91Ö±²¥

AI engineers often recognise ethical risks associated with the systems they build, but many lack the authority, incentives and organisational support needed to act on them, according to new research from The University of Manchester.

Based on in-depth interviews with AI and software engineers working across technology, finance, semiconductor manufacturing and research organisations, the study found that engineers could readily identify issues such as inaccurate outputs, unfair automated decisions, AI systems that are difficult to explain, and the use of automated judgement in areas that can significantly affect people's lives.

However, many also described safeguards they would like to implement but felt unable to put into practice. What they lacked was not awareness. It was the structural capacity to act on it.

The researchers describe this as "ethical awareness without ethical agency": engineers who know the right thing to do but are unable to act on it.

Presented at , the research raises questions about whether current oversight frameworks measure genuine ethical practice or merely the documentation surrounding it.

The study uncovered a series of organisational factors limiting engineers’ ability to act, including tick-box compliance processes, commercial and deadline pressures and reward structures that prioritise speed over rigour. Together, these create what the paper calls "compliance theatre": organisations that signal ethical commitment without consistently putting it into practice.

At a time when governments and organisations are introducing new AI rules and standards, including under the EU AI Act, the study suggests that many current efforts to govern AI focus on producing documents, policies and reports that demonstrate ethical commitment.

However, the findings indicate that unless organisations also change how AI is developed in practice, these measures may amount to little more than a box-ticking exercise.

Professor Caroline Jay, who supervised the research, said: "This is not a story about bad companies or bad engineers. It is a story about the difference between the appearance of ethical practice and the substance of it. The infrastructure of AI ethics has grown faster than the technology it was designed to govern. What this research shows is that the infrastructure is largely aimed at the wrong level."

The study argues that meaningful change must address the way AI projects operate day to day, including how ethical concerns are raised, who is responsible for addressing them and whether careful testing and safety work are recognised and rewarded.

It also calls on regulators to look beyond whether organisations have produced the correct documentation and examine whether ethical safeguards are being followed in practice.

The research forms part of a wider doctoral project examining the relationship between AI engineers’ intentions, public perceptions of AI and the outcomes produced by AI systems.

The researchers now plan to test these findings through a larger-scale survey involving a broader engineering population. They hope the work will help inform future AI governance approaches by focusing greater attention on the organisational conditions that shape how AI systems are developed and deployed.