AI training may be manufacturing its own age gap

Run this audit before diagnosing reluctance

The immediate management response is not an age-targeted remedial course. It is an audit of the opportunity chain. Track conversion at each stage and make ‘not offered’ different from ‘declined’. Use rates, not headcounts, and compare like with like: age band within role, grade, location, contract type and, where available, prior digital exposure. Small groups should be suppressed or combined to protect privacy.

Compare rates within comparable roles, grades, locations and contract types; suppress or combine small groups to protect privacy.

Interpret the funnel within comparable work, not from unadjusted totals. Record offer, access and application as separate events.

Read the result correctly

The useful diagnosis is the first point at which comparable cohorts diverge. Do not jump from a difference at one stage to a story about the people in it.

• Leakage at eligibility or invitation: The programme is filtering people out before they can choose. Review role criteria, nomination patterns and direct communications.
• Leakage at encouragement or equipment: Training may exist on paper while effective permission or approved access does not. Align manager signals, licences and data access.
• Leakage at participation or practice: Investigate timing, protected hours, relevance, format and confidence. Keep ‘declined’ separate from ‘wanted to participate but could not’.
• Leakage at application: Treat this first as an implementation problem: workflow integration, use-case quality, suitable data and manager follow-up.
• A residual gap after comparable opportunity: Only now examine differences in appetite, expected payoff, prior digital exposure or retirement horizon. The audit has narrowed the diagnosis instead of presuming it.

Fix the first point of leakage

If the audit finds leakage, the fix should be operational rather than theatrical. Equalise invitation and licence rates within comparable roles. Ask managers to make permission explicit. Give people paid time to test approved tools on a real task. Use live or facilitated practice where self-service content is not converting. Design examples around the work people actually do, and include experienced staff in deciding where AI should not be used as well as where it should.

Then measure application after 30 and 60 days: not just log-ins or completions, but whether work became faster, better or safer. Older colleagues may bring precisely the contextual judgement needed to spot plausible errors, broken process assumptions and customer risks. Treating them only as recipients of digital remediation wastes that asset.⁶

The point is not to guarantee identical adoption at every age. It is to ensure that any gap left on the dashboard is not a shadow cast by the organisation’s own choices.

Why the evidence comes next: The audit is a proportionate management response even if a gap has another cause in a particular organisation. The research below explains why it is warranted and where the case remains incomplete.

The diagnosis starts too late

A familiar dashboard problem is beginning to appear. AI use is lower among older employees, so the organisation concludes that they are less interested, less confident or less adaptable. It responds with more communications, a motivational webinar or a recruitment plan tilted towards ‘digital natives’.

That conclusion may be wrong because the measurement started too late. Usage data shows who crossed the finishing line. It does not show who was entered in the race, invited to the briefing, given a licence, told that experimentation was permitted, shown a relevant use case or allowed paid time to practise.

Low uptake is an observation. Reluctance is an explanation. If an organisation has not measured opportunity, it has not yet earned the right to diagnose attitude.

The age gap existed before generative AI

The starting prior matters. Age inequalities in workplace learning are not new. OECD analysis published in 2025 found that, in 2023, only about one-third of people aged 60–65 had participated in adult learning during the previous year, compared with more than half of 25–44-year-olds. More tellingly, workers aged 55–65 were seven percentage points less likely to participate in training and 14 points less likely to learn by doing than 25–34-year-olds with the same level of education and in the same occupation.¹

That does not prove that employers discriminate when allocating training. Shorter expected job tenure, occupational mix, prior educational attainment, employee demand and the design of the work itself all matter. But it establishes a sensible baseline: when a new form of learning arrives, the burden of proof should not sit with anyone claiming that the old inequality might recur. We should look for evidence that the new system has broken the pattern.

So far, the available evidence points in the other direction.

AI learning shows the same gradient

An OECD poll conducted in early 2025 provides unusually direct evidence. It covered 14,611 people across 14 countries, using nationally representative samples of roughly 1,000 people per country. About 45% of 26–35-year-olds reported undertaking AI-related training in the previous 12 months. The figure was 39% for 18–25-year-olds, 38% for 36–45-year-olds, 27% for 46–55-year-olds and below 20% for people over 55.²

The measure is broad: it captures AI-related learning, not necessarily training provided by an employer. It therefore cannot tell us who was offered a place at work. Even so, the gradient is hard to dismiss as a quirk of one company or one national labour market.

A separate 2024 survey commissioned by learning-platform provider CYPHER Learning found that 30% of workers aged 18–44 had received AI training, against 11% of those over 55.³ Its methodological reporting is thinner, so it should be treated as corroboration rather than the foundation of the case. AARP research, reported by the American Society on Aging, found that only 12% of workers aged 50-plus had undertaken AI-related training, while 49% expressed interest.⁷ The precise populations and questions differ, but the direction is consistent.

