What happened
A new study from the University of Manchester, presented at the 10th Data for Policy Conference, identifies a structural disconnect in AI development. Based on in-depth interviews with engineers in technology, finance, and manufacturing, the research finds that while AI professionals can identify ethical risks such as unfair automated decisions and opaque systems, they often lack the organizational support to act on these concerns. The study attributes this to 'compliance theater,' where organizations prioritize documentation and speed over substantive ethical safeguards.
Researchers at the University of Manchester conducted a study involving in-depth interviews with AI and software engineers working across technology, finance, semiconductor manufacturing, and research organizations. The central finding is that these engineers are generally aware of ethical risks, including inaccurate outputs, unfair automated decisions, and the use of automated judgment in high-stakes areas.
Despite this awareness, the study reports that many engineers feel unable to implement necessary safeguards due to a lack of authority, incentives, and organizational support. The researchers describe this phenomenon as 'ethical awareness without ethical agency,' where the structural capacity to act on ethical knowledge is missing.
The study identifies specific organizational factors that limit this agency, including box-checking compliance processes, commercial and deadline pressures, and reward structures that prioritize speed over rigor. The authors label this environment 'compliance theater,' where organizations signal ethical commitment through documentation without consistently practicing it.
Alessia Vlasceanu, the lead researcher, stated that raising ethical concerns can carry professional costs for engineers, and that the work required for true ethics is not what is typically rewarded. The study was presented at the 10th Data for Policy Conference and published in the journal Data for Policy CIC—Zenodo.
Why it matters
This research shifts the focus of AI governance from individual engineer training to organizational culture and structural incentives. It suggests that current regulatory frameworks, such as the EU AI Act, may be insufficient if they only verify the existence of ethical documentation rather than the practical implementation of safeguards. By highlighting that ethical agency is constrained by professional risks and reward structures, the study provides a concrete basis for regulators to demand evidence of practical ethical enforcement rather than just policy compliance.
The study challenges the common assumption that AI ethics failures are primarily due to a lack of knowledge or care among engineers. Instead, it points to systemic issues within organizations that prevent ethical action, suggesting that training individual engineers is insufficient without changing the conditions under which AI is built.
This finding has direct implications for current AI governance efforts, including those under the EU AI Act. The research suggests that many regulatory measures focus on producing documents and reports that demonstrate ethical commitment, which may amount to a box-checking exercise if the underlying development practices do not change.
Professor Caroline Jay, who supervised the research, noted that the infrastructure of AI ethics has grown faster than the technology it governs, but is largely aimed at the wrong level. The study argues that meaningful change requires addressing how AI projects operate day-to-day, including who is responsible for addressing ethical concerns and whether safety work is recognized.
By identifying the gap between ethical documentation and practical implementation, the study provides a framework for regulators to look beyond paperwork and examine whether ethical safeguards are actually being followed in practice.
What to watch next
Regulatory bodies may adjust their oversight mechanisms to audit practical implementation of AI ethics rather than just reviewing policy documents. Organizations may begin restructuring internal reward systems to incentivize rigorous safety testing and ethical concern-raising. The researchers plan to conduct a larger-scale survey to validate these findings across a broader engineering population.
Regulators and policymakers may begin to develop new oversight methods that verify the practical implementation of AI ethics rather than just the existence of compliance documents.
Technology companies may face pressure to restructure their internal cultures and reward systems to support ethical agency, potentially leading to new roles or processes dedicated to ethical oversight.
The research team plans to test these findings through a larger-scale survey involving a broader engineering population, which could provide more quantitative data on the prevalence of these structural barriers.
Industry groups may respond by developing new standards or certifications that focus on practical ethical outcomes rather than just procedural compliance.