Making the most of Ada Lovelace Institute’s research into use of AI transcription tools in the social work sphere
1. What the Institute’s report “Scribe and Prejudice” explores
The Ada Lovelace Institute looked at how AI transcription tools are being used in real social work – not in theory, but across 17 local authorities in England and Scotland, interviewing 39 social workers and managers from spring to autumn 2025. They treat AI transcription as a socio-technical change: it is not just a gadget to take notes faster, it shifts workload, risk, accountability and, potentially, people’s experience of care. The focus is adult and children’s social care, but many of the conclusions obviously bleed into other frontline services.
2. Core findings in plain language
The findings fall into five big themes which, together, are the “spine” of the report:
- Why councils are so keen
- Severe and long‑running resource pressures, high turnover and backlogs make anything that promises “more work with fewer staff hours” extremely attractive.
- Many councils are in “test and learn” mode: short pilots, sometimes compressed procurement, and a willingness to try things fast even when they know their evaluation capacity is limited.
- A subset of better‑resourced councils have ethics/evaluation teams and multiple pilots; others move quickly from a small pilot to live use with relatively light oversight.
- What gets measured (and what doesn’t)
- Evaluation is overwhelmingly about efficiency: hours saved, assessments completed, number of visits per worker, waiting lists.
- Some authorities add softer measures (self‑reported wellbeing, record quality) but these are secondary and rarely connected to outcomes for the people drawing on care.
- There is almost no systematic evidence on how AI transcription affects:
- How social workers actually use the tools
- Use is highly varied and pragmatic: some use full record‑and‑summarise; others only record and then write their own reports; others never record at all but dictate or type notes into the tool later.
- Many prefer to use AI for meetings with other professionals rather than with people drawing on care, given concerns about consent, trust, mental health and previous experiences of being recorded by police.
- There are genuine benefits: some workers report major time savings, improved work‑life balance, more time for direct work and better‑captured detail in records; neurodivergent or disabled staff can find the tools transformative for access.
- Those benefits are not evenly realised: local documentation formats, expectations about length/detail, training gaps and managerial demands can turn any time saved into more work rather than less.
- Risk, accuracy and “human in the loop”
- All parties nominally agree that social workers remain accountable; tools are designed so nothing moves forward until a worker has “signed off” the AI output.
- In practice, oversight is highly variable: some staff spend an hour carefully checking; others skim in minutes and paste into the case management system under extreme time pressure.
- The report records concrete problems, not just hypothetical ones:
- clear hallucinations (for example, invented suicidal ideation),
- mis‑transcribed content (especially with accent and dialect),
- irrelevant material (pets, jokes) entering children’s files,
- AI‑generated language that is more formal, educated clinical and less person‑centred than current best practice and thus instantly recognisable as generated.
- There is also a split in risk perception: some practitioners and managers see “loads and loads of risks”; others treat these tools as “light‑touch” AI where the main risk is forgetting to press record.
- No settled view on “where it’s safe”
- Some authorities draw hard lines: no AI transcription for statutory assessments, capacity decisions or safeguarding; only for lower‑stakes work or for multi‑professional meetings.
- Others permit or even encourage use in statutory processes as long as the worker reviews the outputs and the person has consented.
- Cross‑agency contexts (e.g. child protection conferences with police and health) are particularly messy: some police forces prohibit AI recording; others do not, so practice varies even within similar meetings.
- In many places, frontline practitioners are de facto making ethical and legal judgements on when to use these tools, without consistent guidance or shared thresholds for “unacceptable risk”.
3. How the report reframes the “efficiency story”
The report doesn’t deny the efficiency gains – it accepts that many social workers experience meaningful time savings and some backlogs do move. But it pushes three reframes that matter for policy and practice:
- Efficiency is not the same as productivity, and neither is the same as public benefit. Studies in other departments (e.g. Copilot pilots) show time saved on tasks does not automatically translate into better productivity; Lovelace argue that, in social care, you have to ask what replaced that time – more cases, more relationship‑based practice, or simply higher expectations?
- AI‑created records are not neutral “admin”. The risks are not just typos; they include distorted narratives, biased emphasis and language that can affect thresholds, court decisions and people’s ability to understand and challenge what is written about them.
- “Human in the loop” is not a magic fix. It is work. It absorbs scarce capacity. Its quality varies. And if social workers are already overloaded, the temptation to trust or minimally skim AI outputs is entirely predictable. The report is sceptical of treating “there’s a person somewhere in the process” as sufficient assurance. It needs to be a competent expert.
4. Key tensions the report surfaces
Seen as a whole, the report is less about a single technology and more about four structural tensions:
- Speed vs scrutiny
The same austerity‑driven pressures that make AI transcription attractive also undercut councils’ ability to evaluate and govern it properly. Faster pilots and compressed procurement increase the risk that vendor‑led narratives and narrow metrics dominate. - Individual responsibility vs systemic support
Social workers are being cast as the primary safety mechanism (spotting hallucinations, gauging suitability, managing consent) at the same time as they are exhausted and under‑trained in AI. The accountability load is drifting downwards without matched support, guidance or shared standards. - Innovation vs inequality
Performance differences across accents, languages and communication styles, plus patchy governance, risk embedding or worsening existing inequalities – both between areas (better‑resourced authorities vs others) and between groups of people drawing on care. - Promise vs evidence
The report is explicit that the evidence base is thin: no robust causal link yet between AI transcription and better outcomes; very little on long‑term, system‑level impacts on the profession or on people’s lives; and almost nothing that could legitimately justify broad claims about productivity savings across the whole sector.
5. What Ada Lovelace think needs to happen
- Central government infrastructure
- Extend the Algorithmic Transparency Recording Standard to local government, and make AI‑generated care records clearly identifiable (e.g. “watermarks”).
- Fund and coordinate more, and more diverse, pilots, rather than leaving each authority to reinvent evaluation in isolation.
- Create a “What Works Centre for AI in Public Services” to collate evidence, run or commission independent evaluations and establish shared evaluation standards.
- Generate context‑specific evidence on AI and productivity, instead of extrapolating from a small number of pilots to the whole public sector.
- Support longitudinal and qualitative work on systemic impacts: effects on decision‑making, professional skills, workforce, and the experiences of people drawing on care, including potential bias and discrimination.
- Involve people with lived experience and unions meaningfully in this research and in deliberative processes around acceptable uses and red lines.
- Regulators and local authorities
- Develop guidance on acceptable use in statutory processes and formal proceedings, with clear accountability structures that extend beyond the individual worker.
- Train social workers and managers on the specific risks of generative AI in recording (hallucinations, bias, language, consent), not just generic “digital skills”.
- Train social workers and managers on the specific risks of generative AI in recording (hallucinations, bias, language, consent), not just generic “digital skills”.
6. CASCAIDr CIC’s conclusion
- AI transcription tools are already changing social work – often in ways social workers welcome – by easing some of the worst documentation burdens and, in some cases, improving the quality and richness of records.
- Those same tools are being introduced into a system under huge strain, with patchy evaluation and governance, and there is now clear evidence of hallucinations, misrepresentation and uneven benefits that could have real consequences for people and for the profession if not carefully managed.
- We possess a decade of analytical materials that could be added to such tools, as a first source of legal literacy, with humans in the loop – meeting many of the concerns and addressing the report’s Insights. However, if the transcription standard is inadequate, that will do no good at all for people’s legal rights and social workers’ professionalism or sanity.
Scribe and Prejudice can be found at https://www.adalovelaceinstitute.org/report/scribe-and-prejudice/