Why are older adults using generative AI less? OECD data show a fivefold age gap
Across countries with available data, 14% of people aged 55–74 used generative AI in 2025, compared with 68% of 16–24-year-olds. The comparison measures recent use, not benefit, skill or need, and country coverage is uneven.
The short version
- What happened
- The OECD reports that 14% of people aged 55–74 used generative AI in 2025, compared with 68% of 16–24-year-olds; use among people aged 75 and over was as low as 1% where data were available.
- Who it affects
- Older people should not lose access to banking, health or public services because an AI interface becomes the default.
- What remains unproven
- Country coverage, survey years, age bands and reference periods are not fully uniform.
Keep in mind
Country coverage, survey years, age bands and reference periods are not fully uniform.
The spotlight does not publish one common respondent denominator for the cross-country GenAI average.
Use during a reference period does not measure skill, frequency, benefit or safe task completion.
The report is descriptive and cannot establish that age causes the adoption gap.
Source analysis · 1 linked source. Primary analysed source date: 5 November 2026.
Read the full report · 8 min The evidence, methods, implications and limitations
The direct answer: the access gap narrowed, but the frontier moved again
Generative AI use remains sharply divided by age even as older adults have moved much closer to universal Internet access. The OECD's Digital Economy and Society Outlook 2026 reports that 14% of people aged 55 to 74 used a generative AI tool in 2025, compared with 68% of 16- to 24-year-olds. That makes the younger group almost five times as likely to report recent use. Across OECD members with data for people aged 75 and over, use was as low as 1%. The comparison covers tools such as ChatGPT, Copilot, Gemini, Llama, Midjourney and DALL-E used to create text, images, code or video.
The result does not mean that most older adults are digitally disconnected. In the same OECD data, the share of 55- to 74-year-olds who used the Internet during the previous three months rose from 58% in 2015 to 87% in 2025. Daily or near-daily use reached 78%. The evidence instead shows a recurring pattern: access to an established technology can become widespread while a gap opens around the newest services. That distinction matters for public policy because providing a connection is not the same as making a fast-changing tool useful, understandable, accessible or trustworthy.[1]
What the OECD measured—and the denominator it cannot make uniform
The central figures come from the OECD ICT Access and Usage Database rather than a single new survey conducted identically in every country. For generative AI, the indicator asks whether an individual used a tool to create content during the previous three months. The report aggregates the countries for which comparable data were available, but the spotlight does not provide one common respondent denominator. Coverage differs by indicator and country. That is why the percentages should be read as cross-country descriptive averages, not as a census of everyone living in the OECD area.
Two exceptions illustrate the comparability problem. Colombia's GenAI data refer to 2024 and use a 12-month reference period rather than three months. Japan's data also refer to 2024, do not specify a reference period in the note and cover people aged 20 to 69. Other charts in the spotlight mix 2025 data with the latest available year for some countries. The OECD identifies those departures, which is good statistical practice, but they limit fine-grained comparisons. A two-point difference between countries or age groups could reflect measurement as well as behaviour; the fivefold age contrast is large enough to remain material while still requiring those qualifications.[1]
The distinction that matters
One digital divide narrowed; another opened
The OECD compares access to established services with adoption of a newer tool. These measures describe use, not benefit.
Internet
Used in the last three months
87% of ages 55–74 in 2025
GenAI
Created content in the recent reference period
14% of ages 55–74
Younger group
Same GenAI indicator
68% of ages 16–24
Unmeasured
Benefit, harm and sustained useful use
No causal result
Non-use combines preference, skill, awareness and design barriers
The report offers several reasons why adoption may differ. Among older adults, 43% said they saw no need to use AI, 12% said they lacked the necessary skills and 9% were unaware of the technology. Seeing no need is not necessarily a deficit: a person may have no relevant task, prefer a human service or judge the privacy trade-off unacceptable. The survey categories also do not prove which intervention would change behaviour. Training cannot solve an inaccessible interface, and a better interface cannot create value for someone whose needs are already met without generative AI.
Age groups are not homogeneous. People aged 55 to 74 include workers at the height of their careers, recent retirees and people managing very different combinations of health, income and education. The OECD also notes physical and cognitive constraints, technology anxiety, mistrust and limited awareness as possible barriers. Those factors can interact with provider choices: small text, complicated account recovery, voice systems that misrecognise speech, automated support without a human route and uncertain data handling can all turn a technically available service into an unusable one. The report synthesises these mechanisms; it does not estimate how much each one explains the 14% adoption rate.[1]
Essential services show why the gap matters beyond chatbot novelty
The practical concern is not whether every older person adopts a fashionable chatbot. It is whether AI becomes embedded in services people need. In 2025, 68% of 55- to 74-year-olds used online public services, compared with 81% across all age groups. The age gap widened rather than narrowed in half of the countries with data. Online health-information seeking reached 52% among older adults versus 62% overall. Internet banking reached 62% among older adults, nearly double its 2015 level, while online purchasing reached 54%. These are substantial gains, but they also create pressure to preserve usable alternatives as providers automate interfaces and reduce in-person channels.
Generative AI could make some services easier through plain-language explanations, voice control, adjustable interfaces and step-by-step assistance. Those are plausible design opportunities, not outcomes demonstrated by the OECD statistics. The report cites prior research on assistive and conversational systems, but its own cross-sectional adoption figures do not test whether an AI feature improves independence, reduces errors or increases access. The same technology can add friction if it gives unreliable advice, obscures how to reach a person or demands disclosure of sensitive information. Service providers therefore need outcome testing with older users, not an assumption that adding a chatbot constitutes inclusion.[1]
What governments and service providers can use now
The report supports a straightforward operating principle: keep essential services accessible while testing new AI channels with the people most likely to encounter barriers. That means involving older adults in design, offering clear human escalation, measuring task completion rather than login or launch counts, and maintaining non-digital routes where exclusion would carry serious consequences. Training can be delivered through trusted local institutions such as libraries and community organisations, but it should start with practical tasks and informed choice rather than pressure to adopt. Providers should explain what a system does, what data it uses and when its answer needs checking.
The evidence also argues for better measurement. Statistics for people aged 75 and over are sparse, even though that group may face different accessibility and service needs from 55- to 74-year-olds. Country surveys need more consistent age bands and reference periods, and adoption should be separated from effective use. A person who opened a tool once is different from someone who completed a banking, health or government task safely. Governments should measure errors, abandonment, requests for human help, unequal outcomes and the experience of people with disabilities alongside simple usage rates.[1]
What remains unproven—and what would change the assessment
This is an official statistical policy report, not a peer-reviewed experiment. It describes associations across age groups and combines national sources with different fieldwork years. It cannot show that age causes lower adoption, that using generative AI improves well-being, or that persuading non-users would help them. The report's examples of possible benefits—including accessibility, health support, independent living and reduced isolation—are hypotheses and evidence syntheses rather than effects measured by the 2025 usage indicator. It also does not calculate the environmental, privacy or fraud risks that additional adoption could create.
Confidence would increase with harmonised surveys that publish country and respondent denominators, oversample people aged 75 and over, and distinguish experimentation from sustained useful use. Randomised or carefully controlled service trials should measure whether accessible AI interfaces improve task completion without increasing errors, scams, privacy loss or dependence on automated advice. Longitudinal work should test whether today's age gap narrows as cohorts change or reappears with each new interface. Until that evidence arrives, the defensible conclusion is modest: older adults are online in large numbers, but generative AI use is far less common, and inclusive design must be demonstrated rather than assumed.[1]