Social Care Files in 58 UK Councils Contain AI-Invented Reports of Suicidal Ideation
What happened
AI transcription tools deployed by local councils across England and Scotland are generating fabricated content in official social care records, including false reports of suicidal ideation that never occurred. An eight-month study by the Ada Lovelace Institute, published in February 2026, found that tools now in use across 58 UK councils routinely produce hallucinations when transcribing conversations between social workers and the vulnerable adults and children they are meant to protect.
The errors range from the absurd to the dangerous. One social worker reported that an AI tool inserted a note of suicidal ideation into a client's record after a conversation in which no such thing was said. Other cases produced nonsensical substitutions, with one tool rendering "parents fighting" as "fishfingers." Because these records inform decisions about child protection, care plans, and risk assessments, an uncorrected hallucination can follow a family through the system for years.
The Ada Lovelace Institute traced the problem to two converging failures. The first is technical: large language models predict likely text sequences rather than verify what was actually said, and in emotionally charged conversations they tend to fill silences and unclear audio with formal, clinical-sounding language that was never used. The second is organizational. Councils rolled out these tools with efficiency and cost savings as the primary metrics, some workers received as little as one hour of training, and review practices varied so widely that some practitioners spent only minutes checking AI-generated text before signing off on legal documents.
For the individuals recorded, the consequences are asymmetric and lasting. A false flag for suicidal ideation can trigger intrusive interventions, generate stigma within the system, and shape every subsequent assessment a family encounters. Equally, transcripts full of nonsense risk burying genuine concerns under noise that nobody fully processes. Social workers who relied on the tools without knowing they could invent content now face questions about their professional liability for documents they did not, in any meaningful sense, write.
What the study makes visible is a documentation gap that precedes accountability entirely. When a social care record is challenged, the current state of most deployments offers no way to establish what the AI system actually heard, what it substituted, and who reviewed what before signing. A provable record of what a system did, at what step and with what input, is the basic instrument that would let errors be traced, corrected, and attributed. Without it, a wrongly flagged client and a council that deployed a flawed tool occupy the same fog of uncertainty, and the person with the least institutional power is the one who bears the consequences.
Reported impact
- Affected parties
- Not publicly disclosed
- Harm type
- Not publicly disclosed
- Scale
- Not publicly disclosed
- Financial impact
- Not publicly disclosed
- Regulatory action
- Not publicly disclosed
Classification
Relevant governance controls
Governance control mapping is not available for this record.
- No controls mapped
Not publicly disclosed
Control mapping is analytical. It does not state that any control would have prevented the incident.
Sources and evidence
This record was researched and written by the Index. The event is also catalogued in the following database, which is listed for cross-reference.