Search

4 results for “sarahalsherif”

1. 

   Sara Alsherif سارة الشريف @sarahalsherif

   2026-10-12

   Neither the ACS nor the APS tools are referenced in the Algorithmic Transparency Recording Standard Hub.

   There is no quality assurance process in place for the AI-generated summaries from ACS.

   Similarly, the APS tool's generated text is not subject to quality checks.

   Only 0.3% of all summaries produced between April and July were reviewed by Home Office Subject Matter Experts, with error rates between 15% and 23%!

   There are no plans to conduct an independent audit of the system.

   And there has been no effort to align the use of these tools with the UK Government AI Playbook.

2. 

   Sara Alsherif سارة الشريف @sarahalsherif

   2026-10-12

   In two accounts the Home Office stated that asylum applicounts “are not specifically told AI has been used in the consideration of their claim” and cannot see or correct AI summaries of their interviews.

   The DPIA clearly states that applicants are informed of data processing via the Borders, Immigration and Citizenship privacy information notice on GOV.UK. However, this notice, updated on August 6, 2026, does not mention AI tools.

   The DPIA for the APS tool that supplied to ORG through FOIA as the APS DPIA describes ACS. Its
   internal title is ‘Interview Summarisation Tool – Pilot’ and every substantive
   section concerns the summarising of interview transcripts. It is not apparent from the disclosed material that a data protection assessment specific to APS has been carried out.

3. 

   Sara Alsherif سارة الشريف @sarahalsherif

   2026-10-12

   Out of 70,000 summary texts generated by the summarisation tool, only 0.3% were examined for accuracy. The error rates observed in these summaries exceeded the 9% figure reported in the Home Office's published pilot evaluation.

   Of the 203 summaries reviewed, 139 were deemed acceptable, 19 required further review, and 45 failed (constituting 22%). In the second stage of review, 30 of 56 summaries failed. Overall, between 15% and 23% of sampled summaries were assessed as unsuitable for use.

4. 

   Sara Alsherif سارة الشريف @sarahalsherif

   2026-10-12

   ERRORS, INACCURACIES AND HOME OFFICE INCOMPETENCE
   Check the findings from our latest report @openrightsgroup
   All the resources are now publicly accessible.
   Review them and verify the findings.

   We conducted a six-month investigation to understand the actual use of AI tools in the asylum decision-making process.

   Each statement from the Home Office was found to be contradicted by alternative accounts obtained through Freedom of Information Act (FOIA) requests. It took eleven Written Parliamentary Questions (WPQs), two pieces of ministerial correspondence, and ten FOIA requests to establish that these tools are deployed without adequate internal oversight, accountability, or transparency.

   https://www.openrightsgroup.org/publications/the-use-of-chatgpt-in-asylum-decisions/