Ten lessons on public sector AI from our work with education unions

Between July 2025 and April 2026, as part of our project to build worker power to shape public sector AI, Connected by Data and the TUC worked with nine education unions on an action learning programme to test how to catalyse the union movement’s capacity to understand, negotiate and advocate on AI.

This work threw up several lessons that apply more generally to the adoption of AI across the public sector – the risks and benefits; the challenges and opportunities; and the roles that workers can and should play in AI’s development, deployment and evaluation – which I’ve thrown together into this listicle.

1. Public services are not in control of how the public uses consumer-facing AI, but have to respond to it.

Educators are already having to respond to the ways in which students use consumer-facing AI such as ChatGPT to help them with homework and how parents use it to write complaint letters. Their duty of care means they need to understand and respond to the impacts of the use of AI on children’s mental health and resilience (just as they have had to respond to children’s changing media diets, and use of social media). They also need to be thinking about how to equip children with the right skills for the future.

These examples illustrate how public services don’t just need to focus on how they adopt AI, but also on how they respond to the societal adoption of AI, which may change what they need to do and the volume of work they need to handle. In the medical profession, mental health practitioners need to understand how to treat AI psychosis; frontline doctors need to learn how to diagnose patients who report their symptoms using language shaped by AI or have been taken in by AI-generated misinformation. Reduced friction in interacting with public services is leading to additional demands on Employment Tribunals and an increase in quantity and decrease in quality of Freedom of Information requests.

All this means that the government’s approach to regulating consumer-facing AI matters to public service provision, and that at the same time as public bodies are considering how to adopt AI, they also have to work out how to adapt to demands that have changed because of AI.

2. Public services are in a poor state and AI cannot solve some of the deep challenges they face.

The education union’s joint statement produced by this project said “We recognise that AI and education technology (EdTech) can complement human-centred education. But we know that AI cannot solve the education system’s deeper problems: underfunding, teacher shortages, overwhelming workloads, or the social and economic pressures educators face every day.”

There are similar deep challenges across the public sector, brought about by under-investment and changing demographics and requirements.

AI should not be used to work around and temporarily alleviate challenges that require more radical rethinking. Having a clear agreed vision and values, and taking a problem-first, rather than technology-first, approach, helps to align everyone’s incentives towards prioritising changes that will make a difference.

3. The impacts of the adoption of AI by public sector workers go beyond their immediate interactions with AI systems.

The most transformative uses of AI (and other technologies) require end-to-end reinvention of services and processes: the introduction of technology is a trigger for rethinking existing ways of working.

Non-transformative introduction of AI simply automates pieces of processes or layers systems over existing (broken) services. For example, the IT infrastructure in many schools is outdated, so simply introducing an expectation to use AI introduces more work – getting laptops working in a classroom; helping pupils to reset the passwords they’ve forgotten and so on.

The effective introduction of new technologies like AI are organisational change processes. These require the involvement and buy-in of the workforce, which has intimate knowledge of what works and what doesn’t on the ground.

4. All public sector workers are feeling the impacts of AI.

There’s a tendency to focus on AI systems specifically supporting teachers, doctors, or social workers. However, AI is being built into systems across the board, from CoPilot in Microsoft Office, to HR systems and sector-specific administrative software, and therefore affects all public sector workers.

Specialised software, targeted at particular public sector workers, should be co-designed with them. (Beam’s Magic Notes was designed with social workers, for example, and the government has committed to AI tutoring systems being co-designed with teachers.)

On the other hand, when non-specialised software, such as office suites, start to incorporate AI, there needs to be workplace-specific negotiation to co-design policies around its use. This is particularly the case around HR systems that may be introduced to support decision making around recruitment, and job progression, but also applies to developing workplace norms for recording and transcribing meetings, answering emails or writing reports.

5. The introduction of AI in the public sector affects the public as well as public sector workers, and the relationships between the two, which in turn affects public attitudes to the state.

One narrative about AI, evident in the Blueprint for modern digital government, is that introducing AI will allow “public sector workers to focus on delivering the crucial front-line work that will always require human relationships, judgement and empathy”.

