Responsible AI depends on whose knowledge shapes the design

More than a failure to follow principles

Responsible AI is often described as a challenge of translating ethical principles into practice. In a November 2026 preprint, University of Bath researchers Thao Do-Ngoc and Emma Carmel propose a different emphasis: responsibility itself takes shape through development work, professional experience and institutional priorities.

Their argument is not that ethical commitments are necessarily insincere. Rather, people committed to addressing exploitation can disagree about what responsible action requires. An investigator may see linking records as essential to finding hidden abuse, while a support organisation may see keeping records separate as essential to protecting people.

The paper also connects ethics with questions about knowledge: what data can represent, which categories are defensible and what a system can legitimately infer. These choices may look technical, but they help determine whose circumstances become visible and which interventions appear justified.

What the researchers examined

The study draws on 14 semi-structured interviews conducted between October 2025 and February 2026, lasting 30–90 minutes. Participants included developers, researchers, law-enforcement practitioners, commercial actors and people working in civil society and survivor support.

Applications ranged from investigative tools and supply-chain risk assessment to worker screening, conversational support and privacy-preserving data sharing. Projects were at different stages, including prototypes, deployed systems and discontinued work.

The researchers deliberately sought varied settings rather than a representative sample. Their analysis compared participants’ reasoning about evidence, harm, competing priorities and development decisions, using institutional and professional contexts to interpret recurring patterns.

Anti-trafficking makes these questions particularly difficult. As the paper explains, records of identified or assisted victims do not measure everyone experiencing trafficking. Digital indicators can also be ambiguous: features treated as warning signs may occur in consensual sex work or ordinary migrant recruitment advertisements.

Different understandings of responsibility

The authors identify four overlapping orientations, which they call moral economies: shared understandings of valuable knowledge, legitimate action and good professional work.

One foregrounded exposure, agency and safety. Participants discussed whether collecting information could leave vulnerable users traceable, and whether developers understood enough about workers’ circumstances to design appropriate tools.

Another emphasised hidden exploitation, evidence and actionable investigations. Aggregating information could appear ethically necessary because missing exploitation was the central concern. Yet participants working with investigative applications differed: some questioned profiling, biased training data and the boundaries between trafficking, smuggling and immigration categories.

A third orientation translated exploitation into commercial risk, compliance and reputational exposure. The accounts raised a distinction between reducing a company’s risks and improving workers’ conditions; the two were not treated as automatically equivalent.

The fourth questioned the authority and adequacy of knowledge itself. Participants challenged what datasets actually measured, how context disappeared when people became data, and whether affected communities helped produce knowledge rather than merely becoming objects of analysis.

Four places where concerns can weaken

The paper organises its findings around four interconnected moments, not a fixed sequence.

Perception concerns whether a problem becomes recognisable. Professional routines can make some harms readily intelligible while leaving others difficult to articulate.

Judgement concerns how recognised problems are weighed. One participant described extensive data aggregation as justified by the seriousness of exploitation; another considered restrictions on identifying individuals justified by human rights.

Situated action concerns whether people can act on their judgements. A researcher developing a maritime forced-labour risk prototype wanted sustained community engagement but reported lacking resources, connections and a mandate. Other participants assigned ethical responsibility to downstream users or courts.

Socio-technical materialisation concerns what becomes embedded and sustained in systems and working arrangements. According to one participant, a support chatbot avoided accounts and IP collection, included a quick-exit button and reset sessions after two hours of inactivity. A worker-facing tool excluded photographs after its developers judged the potential harm too great. Elsewhere, commitments remained in undeployed models or unfinished prototypes, while staff departures threatened continuity.

Governance—and the limits of the evidence

The authors recommend durable arrangements that let technical specialists, domain practitioners and affected communities influence decisions together. Their proposed “epistemic infrastructure” includes sustained engagement, resources, decision-making authority and documentation of assumptions, trade-offs and rejected alternatives. Consultation alone is insufficient if the resulting knowledge cannot change development choices.

These are governance proposals drawn from qualitative analysis, not interventions tested for effectiveness. The study relies on participants’ interpretations and justifications rather than observation of development or deployment. It does not assess accuracy, downstream harms or benefits, and it includes no interviews with survivors, workers or other affected communities.

The small, heterogeneous sample supports theoretical development, not generalisation across anti-trafficking or AI practice. The preprint’s contribution is an account of where responsibility can take shape or weaken—and why ethical awareness alone may not give practitioners the conditions needed to act.