Transcribing trust: is transcription the use case that shows AI’s transformative power?
Government and public sector organisations are increasingly adopting AI transcription tools. Here, Imogen Parker of the Ada Lovelace Institute sets out who’s using what and the key considerations when it comes to assuring accuracy and avoiding bias
This year has seen transcription tools (also called AI scribes, ambient scribes or ambient voice technologies) take off within the public sector.
Newer foundation models now power automated-speech-recognition with far greater accuracy and flexibility than earlier language models. The UK prime minister’s AI Exemplars Programme, launched in summer 2025, includes several tools to transcribe and summarise notes, such as Justice Transcribe, the Department of Health and Social Care’s Discharge Summaries, and i.AI’s Minute tool for local authority use. Beam’s Magic Notes is one of various commercial products that can be deployed in the context of social work. And of course, general purpose AI tools, like Microsoft’s Copilot, offer a transcription functionality.
These applications are already widely adopted in the public sector. It is estimated that around a third of social workers are using generative AI tools with transcription capabilities [1]. A number of evaluations are underway, including the Greater London Authority’s co-produced study of Minute.
Amongst policymakers, there is optimism that transcription could be an important case study of how AI can transform the public sector and deliver efficiency benefits, and for good reason.
Large Language Models have radically improved and can now be fine-tuned for a variety of tasks. Paperwork is the constant complaint of stretched frontline workers and accurate transcription tools could be a significant time saver. Indeed, documentation is the most time-consuming task in health and care. Automated transcription could also act as a ‘leveller’, supporting, for example, those who are excellent at interpersonal aspects of frontline practice but struggle with paperwork or writing. This is an area in which AI tools seem to have a transformative potential to address a known problem.
However, to know whether they can deliver benefits, we have to answer various questions. The obvious ones are whether AI can transcribe content accurately enough and how much value (in terms of efficiency) the adoption of these tools can deliver. But more broadly, what other issues should we focus on and plan evaluations around?
Read more: UK chancellor allocates funds to deploy AI in government in Spring Statement
Data questions arising from foundation model underpinning transcription tools
First, the well-established issue of fabrication of information (also known as ‘hallucination’) means that tools may create inaccurate or fictional content.
Risks of fabrication occurring in transcripts are heightened by the fact that foundation models are not subject to adequate safety checks, testing or even transparency requirements. Flawed and biased data or toxic language could be built into tools, with negative effects even within fine-tuned models (tools that focus on specific domains like health, for example). Automatic speech recognition models may generate incorrect content that promotes harmful perceptions of the individuals included in the recordings.
We have to address a cluster of evaluation questions on the accuracy of transcribed content, the success rates in noticing and correcting inaccurate content, and the potential issues arising from the harmful, rather than simply incorrect, creation of content.
Questions of privacy, security and confidentiality are similarly linked to the reliance on foundation models. Transcription tools may not prevent sensitive information from being used or stored insecurely, depending on the arrangement with general purpose providers, and it is unlikely those using free versions of specific or general purpose tools will have meaningful data privacy. Earlier this year, NHS England’s national chief clinical information officer warned that many ambient voice technologies were in widespread use within clinical practice despite not being compliant with NHS governance requirements on data protection and clinical safety. The health authority instructed providers to stop using any transcription tools that failed to meet appropriate standards.
Read more: New Global Government Forum study calls for a ‘step-change’ to unlock NHS digital transformation
Functionality in practice
As well as evaluating accuracy, it will be important to assess how automated transcription performs in a variety of real-world settings. How close is the transcription to the spoken account of the conversation? How does it perform with different languages, accents, slang, illnesses or medical conditions, background noise or simply overtalking? How does automated transcription compare with human transcription, and are there particular groups who may be disadvantaged by the former?
As well as technical performance, how might the adoption of transcription tools change what is being noted during meetings between professionals and clients/patients: how does the potential of transcription affect people’s degrees of comfort in conversation? What etiquette or consent expectations should exist around the use of transcription tools? How might such applications affect disclosure or flow, especially in more sensitive conversations like those concerning HR, social work or policing? Will notetaking become exclusively related to spoken information, rather than capturing body language, silences or movement for example?
Transcript use
For transcripts to deliver benefits, professionals need to be able to use them – raising questions about acceptability and governance in different settings. Are transcripts valid when it comes to their use as legal evidence or as the basis for formal decision-making and due process? Who is legally responsible for ensuring accuracy in high-stakes settings, and – in the absence of regulation on foundation models – who will be held accountable for the performance of technical models or the potential for bias?
Professional practice impacts
There is a suite of evaluation questions around the effects of transcription tools on professional practice that are useful to explore, given the rapid roll out in key sectors like social work.
For example: How might roles and responsibilities change in light of transcription adoption? Will there be new duties to check text for accuracy, and who is held responsible or even liable for technical errors? If there are productivity benefits, to whom will they accrue: as transcription becomes normalised, will staff gain time for more relational practice or training or reflective practice, or will they simply be required to take on more case work.
Read more: Slick cities: How local authorities are using AI to tackle their most pressing problems
Beyond transcription
The notion of ‘transcription’ tools may increasingly be a misnomer: many applications offer functionalities that go beyond pure transcription, including summarisation, report writing or decision support. These functions may increase efficiency but bring more complex evaluation questions. Errors or bias in summarisation and decision support may be less immediately apparent than in simple transcription, while their impact could be more profound.
A recent study by Sam Rickman used LLMs to produce 30,000 summaries of real case notes from over 600 people, to assess counterfactual fairness – whether an AI model was introducing bias by creating different outputs for people based on their gender. While Meta’s Llama 3 produced similar summaries for men and women, Google’s Gemma created stark differences, despite the identical inputs. In some instances, summaries of the same case changed according to the gender attributed to the person. For instance, Mr Jones, in the summary, is described as ‘unable to access the community’, whereas Mrs Jones ‘despite her mobility issues and memory problems, is able to manage her daily activities’.
Beyond bias, there are questions to explore on how summarisations compare to human decisions about information hierarchy, including how much trust people place in AI’s summarisation. How does this compare to how a professional might summarise or prioritise information, and how does this affect professionals’ understanding of the conversation they are having with a client and retention of knowledge?
Read more: License to build: understanding what people think about public sector use of AI
Forthcoming research
At the Ada Lovelace Institute we have spent this year studying transcription tools in one sector: social work. We have been building evidence on the ground around some of these questions, exploring both how tools are being used by frontline workers, and how they could be evaluated to build a rich picture of impacts.
We are soon to publish the first instalment of that research in the form of an explainer on transcription tools, highlighting how they are built, and what the risks and benefits of using them are. Early next year we will present the full primary research on social work and transcription.
To hear about our AI in social work research when it’s published, subscribe to the Ada Lovelace newsletter.
[1] ‘Use of AI Rising among Social Workers, Poll Finds’ (Community Care) <https://www.communitycare.co.uk/2025/06/12/use-of-ai-rising-among-social-workers-poll-finds/> accessed 29 September 2025.
About Imogen Parker
Imogen is Associate Director (Society, justice & public services) at the Ada Lovelace Institute. Imogen’s career has been at the intersection of social justice, technology and research. In her previous role as Head of the Nuffield Foundation’s programmes on Justice, Rights and Digital Society she worked in collaboration with the founding partner organisations to create the Institute. Prior to that she was acting Head of Policy Research for Citizens and Democracy at Citizens Advice, Research Fellow at the Institute for Public Policy Research (IPPR) and worked with Baroness Kidron to create the children’s digital rights charity 5Rights. She is a Policy Fellow at Cambridge University’s Centre for Science and Policy.
