TL;DR
The Department for Work and Pensions has published an algorithmic transparency record for Work Hub, the AI assistant inside its Jobs and Careers Service. Dated 30 October, it says the tool currently uses Claude Sonnet 4.6 through Amazon Bedrock, has several thousand private-beta users, and is expected to reach a national audience over time. No human reviews each interaction as a matter of course.
The service is voluntary. Users talk to a chatbot about their goals and skills, and it suggests jobs, training and career paths, reworks CVs and draws up action plans. The record says it does not decide benefit entitlement, and that it is not meant for enforcement or compliance. We reported in August on an Open Data Institute evaluation of the same assistant.
What the record discloses
The record names both the supplier and the model, while noting that the configuration may change.
On oversight, DWP says outputs are advisory and so not checked individually. Instead it relies on evaluation and quality assurance, with people involved during testing. Users “remain responsible for acting on advice”.
On data, the record says sensitive details may come up, including health information that people volunteer. Guardrails are meant to strip personal identifiers before processing, and processing at AWS is described as transient, with nothing retained for model training. A data protection impact assessment and an equality assessment were both completed in May 2026.
On risks, it acknowledges that the AI “may produce incorrect or misleading outputs”, answered by testing, evaluation and disclaimers. For bias, DWP describes paired test profiles that vary demographic signals, to see whether equivalent customers get materially different recommendations.
Why it matters
Advice can change decisions even when it decides nothing. Someone who is out of work and told by a government-branded assistant which jobs suit them may reasonably give that weight, in our view, whatever the disclaimer says. The previous route, the record notes, relied mainly on work coaches, job boards and static guidance.
The appeals and review section reads “N/A”. That follows from the department’s position that nothing here is a decision about legal rights, but it does leave open what a user can do if advice is poor.
Looking Forward
National rollout is the point to watch. In our view, several thousand beta users is a small base for checking fairness across every group the service would reach. The most useful next disclosure would be DWP’s bias test results and any evaluation of whether users find work more easily. Other departments’ records will show whether this level of detail, including a named model, becomes normal practice.