Digital Privacy
AI in Asylum Decisions: Transparency and Accountability Failures
The Home Office is using two AI tools in asylum casework which could be unlawful. The Asylum Case Summarisation (ACS) tool uses GPT-4o to summarise asylum interview transcripts, and the Asylum Policy Search (APS) tool is a GPT-4o chat-based interface used to search country policy documents.
Both generate new text rather than simply retrieve existing information, meaning they can filter or omit facts relevant to a caseworker’s decision. APS was fully rolled out in July 2025; ACS followed in April 2026.
In response to a parliamentary question tabled in April 2026, the Home Office subsequently confirmed it did not follow the UK AI Playbook principles – even though the February 2025 Playbook preceded the full rollout of both. (The Home Office referenced the ‘project pilots’ in their explanation.)
In March 2026, the Open Rights Group had already published a legal opinion which concluded that the AI tools are likely unlawful.
Using these tools creates several problems, which include:
- No transparency. Applicants are not told AI is being used in their case. They cannot see or correct AI-generated summaries of their own interviews. Neither tool appears in the Algorithmic Transparency Recording Standard (ATRS) repository. The Home Office has also refused to disclose the ‘prompt’ used in ACS. See also this Home Office response.
- No published assessments or guidance. No Data Protection Impact Assessment (DPIA), no Equality Impact Assessment (EIA), and no Standard Operating Procedure (SOP) has been published for either tool. Without an EIA, the Home Office cannot demonstrate compliance with the Public Sector Equality Duty.
- Training is inadequate. The minister cited a generic ‘AI for all’ learning package, giving no detail on content, duration, or assessment. Case-specific procedural guidance is imperative.
- Oversight is internal and reactive. The only oversight mechanism is a feedback inbox set up after APS was already deployed. Caseworkers are not required to verify AI outputs against the original source material, so the quality of oversight depends entirely on each individual.
- ‘Human-in-the-loop’ is inadequate. Both tools generate the text a caseworker reads before making a decision, so they shape the information that is seen. Stipulating that a human signs off the final decision does not address this. Pilot evaluations found that 9% of ACS summaries were too flawed to use, and 23% of caseworkers lacked full confidence in the outputs.
We recommend that the Government:
- Suspends both tools immediately. The full DPIAs and EIAs must be published. Tools must be registered in the ATRS repository.
- Stipulates full transparency. The office of the Chief Inspector of Borders and Immigration should conduct an independent investigation. The Independent Advisory Group on Country Information should consider the use of AI for country of origin information summaries. Remedies for affected applicants must be implemented.
- Mandates civil society consultation before any future deployment of AI in the asylum system, with findings published and independently reviewed. Any future iteration must come with comprehensive caseworker guidance.
SOAS ICOP Policy Briefings to Inform Government and Parliamentary Debates. By Sara Alsherif (Open Rights Group), with Sanjana Deen (SOAS University of London) (02 June 2026)
For further information, contact the author at sara@openrightsgroup.org/ 07902823294. Contact Professor Alison Scott-Baumann for access to other experts at as150@soas.ac.uk, and visit our website for more information. The views expressed in SOAS ICOP Briefings are those of the authors and do not necessarily represent those of SOAS University of London.
