Mark Wong, Zoe Bartliff, Patrizia Di Campli San Vito, Yunhyong Kim, Emily O'Hara, Kate Simpson, Cagri Yildirim, Simone Stumpf · ENLIGHTEN (Jurnal Bimbingan dan Konseling Islam) 2026 · 2026
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
AI is widely expected to transform government and public services by improving productivity and cutting costs: from health, social care to welfare eligibility assessments. However, evidence of achieving these benefits in real-world contexts is limited, whereas evidence of their harms and exacerbation of systemic inequalities is growing (Nature Medicine, 2026; Raji et al. 2020). Key to this gap is the lack of public accountability and participation. In the UK, the public feels increasingly disenfranchised from government decision-making about AI use (Modhvadia et al., 2025). While more than half of the respondents of a Responsible AI UK survey (N=2,055) believe AI risks outweigh their benefits, showing a growing 'AI pushback' (Devlin et al., 2026). The need to strengthen accountability and AI assurance, especially auditing, is highlighted in international policy debates, e.g., the EU AI Act and the UN Global Dialogue on AI Governance. But there is a lack of consensus on how audits should be conducted, what good practice looks like, and what role end-users and people impacted by uses of AI should play. The evidence and tools for AI auditing are growing, but there is a critical lack of accountability, what Birhane et al. (2024) call the 'broken bus' of audit practice. Recent research also shows that assessments of sociotechnical aspects are typically done late in the AI development lifecycle. Our presentation critiques this AI audit process in the public sector: who is in control, who should audit, why should they audit, and whose expertise has power and legitimacy? Extant research and tools are heavily focused on technical instead of socio-technical audits, which means current auditing approaches rely on evaluating technical performance, optimising AI models, and supporting AI developers, researchers, and corporations. Very few AI audit practices redress the exclusion of end-users and impacted communities. This exclusion is not only in tension with the public sector's ethos but also fails to ensure public benefits and interests are safeguarded. Ojewale et al.'s (2025) review of 400+ auditing tools highlight that most tools focus on technical 'standards identification' and 'performance analysis', while very few address the 'harm discovery' and 'audit communication and advocacy' stages of auditing, and overall lack meaningful support for active participation. Increasing accountability by putting previously under-represented people at the centre of auditing and make AI adoption more ethical and trustworthy in the public sector is vital. We present a new framework ('methodology') developed by our project, Participatory Harm Auditing Workbenches and Methodology (PHAWM), which tackles this urgent gap through participatory AI auditing. This fulfils a key requirement of a recent UN AI Governance for Humanity Lab's (2026) white paper: "Member States and the international community could co-fund and institutionalise standardised participatory harm auditing methodologies...These resources provide the tools and training necessary to empower citizens without technical AI backgrounds to formally audit systems for societal harms." (ibid, p. 36) Our multi-year research project provides a "standardised participatory harm auditing methodolog[y]", bedded in an interdisciplinary approach by combining computer science, social policy, digital humanities, and human-computer interaction expertise, and supported by a series of co-design stakeholder workshops in the UK. We will discuss results from our subsequent field trials with two local government authorities in England and Scotland, who conducted participatory audits using our audit methodology and the tool it supports over four months in 2026. The local authority in England involved 11 young people to audit the impact of an AI chatbot developed internally by the local authority, which aimed to improve adult social care information and service access. In Scotland, three practitioners were involved in the participatory audit of a commercial AI scribe / ambient voice technology. In this presentation we argue that participatory audits help inform responsible AI adoption and diversify knowledge and perspectives represented in auditing in local governments. Our key findings include: -The same methodology supported audits in two very different settings. In England young people audited a chatbot they might use themselves; in Scotland senior practitioners audited a clinical tool they already used daily. Neither group had audited an AI application before, and both were able to produce audits that covered the key participatory auditing steps of impact discovery, standards identification, and performance analysis. This suggests portability across diverse groups of auditors and audit settings, a key design requirement of a successful participatory auditing methodology. -Auditors identified harms the research team had not anticipated. The young auditors added audit criteria that included evaluating the environmental cost of the underlying model, testing the chatbot's use at a highly local level of service provision, and finding signposting failed once services had closed. They felt successful as auditors and located their legitimacy in their own service knowledge and lived experience. This also illustrates how non-AI-expert auditors can uncover harms that may be missed in a traditional audit. -Disagreement between auditors was informative rather than noise. The English cohort reached different verdicts on the same chatbot. The divergence was traceable to which criteria each group of young people chose to test, and it records the range of judgements a service might meet from the people it serves rather than a single expert conclusion. -The barriers were often organisational, not methodological. Organisational AI policy prevented some intended auditors from participating, an initial pool of around thirty yielded three for the auditors in Scotland, and information governance and procurement standards were raised as preconditions for deployment. This evidences the kind of conditions and organisational cultures a participatory audit has to account for rather than assume. In conclusion, this presentation will demonstrate how our methodology and its underlying principles support participatory AI auditing in local government and the public sector more widely. Our findings will add new understandings to how people with lived experience expertise and policy domain expertise can be empowered and can identify blind spots overlooked by current auditing approaches. We argue participatory auditing is key to create more equitable and responsible public sector AI governance, and we recommend that public sector policies, regulations, and assurance need to include participatory auditing methodology as an essential complement to technical AI auditing.
No comments yet — start the discussion below.