The reality of the impact on relationships is more nuanced; other work on pupil attitudes to EdTech reveal that some pupils find it disrespectful for teachers to mark their work using AI, for example – something many technologists (though not many educationalists) might otherwise view as an admin task. Doing a relational activity more efficiently changes the quality of that activity for both parties; optimising for respect may be a way to ensure that AI does not end up diminishing the public’s experience of the public sector.

6. There is the same variability in how public sector workers (and union members) feel about AI as there is in the general population.

Some public sector workers are early and enthusiastic adopters of AI, trying it out even in places where it hasn’t been proven. Some are willing to adopt it but need help and support to feel confident to do so. Some are deeply skeptical, believing that AI is immoral, actively damaging and should be resisted.

This diversity in worker attitudes inevitably leads to variability in use and impacts across settings and brings into play questions around the balance between professional autonomy (both to adopt fast and to not adopt at all); the need to innovate and learn-by-doing; rules and safeguards; and training and support.

7. AI products and services implicitly encode values and political choices, particularly as they reduce relationships between citizens and the state to numbers.

In education, student-facing EdTech can only measure certain things, such as time taken on tasks, percentage accuracy on questions, or when homework (or preparation) is started and completed. This data is readily available, and it therefore becomes something that appears on teacher- and management-facing dashboards.

Goodhart’s Law states that when a measure becomes a target, it ceases to be a good measure. Similarly, when something starts to be measured, it becomes what’s important, especially when evaluating and managing performance. But many things in education are not easily measurable, particularly within applications or through screens. It’s hard for an algorithm to measure conceptual comprehension or identify the missing piece that’s blocking understanding in a student; hard to measure their critical thinking or creative skills; hard to measure how happy they are, how frustrated, how fulfilled, or how much they trust the adults in their lives.

The voices of public sector workers and unions, as well as academics and other civil society voices, are important to work out what’s useful to measure, as well as to challenge the dominance of measurable things in the way we think about what’s important from public services.

8. This all means that the impact of AI interacts in complex ways with the public sector context in which it is used.

The hope is that AI can help under-resourced public bodies catch up; the reality is messier than that.

Some struggling public bodies may be able to use the introduction of AI to strengthen failing IT infrastructure and transform practices and processes that weren’t working, or find that AI fills a genuine gap in their provision, whereas public bodies that are already doing well may find that AI doesn’t add a meaningful benefit on already-great services.

On the other hand, other under-resourced public bodies may be the ones that struggle most with AI adoption, precisely because they are under-resourced. For example, the leadership of under-resourced schools are likely to be too busy fire-fighting to have the time, capacity, or imagination to incorporate AI thoughtfully. This can either mean they don’t adopt AI at all, or more likely adopt it in ways that are more likely to harm, and less likely to help, their students or staff. Meanwhile, more well-resourced schools may be able to have the time and resources to target their adoption of AI more effectively, building on top of strong existing capacity, thereby increasing the gap between provision between the best and worst schools.

9. The evaluation of AI tooling is particularly challenging in these circumstances.

First, real-world evaluations of AI tools are lacking, and public bodies such as schools are adopting AI without evidence, as the result of persistent sales tactics, political pressure, and a fear of missing out.

Second, evaluations necessarily trail implementation, and traditional evaluation timelines and techniques do not meet the demands of rapid innovation and adoption. Individual applications also change underneath the feet of public bodies as new versions are developed to take advantage of new technologies: something with proven benefits that they adopt today, can change tomorrow as new features are added or algorithms tweaked.

Finally, evaluations tend to be narrow and seldom account for the varied levels of impacts discussed above. All this means that evaluation has to be rethought, to get meaningful and timely evidence into the hands of decision makers faster. Worker and union voice is key to reflecting the felt impacts of AI adoption, and mechanisms are needed for them to be heard within the real-world evaluation of these tools.

10. Unions have a particular role to play in looking beyond the near-term and individualised impacts of AI adoption, towards the systemic effects on the workforce they represent.

For example, unions are attuned to situations in which AI service providers are capturing data that may then be used to create AI systems that replace human workers. Unions think about both deskilling of the existing workforce, and building pipelines of new workers to support future sector needs.

As such, unions are well placed to assist the government to think through the transformational changes brought about by AI, and the steps that need to take place to ensure these are beneficial in the long term.