we also need the prediction: how much we think others use AI, and the measure against which that prediction is too high.

The false consensus effect, studied by Ross, Greene and House, concerns the tendency to judge responses similar to our own as relatively common. It offers a possible framework for the question about AI; the original study did not examine IT professionals or the adoption of these tools. Source: Ross, Greene and House, 1977.

Some studies of AI measure these predictions directly. In a 2026 preprint, Albers and colleagues ask university students and teachers in Germany how often they use AI for six activities, from information research to writing, and how much of the work they delegate to it. They then ask each group to predict the other’s answers for the same activities. Both groups overestimate frequency and delegation relative to the other group’s self-reports. Those self-reports, collected from selected samples, remain the reference: the result concerns those activities and participants. Source: Albers et al., methods and RQ3 results.

The opposite direction appears in Innovation without Borders?, a working paper published in June 2026. Business executives in twelve European countries estimate how many firms in the same country, sector and size category had invested in AI by June 2025. Compared with responses collected in an earlier survey wave, underestimation predominates. That reference also carries uncertainty: it is constructed from a sample, with few respondents in some categories. Source: Baumann et al., §§3–5.

These studies measure different things: reported use and delegation among university participants, and firms’ AI investment. They give us a reason to check the direction of error against the relevant measure rather than assume it. They do not establish how IT professionals in Italy estimate use across the population. Greater familiarity might also improve accuracy. That is an alternative explanation to test before attributing a bias to the profession.

What we see also depends on what people disclose

A feed full of discussions about agents and assistants may suggest that those habits are common. Two experiments by Luzsa and Mayr, using constructed feeds about social and political issues, show that content supporting people’s own opinions can strengthen false consensus. These experiments concern brief exposure to feeds about opinions. They do not establish whether algorithms lead developers to overestimate AI use. In two other studies, Bunker and Varnum find a small association between social media use and false consensus, much smaller than participants in a third study estimate. Sources: Luzsa and Mayr, 2021, Bunker and Varnum, 2021.

Existing use can also escape our notice. Across four preregistered experiments, Reif, Larrick and Soll find negative social evaluations of AI users. In the first, people imagining AI use at work report less willingness to disclose it than those using another tool. The research makes it plausible that fear of judgment influences what people disclose; it does not measure how many workers actually conceal their use. Source: Reif et al., PNAS, 2025.

Within a team, then, “nobody talks about it” and “everyone talks about it” are observations about the conversation. Understanding use requires confidential, comprehensible questions about specific behaviours. Understanding usefulness requires a further check: what people can accomplish.

The assumption can end up in the interface

Consider a hypothetical case: a team is developing an assistant that helps people write support requests. The intended result is a ticket containing the problem, the conditions under which it occurs and the information needed by whoever will resolve it.

For the product’s developers, a box labelled “Describe the problem” may be enough. They are used to providing context, adding an error message and correcting an incomplete response. An occasional user might write only “it doesn’t work”. The assistant could then return a fluent description containing details the user never supplied.

Knowing how widely chatbots are used in the population does little to settle this interface choice. We need to know whether the intended users recognise missing information and can distinguish a useful suggestion from an invented detail. Even someone who uses a chatbot daily may lack that familiarity with support requests.

The team can compare free text with a guided flow that first asks what the user intended to do, what happened and which message appeared. The generated text remains editable and is presented as a draft to check before submission. Guidance has a cost: it takes up space, needs maintenance and may slow down people who can already write a complete request. Free text could be the better choice for an experienced audience; for a mixed audience, optional guidance may be worth testing.

An initial test can use the same simulated problem and the same criteria for both versions. The person running it records whether the ticket contains the required information and what help users ask for. They also observe whether users check the draft before submission without being prompted, and whether they notice and correct an incorrect detail inserted for the test. If users are then explicitly asked to look for errors, recognition after that prompt is recorded as a separate observation. They also record the time needed to produce a usable ticket, so any slowdown can be assessed alongside the quality of the request. Assigning the two versions to separate groups with similar experience can limit learning effects. A simulated case avoids having to collect real tickets for this initial check.

Observations need to retain their context: who took part, what task they performed, which version they used and what errors they encountered. A few tests can reveal a specific obstacle; they do not estimate how common it is across the population. They are nevertheless more relevant to this interface choice than a national percentage for AI use in general.

Before choosing, the team could use a worksheet like this:

The assumption can end up in the interface
AssumptionRelevant checkDecision to make
Users know what information to provideTickets produced without help on the assigned taskFree text, guided questions or optional guidance
Users check the draftUnprompted checking before submission, distinguished from recognising errors when askedChecks before submission and instructions on what to verify
Frequent AI users need less supportResults separated by reported experience and observed abilityDifferent paths only if the difference emerges

The aim is to identify what help is useful and who needs it. If someone does not use AI because the task does not benefit from it, adding training or pushing adoption may solve no problem at all.

Testing perceptions without changing the question

Investigating the possible bias requires a different check. A team could first estimate how many of its users deliberately used a generative chatbot to write a support request in the preceding thirty days. A separate survey would then ask users about exactly that behaviour over the same period.

The reference group must stay the same. An estimate about the product’s users cannot be checked against a figure for everyone in Italy; a prediction about weekly use cannot be checked with a question about whether someone has ever tried the tool. Users’ responses also remain self-reports, subject to recall error and nonresponse.

The difference between the prediction and the reported proportion indicates the direction of the error relative to that measure. To compare people working in IT with those outside it, both groups must estimate the same audience and the same behaviour. An average close to the reported figure can conceal individual overestimates and underestimates. Retain the distribution of responses and, for each person, both the signed difference and its size regardless of direction. Recording estimators’ own AI use and the experience informing their estimates helps interpret the results. Keep “don’t know” responses separate from zero.

A correlation between personal use and estimates of others’ use could reflect genuinely different environments. It is therefore insufficient to establish false consensus. The hypothesis that IT professionals systematically overestimate adoption would be weakened if, for the same audience, their predictions were accurate or predominantly too low. Greater accuracy relative to others would instead challenge the idea that technical familiarity makes estimates worse: a group can be more accurate and still overestimate.

If respondents were reached through acquaintances or a community, the result concerns that group. A national conclusion requires a sample appropriate to the population being discussed. A difference between professions alone does not show that working in IT caused it: age, education and context may contribute.

Designing the ticket assistant need not wait for that general question to be resolved. We can already check whether the intended users can produce and review a useful request. My recommendation is to make explicit the abilities we are taking for granted and look for evidence on the task itself. When we decide to omit an explanation, that choice should rest on evidence of what the intended users can do without it.

Back to contents

Share this article

Have a question or a comment about this article?

Message me on LinkedIn
All articles