The evidence is not sufficient to say that employers systematically offer older workers less AI training after accounting for role, grade, location and contract type. Few public datasets measure the offer itself with that precision. It is sufficient to reject the more convenient assumption that older workers are somehow receiving more AI training and still choosing not to engage. No evidence reviewed for this article showed that reversal.

A course catalogue is not an opportunity

Training is often treated as a binary management variable: available or unavailable. For the employee, it is a chain. A course can exist without being visible. It can be visible without being relevant. It can be relevant without a manager’s permission. It can be completed without access to the approved tool. A licence can be issued without time to practise on real work.

A representative US survey by the Federal Reserve Bank of New York illustrates the scarcity of the formal offer. In late 2025, 38% of employed respondents said AI training was important, but only 15.9% said their employer currently offered any. Thirty-seven per cent said their workplace did not offer AI tools and another 11% said use was actively prohibited.⁵ That is an access problem before it is a motivation problem.

Research on workers in the United States and six European countries goes further. Bick and colleagues found that firm encouragement was the strongest organisational predictor of AI adoption in their data. Among workers who had neither employer-provided training nor tools, 47% had adopted AI when their firm encouraged it, compared with 10% when it did not. Among workers who were not encouraged or trained, adoption was 21% for those given tools and 10% for those without them. Training was not independently predictive once encouragement and tool provision were taken into account.⁴ These are correlations, not proof of causation, but they change the practical diagnosis.

The managerial unit is not ‘a course’. It is effective opportunity: a relevant invitation, explicit permission, suitable tools, protected practice and a route into real work.

How a business can manufacture the gap without intending to

No discriminatory policy is required. An age pattern can emerge from ordinary operating choices:

• Pilot selection: Early seats go to digital teams, innovation volunteers or managers’ usual nominees. If those groups skew younger, the pilot inherits the skew.
• Permission: Some employees hear ‘experiment’; others hear only risk warnings. Informal managerial signals decide who feels authorised to try.
• Tool access: A generic webinar is offered widely, while approved licences, integrations and useful data remain concentrated in a small cohort.
• Practice time: Learning is technically available but expected to happen outside paid work. People with less discretionary time or less confidence in the payoff opt out.
• Relevance: Generic prompting content is detached from the learner’s role. The employee declines the abstraction, not necessarily the technology.
• Measurement: The dashboard begins at enrolment or log-in, making everyone filtered out upstream invisible.

A 2026 Training magazine and Panopto report, based on more than 400 learning-and-development leaders, describes a gap between available training content and employees’ ability to find and apply it. Its public summary does not report results by age. It illustrates possible organisational friction, not an age-specific effect.⁹

The honest counterarguments

It would be equally lazy to assume that every difference is created by the employer. OECD data shows that the desire to participate in learning falls with age: roughly 60% of 25–44-year-olds wanted more learning, compared with 37% of 60–65-year-olds.¹ Some people approaching retirement will rationally place a lower value on a course with a long payback period. Job content, education, confidence with existing digital tools and expectations about whether AI will affect one’s role also vary.

Those are alternative explanations, not a licence to stop measuring. Expressed demand is partly shaped by what people believe will be useful, permitted and rewarded. An employee who has not been shown a relevant application, sees no approved tool in their workflow and expects a manager to regard experimentation as frivolous may quite reasonably say they do not want ‘AI training’. That answer is not a stable personality trait.

There is also constructive counterevidence. AARP analysis reported that the age gap in participation on LinkedIn Learning narrowed sharply between 2022 and 2025, while technology-topic learning among older users increased.⁷ That is not proof of equal workplace access; platform learners are self-selecting. It is evidence against the idea that low learning appetite is fixed by age.

Evidence maturity: Adult-learning and AI-training age gradients are strong signals. Employer encouragement, tools and practice conditions are associated with adoption. The share of an age gap caused specifically by unequal employer offers remains unestablished; treat it as an audit trigger, not a verdict.

Seven questions for the next workforce or AI meeting

• Are we measuring an attitude gap, or an opportunity gap?
• What proportion of each comparable cohort was eligible, directly invited and given an approved licence?
• Can we distinguish ‘declined’ from ‘never offered’ and ‘offered but unable to make time’ in our data?
• Are managers giving the same permission signal across roles, locations and age bands?
• Does the training use real workflows, or require employees to imagine the relevance for themselves?
• Who is represented in the pilot and workflow-design group, not just in the learner population?
• What would we need to observe before concluding that lower use reflects lower appetite?

Meeting-ready line: “Are we measuring an attitude gap, or an opportunity gap?”