Social workers liable for AI use amid scarce training, report warns

Social workers are being held fully accountable for their use of AI transcription tools, despite receiving inconsistent training and guidance and facing significant risks of inaccuracy, a new report has found.

The study, published by the Ada Lovelace Institute, found that practitioners recognised clear benefits from using AI-powered transcription and summarisation tools. These included reduced administrative workloads, more time for direct relationship-based practice, improved wellbeing and better work-life balance.

However, the report warned of a “great risk” that inaccurate or fabricated information could be entered into care records, amid inconsistent safeguards and a lack of agreed standards on appropriate use.

Rapid rollout amid resource pressures

The research, based on 39 interviews with practitioners, senior leaders and technology providers across England and Scotland, found that adoption of transcription tools had been “rapid”. In some cases, councils implemented them following only limited pilot testing, motivated largely by the promise of efficiency savings in a context of acute resource pressures.

Most practitioners interviewed had used Beam’s Magic Notes, a tool designed specifically for social care, or Microsoft’s general-purpose Copilot. The report emphasised that not all AI transcription tools are equivalent, noting that sector-specific products may include tailored safety features and language capabilities.

Council evaluations primarily focused on efficiency metrics, such as time saved on case recording or increased visit capacity. There was far less evidence of systematic assessment of impacts on the quality of records, practitioner wellbeing, or—critically—on people receiving support.

Reported benefits

Practitioners described saving from several hours per week to, in some cases, halving their administrative workload. This enabled:

• More time spent in direct work

• More natural, less disrupted conversations during visits

• Improved ability to observe non-verbal cues

• More comprehensive documentation

Some participants also highlighted accessibility benefits, particularly for neurodivergent or disabled practitioners, who found transcription tools reduced cognitive load and administrative stress.

These findings echo a separate review commissioned by Social Work England examining AI’s impact on the profession.

Frequent inaccuracies and “hallucinations”

Despite these advantages, practitioners reported recurring problems with:

• Fabricated or “hallucinated” content

• Incorrect summaries, including serious misstatements (for example, indicating suicidal ideation where none had been expressed)

• Misspelled names

• Difficulty handling regional accents

• Overly formal or impersonal language

Such errors carry significant safeguarding and legal risks, particularly in statutory contexts. The report cautioned that relying solely on individual practitioners to detect inaccuracies may be insufficient to ensure faithful case records.

Inconsistent oversight and accountability

Training provision varied significantly between councils. Some provided structured webinars, e-learning modules or vendor-supported pilots; others offered minimal or no specific guidance. As a result, practitioners’ understanding of risks and appropriate oversight differed considerably.

All practitioners interviewed stated that reviewing AI outputs was essential, but the time spent checking transcripts ranged from a few minutes to up to an hour per case. There are currently no agreed professional standards governing verification processes.

Approaches to prompting—used to shape or refine AI-generated summaries—also varied. Some practitioners actively used prompts to improve person-centred language or remove repetition; others were unaware of or unconfident in using such functionality.

Policies on appropriate use were equally inconsistent. Some local authorities prohibited transcription tools in statutory assessments, such as Care Act processes, while others left decisions entirely to individual practitioners.

Despite these inconsistencies, social workers reported assuming full professional responsibility for AI-generated outputs. Most tools require practitioners to formally sign off on documentation, and managers reinforced this accountability.

The report questions whether this model is proportionate, given the variability in training, absence of sector-wide standards, and inherent technological risk.

Key recommendations

The Ada Lovelace Institute called for:

• National guidance from regulators and sector bodies on AI transcription use in statutory processes, with clear accountability frameworks informed by people with lived experience

• Mandatory council reporting on AI tool usage through the Algorithmic Transparency Recording Standard

• Government-funded, multi-site pilots using diverse evaluation methodologies

• Establishment of an evidence centre for AI use in public services

• Greater transparency when care records have been produced using AI

In response, Beam co-founder Seb Barker argued that sector-specific tools are essential and stated that generic AI systems are not adequately designed for the complexities of frontline social work.

Overall, the report concludes that while AI transcription tools offer tangible operational benefits, their rapid deployment—combined with inconsistent governance and persistent inaccuracy risks—creates structural accountability challenges for practitioners who remain professionally liable for outputs they did not fully author.

Disclaimer: The accuracy and information of news stories published on this website is accurate on the date of publishing. We endeavour to update stories if information change. You can contact us with change and update requests. Where possible, we will link to sources. Content on this website is for guidance purposes only. We cannot accept any responsibility or liability whatsoever for any action taken, or not taken. You should seek the appropriate legal advice having regard to your own particular circumstances.