The policy map has a clear bright spot

Governments have not been idle on AI and work. They have mostly been busy in the same place.

A new OECD review looks across the G7, the European Union, Chile, Colombia, Costa Rica and Mexico. It covers adoption, skills, privacy, discrimination, workplace safety, transparency and social dialogue.

The strongest activity is around helping companies use AI and teaching people to work with it. Every surveyed country promotes adoption through a national strategy, and nearly all provide targeted support to smaller businesses.

All also have reskilling or upskilling programmes. Some fund courses, some work with technology companies, and some use public employment services to match people with jobs.

That is the easy part to announce and count. It is much harder to say what an employer must explain when an algorithm screens a candidate, watches a worker or changes how their performance is judged.

Training still leans towards the specialists

The OECD finds a second gap inside the skills push.

It cites an earlier survey of 21 OECD countries in which 14 governments had publicly funded AI training. Nine programmes aimed to create AI professionals. Seven focused on general AI literacy.

Among G7 countries, the review says France and Japan reported general literacy training, Germany supported both literacy and specialist skills, while the United Kingdom and United States concentrated on professional skills.

That balance may miss the larger group. Most workers will not build a model. They will need to know when to trust one, what information not to give it and how to challenge a decision shaped by it.

A short technical course is not the same as practical confidence. The report points to career guidance, train-the-trainer programmes and public-private partnerships as useful approaches that remain underused.

Old rules are carrying much of the load

On privacy, discrimination and occupational safety, most governments are starting with laws that already exist.

That is not necessarily a weakness. A hiring model should not get a new right to discriminate because its scoring method is complicated. Workplace data does not stop being personal data when a prediction is built from it.

The problem is translation. The OECD says the implications of general privacy rules are not always clear to employers or workers when AI is involved. Many countries have added workshops, guidance and self-check tools, but the level of detail varies.

Risk assessments are recommended or required across every surveyed country in the areas the report examined. Audits, certification and security measures are also common on paper.

The review does not test whether those checks are good, independent or routinely used. It maps the framework. It cannot yet tell us whether a worker gets a better outcome because the framework exists.

Being told about AI is not the same as having recourse

Most surveyed countries require or encourage employers to tell workers when AI is being used. Some extend that notice to job applicants. Explicit provisions for platform workers are less common.

The stronger frameworks add human oversight, records and a way to contest a decision. Those details matter when a system ranks applications, assigns shifts or flags a worker as a risk.

Broad promises of transparency can still leave someone with a notice they cannot use. Knowing that software helped reject an application does not explain what information mattered, whether it was correct or who can change the outcome.

The OECD's own AI principles call for people to understand when they are dealing with an AI system and, where useful, to receive information that lets them challenge its output.

The new review says practical, sector-specific examples are still missing in many places. The law may point in the right direction while the workplace process remains vague.

The people already in the job are easy to overlook

Governments usually rely on standard employment support and social protection when AI displaces a worker. Far fewer measures address people who keep their job but see it substantially reorganised around AI.

That group may face quieter changes: less control over pace, new monitoring, unfamiliar performance scores or the gradual removal of tasks that once helped them learn the role.

Worker consultation and collective bargaining can turn general rules into decisions about a real workplace. Yet the OECD finds that relatively few countries invest in giving unions and other social partners enough AI expertise to do that well.

The report is a policy inventory, not an impact study. It does not rank countries and cannot show which measures raise wages, protect autonomy or improve productivity. Those are open questions, and governments have done little evaluation so far.

The pattern is still useful. Training and adoption have moved from speeches into programmes. The rights and routines around workplace AI now need the same practical attention.

Sources

  1. OECD — Recent policy developments on AI in the labour marketPrimary publication page for OECD Artificial Intelligence Papers No. 63, released 24 July 2026.
  2. OECD — Full workplace AI policy reviewPrimary 64-page comparative review. Source for scope, methods, findings on adoption, skills, safeguards, transparency, social dialogue and remaining policy gaps.
  3. G7 — Action Plan for AI in the World of WorkPrimary G7 labour ministers' declaration and action plan that frames the six policy areas used in the OECD review.
  4. OECD AI Principle — Human capacity and labour transitionPrimary OECD principle on skills, fair transition, social dialogue and responsible workplace use.
  5. OECD AI Principle — Transparency and explainabilityPrimary OECD principle on disclosure, understandable information and the ability to challenge AI outputs.