An article from the Human Era programme

How many Europeans no longer decide without AI? An estimate where no measurement exists

An article from the Human Era programme — on a number nobody measures, and how to estimate it honestly.

Authors

The EPIQ Foundation team

Published

21 May 2026

Document ID

07/2026-0055-677-763

Publisher

EPIQ Foundation

This article was prepared by the EPIQ Foundation’s subject-matter team and put through internal editing and review.

This text is educational and review in character. The figures on decision dependence on AI are model-based estimates built from explicitly stated assumptions, not the result of a direct population study. They are not a clinical diagnosis nor a basis for judging any individual.

Introduction

How many Europeans can no longer make a decision without consulting AI? The question sounds as if it ought to have an answer in a Eurostat table. It does not. No major institution — Eurostat, the OECD, the European Commission, the commercial pollsters — directly measures decision dependence on AI today. What gets measured is tool use, frequency, applications; what does not get measured is what happens to the decision when the tool is taken away. This article sets out what is known for certain, how the number can be estimated honestly — and why the estimate is at the same time a diagnosis of a research gap.

What we know for certain

The starting point is prevalence. According to Eurostat, in 2025 32.7% of EU residents aged 16–74 used generative AI tools, and around 30% of EU workers use AI at work (Eurostat, 2025; JRC, 2025). Against the EU population in that age band this means roughly 110 million users; with the United Kingdom and the rest of Europe, on the order of 140–150 million. Consumer surveys, counted more loosely than Eurostat’s, point higher still.

Use, however, is not dependence. The research literature does not in fact speak of “AI addiction” — it describes a spectrum: reliance, overreliance, automation bias (accepting a system’s output without verification) and decision delegation. Scales measuring the degree of reliance on AI already exist, but they are used in academic experiments; European population statistics still do not exist.

From reliance to dependence

Partial studies show that movement along this spectrum is real. In an experiment on AI advice, participants followed the model’s recommendation in close to 80% of cases even when it was wrong (Klingbeil, Grützner & Schreck, 2024) — the classic picture of automation bias, documented in computer-supported decisions for decades (Skitka, Mosier & Burdick, 1999; Logg, Minson & Moore, 2019). In a survey of knowledge workers, 46% reported regularly using AI to support decisions, and 44% said they begin to doubt their own judgement when AI disagrees with them (Human Clarity Institute, 2026). In the GoTo and Workplace Intelligence study, half of employees admitted relying on AI too much, and 30% that they cannot function without it (GoTo, 2026). A cross-national study of chatbot attachment finds attachment-related behaviours in at least a third of users (Technology in Society, 2026).

All of these figures call for caution: they come from self-report, often from English-speaking or workplace samples, and they measure different things. The same literature also shows that most people can still reject a model’s wrong suggestion — full dependence is not the dominant phenomenon. It is a marginal one, but already countable.

Three estimation models

Since nobody has measured it, a model must do. Take a base of roughly 140–150 million generative-AI users in Europe and ask what share of them could defensibly be described as strongly decision-dependent.

The conservative model. Research on technology behaviour suggests that the strongly dependent group is usually a small fraction of users. At 3–5% this gives around 5–8 million people.

The middle model. If strong dependence affects around 10% of users — a level consistent with the shares reporting attachment and loss of confidence in their own judgement, after discounting overstated declarations — we get around 15 million people, or 2–3% of Europe’s adults.

The aggressive model. Carrying declarations like “I cannot function without AI” (30% of heavy users) straight over to the whole user population — and taking the higher survey-based prevalence estimates — pushes the ceiling to 40–60 million. This scenario has no strong support today; we treat it as a ceiling, not a result.

Two answers cannot be defended: “zero”, because the phenomenon is already visible in partial data, and “30–40% of Europeans”, because there is no evidence for it. Everything in between is a function of assumptions — which is exactly why measurement is needed.

How it should be measured

“Are you addicted to AI?” mostly measures self-presentation. Functional dependence is measured by questions about behaviour: Do you find yourself postponing a decision until you have an answer from AI? — How often do you change a decision you had already made after consulting AI? — When AI is unavailable, do you feel discomfort making decisions? Indicators like these separate using a tool from shifting the weight of the decision onto it — and they allow change to be tracked over time, which at the current pace of AI adoption matters more than any single reading.

The area is unclaimed research territory: no official European indicator of decision dependence on AI exists. A representative population study — working title: AI Decision Dependence Index (ADDI) — would turn the “5–60 million” bracket into a number. Until then, every answer to the title question, including ours, remains an estimate.

Conclusions

1. The phenomenon exists, but it has not been measured. Partial studies show real movement from reliance towards dependence; no European population statistic exists.

2. The most defensible estimate is on the order of the teens to thirty million people in Europe (2–4% of adults), within a wide 5–60 million bracket. It is the output of a model, not a measurement — and should be quoted as such.

3. What needs measuring is functional dependence, not use. The indicator is behaviour — postponing, changing and feeling discomfort about decisions without AI — not a declaration of “addiction”.

Written as part of the EPIQ Foundation research and education programme: Human Era.