Bath and North East Somerset Council has deployed Microsoft Copilot to read and summarize over 5,500 public comments on the proposed Bath Rugby stadium, embedding the AI-generated analysis directly into a 121-page planning report. The move, revealed in committee papers ahead of a crucial vote, immediately sparked a fierce debate over transparency, democratic legitimacy, and the proper role of artificial intelligence in local government decisions.

The council’s planning officers used the enterprise AI assistant to sift through thousands of submissions—overwhelmingly in favor of the 18,000-seat permanent stadium but including hundreds of objections—and then manually refined the thematic summaries. While the full text of all comments remains publicly accessible on the council’s planning portal, the reliance on AI for the official précis that will guide councillors’ understanding of public sentiment has drawn scrutiny from campaigners, data protection advocates, and governance experts.

The Stadium Proposal and Public Response

Bath Rugby’s push for a year-round stadium at the Recreation Ground—a site gifted to the city in 1956 and set within a UNESCO World Heritage landscape—has polarized the community. The club argues the project will generate economic benefits, while opponents, including filmmaker Ken Loach, decry it as a “travesty” that threatens the historic riverside setting.

During the latest consultation, the council received 5,590 representations. According to the officer’s report, 5,086 supported the scheme, 368 objected, and 136 were uncategorised. Given the volume, the planning department turned to Microsoft Copilot to help them process the influx.

How the Council Used Microsoft Copilot

The officer’s report states: “Given the very large number of representations submitted via the council’s online comments form, these have been reviewed and summarised by Microsoft Copilot. This is an artificial intelligence tool that was instructed to identify reasons for objection/support. The programme reviewed all comments received since the application was submitted.”

The AI generated topic headings, which the case officer then refined by sampling and checking a subset of comments. Representations sent directly to the case officer were read individually and also fed through Copilot for structured summarization. The final report listed themes such as heritage impact, economic benefits, traffic, and flood risk—all derived from the AI-assisted review.

Efficiency Gains vs. Democratic Accountability

Proponents argue that using AI for routine bulk analysis is a pragmatic response to chronic under-resourcing in local planning departments. Other UK councils that have piloted Copilot report measurable productivity improvements, allowing officers to focus on complex statutory assessments rather than administrative triage. In Bath’s case, manually processing thousands of comments would have taken weeks; the AI compressed this into a fraction of the time.

However, critics point out that consultation summaries are not neutral clerical tasks—they frame the issues that will shape a decision. If an algorithm weights themes by word frequency alone, a flood of templated support emails could drown out a lone ecologist’s detailed objection about a protected species. The fairness of the process hinges on whether the summary faithfully captures the material content of all representations, not just the loudest.

The Risks: Hallucination, Bias, and Opacity

Large language models like those powering Copilot are prone to hallucination—generating plausible but incorrect outputs. A mis-summarized comment could omit a crucial technical objection or incorrectly categorize a submission as supportive. Without a clear audit trail, neither councillors nor the public can verify the AI’s accuracy.

Frequency bias is another concern. Mass template responses, whether in support or opposition, can dominate automated thematic analysis, while sparse but substantive expert evidence may be sidelined. In this case, statutory consultees such as the Environment Agency and Historic England submitted separate detailed responses, but the AI’s thematic roll-up of public comments might have blended these with general sentiment unless explicitly weighted.

Opacity compounds these issues. The council has not proactively published the specific prompts given to Copilot, the unedited AI output, or a record of which themes were modified by human reviewers. While the original comments are available online, the intermediary step—the AI’s interpolation—remains a black box. This lack of traceability undermines public trust and exposes the council to legal challenge.

Public representations often contain names, addresses, and other personal data. Feeding such information into a cloud-based AI raises immediate data protection questions. The council would need to conduct a Data Protection Impact Assessment (DPIA) and ensure its contract with Microsoft prohibits the use of input data for model training, mandates data residency, and provides for deletion. Several councils that have deployed Copilot emphasize DPIAs and governance boards as essential safeguards—but it is unclear whether Bath and North East Somerset has done so publicly.

Moreover, the Secretary of State for Housing, Communities, and Local Government has already signaled the stadium decision’s sensitivity by instructing the council not to grant permission without “specific authorisation.” If the application is refused or approved and then challenged in court, the AI-assisted summary could become a central exhibit. An appellant might argue that the council failed to properly consider material planning considerations because an opaque algorithm distorted the public’s input.

Best Practices for Municipal AI

The Bath case underscores the urgent need for clear guidelines when public bodies deploy AI in democratic processes. Experts recommend a minimum set of safeguards:

  • Human-in-the-loop sign-off: Every AI-derived summary that informs decision-makers must be reviewed, annotated, and certified by a named officer who confirms that all material points were captured correctly.
  • Publish the audit trail: Release the prompts, raw AI output, and final edited summaries as an appendix to committee papers, along with a clear explanation of sampling methodology.
  • DPIA and procurement protections: Conduct and disclose a specific DPIA; ensure contracts explicitly prohibit third-party model training on council data, define retention periods, and lock data residency.
  • Weighting and flagging rules: Present technical submissions from statutory consultees and experts separately from general public opinion, so they are not lost in frequency counts.
  • Representative sampling at committee: During the meeting, require officers to read out verbatim original comments that exemplify each major theme, rather than relying solely on tables of AI-generated counts.
  • Proactive public communication: State clearly in consultation materials if AI will be used and explain the safeguards in plain English.

These steps do not ban AI; they insist on traceable human stewardship. Councils that have adopted such frameworks report that they preserve the efficiency benefits while maintaining legitimacy.

The stadium proposal sits at the intersection of heritage preservation, urban development, and local identity. ICOMOS-UK and the Environment Agency have raised concerns, and the ministerial holding directive places the final decision in a politically charged limbo. If the planning committee grants approval without robust documentation of how public representations were processed, the decision could be vulnerable to judicial review.

Cllr. Paul Crossley, a Bath councillor with a technology background, told local media that he was “comfortable” with the AI use because “on this occasion, I can’t see anything that is likely to cause a problem.” But many campaigners demand more than personal reassurance; they want a verifiable system that withstands legal and public scrutiny.

What Happens Next

The planning committee meeting on 17 September may not be the final word. If the committee defers or refuses, the developer could appeal. If it approves, the Secretary of State may call in the application. Either path will subject the AI-assisted process to extraordinary review.

For now, Bath’s experiment serves as a live stress test for municipal AI governance. The technology offers undeniable efficiency gains, but those gains cannot come at the expense of due process. As councils across the UK explore similar tools, the Bath case provides a crucial lesson: AI can scale human judgment, but it must never replace it.

Citizens, journalists, and councillors should demand the full AI audit trail—prompts, raw outputs, and human edits—and verify that officer oversight was more than cosmetic. The future of democratic planning in the age of AI depends on getting this balance right.