Opinion: What AI regulation can learn from a pint of milk

After the president of Microsoft reportedly said AI is less regulated than milk, Professor Elena Gaura, Associate Pro-Vice-Chancellor (Research) and AI Lead at Coventry University Group, explains what food regulation can tell us about how to make AI safer.

When we buy a pint of milk, we do not inspect the dairy ourselves. We rely on a system of hygiene requirements, inspections and responsibilities that extends from production to sale. That system cannot eliminate every risk, but it makes safety an obligation rather than a promise.

What should we expect when artificial intelligence influences decisions about our education, employment or access to services?

The comparison needs careful treatment. AI is already subject to laws, including data protection and equality requirements, and to rules in regulated sectors. The European Union has also introduced its AI Act. Milk and AI present different problems, yet the everyday expectation behind food regulation is relevant:

People should not have to become technical experts to receive basic protection.

Responsible AI requires regulation and organisational competence

For AI, people need evidence that systems are fit for their intended use, have identifiable responsibility, and there is a clear way to challenge harmful decisions.

Delivering those protections requires more than passing legislation. Organisations must be able to understand the systems they buy, evaluate how they perform and act when something goes wrong.

Regulation and organisational competence belong together. Responsible AI adoption requires enforceable safeguards, organisations capable of applying them and evidence that the technology actually improves the task it is used for.

A policy cannot compensate for a team that lacks the skills or authority to question a system. Equally, responsible teams should be able to expect credible evidence and cooperation from suppliers.

Putting responsible AI into practice

Our experience at Coventry University brings this into focus. Through our AI Adoption Lab, researchers work alongside operational colleagues to address specific organisational problems. In our Early Student Support project, the initial question is set deliberately narrow: can available institutional data help staff identify students who may need support earlier than the current process achieves?

Earlier identification could give colleagues more time to help. But a technically impressive prediction has little value if it arrives too late or generates more alerts than staff can handle.

The first phase tests identification; it does not establish which interventions work or automate contact with students. Those questions require separate evaluation. Keeping that distinction clear is part of responsible adoption.

Before wider use, we should also ask who might be missed, who might be flagged unnecessarily and whether performance differs between student groups. Students should be able to correct inaccurate information and question decisions affecting them. Success ultimately means better support and demonstrating that requires evidence beyond a model’s accuracy score.

This might suggest a practical agenda for government and regulators: requirements for AI systems should be proportionate to the consequences of failure. An application influencing access to education or employment deserves closer scrutiny than one suggesting a meeting agenda.

For consequential uses, providers and deploying organisations should be required to produce evidence of suitability, document apps limitations and commitment to monitor performance after deployment. Changes to a system or its use should trigger renewed assessment where they could alter risk.

Clear accountability across the AI supply chain

Responsibility must also be clear across the supply chain. Organisations need enough information from suppliers to make informed decisions, including when systems change or when raw data ownership, control and access are affected.

Suppliers must answer for their products, while deploying organisations remain accountable for how they use them. Regulators need the technical expertise and resources to investigate failures and enforce obligations. People affected need an accessible route to human review and, where harm occurs, meaningful remedy.

Organisations can and should act now. Leaders should appoint accountable owners, and govern institutional AI so that adoption is always two-fold justified: via clear problem definition and expected improvement.

Operational staff must help design and evaluate the AI systems because they understand the work and will maintain it after specialist support ends. Procurement should examine evidence and limitations as carefully as functionality and price.

Building the right skills and human oversight

Training is essential to making those arrangements work. Asking someone to provide “human oversight” is meaningless unless they have the knowledge, time and authority to disagree with the system.

Staff need to recognise uncertainty, investigate errors and know when to stop using a tool. Leaders need to resource that work and listen to concerns. These capabilities must survive beyond an initial pilot or an enthusiastic project champion.

This also means giving people a voice in how AI is introduced. Staff and those affected should help identify what could go wrong and what acceptable performance looks like. Their experience can expose problems that technical testing misses, and help organisations decide whether an application should be used at all.

There is public responsibility too. Smaller organisations need practical guidance and access to expertise so that responsible adoption is achievable. Universities can contribute through applied research, workforce development and independent evaluation. This should be part of the national ambition for AI.

The lesson from a pint of milk is that trust depends on obligations supported by competent people and working institutions. We should expect the same seriousness when AI affects people’s opportunities and wellbeing. AI needs clear accountability – and organisations cannot wait for it.

Find out more about AI and Digital Technologies at Coventry University Group.