Bath Rugby stadium: AI summarises 5,500 public comments for council

Bath and North East Somerset Council used Microsoft Copilot to read and summarise thousands of public comments on Bath Rugby’s proposed 18,000‑seat stadium at the Recreation Ground, folding an AI‑assisted thematic analysis into the officer’s 121‑page planning report that went to councillors ahead of a high‑stakes committee meeting — a step that illuminates both practical gains and serious democratic, legal and data‑governance questions. (feeds.bbci.co.uk)

---

Background​

The Recreation Ground (the Rec) lies in the heart of Bath’s UNESCO World Heritage setting and has been the club’s home for more than a century. Bath Rugby has submitted revised plans for a permanent, year‑round 18,000‑seat stadium, a proposal that has generated thousands of public representations and drawn scrutiny from conservation bodies, local politicians and national ministers. The council received roughly 5,590 representations in the most recent consultation period; reporting consistently puts the bulk of those comments in support of the development (around 5,000+ in favour), with hundreds of objections and a smaller number of uncategorised entries. (bathecho.co.uk)
The planning officer’s committee report — a detailed, 121‑page document prepared for councillors — states that the sizable set of public comments submitted via the council’s online form were reviewed and summarised using Microsoft Copilot, with the planning officer subsequently sampling and refining the AI’s outputs. Representations submitted directly to the case officer were likewise described as read individually and summarised with Copilot. That procedural detail is the single most consequential AI‑related fact in the public record for this application.
The application was scheduled for a planning committee meeting on 17 September. The national government has written to the council directing it not to approve the plan without "specific authorisation" from the Secretary of State, signalling that ministers are reserving the right to call the decision in for central determination. The intervention underscores how politically and legally sensitive the matter is. (feeds.bbci.co.uk)

---

What the council actually did — the facts as reported​

• The council’s planning officer prepared a 121‑page committee report that described how the council handled the public representations on the stadium application.
• That report states that the online comments (more than 5,500 entries) were reviewed and summarised using Microsoft Copilot; the officer then refined AI‑generated topic headings based on sample checks. Representations sent by other channels were read individually and also summarised using Copilot.
• The headline tallies reported in council papers and local reporting were approximately 5,590 total representations, ~5,086 in support, ~368 objecting, and ~136 uncategorised. Those raw comments remain available on the council’s public planning portal for independent scrutiny. (bathecho.co.uk)

These three points are the operational core of what happened: an enterprise AI assistant was used to process a very large set of public comments, and those summaries were incorporated into the officer’s committee report for councillors to consider.

---

Why this matters: public consultation, evidence and legitimacy​

Public consultation performs a specific legal and democratic function in planning: it identifies material planning considerations (heritage, highways, flooding, ecology, noise, amenity, etc.) that decision‑makers must weigh. The mechanics of how representations are summarised and presented matters because committee reports frame councillors’ understanding of the issues and the perceived balance of local opinion.

• Scale and capacity pressures. Planning teams are chronically under‑resourced. Processing 5,000+ individual comments by hand is labour‑intensive; automated theming can deliver time savings and ensure officers can focus on technical and statutory issues. Evidence from other councils deploying Copilot shows measurable productivity gains in drafting, minutes and routine summaries. (microsoft.com)
• Framing and emphasis. Summaries shape narratives. If an AI tags and groups information primarily by word frequency, repeated short template responses can dominate the theme extraction while sparse but technically significant submissions (e.g., an ecologist pointing to a previously unrecorded protected species) may be down‑weighted by the algorithm unless flagged. That raises a fundamental fairness question: does an AI summarisation preserve the material content of public representations?
• Transparency and traceability. Councillors — and the public — should be able to trace summary headings back to representative original comments. Without an auditable trail showing the prompts used, the AI output and the officer’s edits, trust can be eroded. The planning portal still hosts the full texts of representations, but committee papers ideally should include a clear, accessible audit trail describing how automated theming was performed and verified.

In short: using AI to process public comments is not intrinsically improper, but the legitimacy of the planning process depends on governance, documentation and human verification.

---

The technology: what is Microsoft Copilot and how are councils using it?​

Microsoft Copilot is an AI assistant embedded into Microsoft 365 and other Microsoft services. In enterprise deployments it can be provisioned under organisational controls, and councils across the UK have trialled Copilot for internal tasks—drafting, meeting summaries, simple analyses and thematic grouping of consultation responses. Local authority pilots (for example Somerset Council and others) have reported time savings and productivity improvements when Copilot is used with governance and DPIAs in place. (local.gov.uk)
Typical municipal uses include:

• Converting long submissions or minutes into concise summaries;
• Applying consistent tagging or thematic labels across a large body of text;
• Extracting frequently recurring issues to prioritise officer investigation.

These are useful, pragmatic tasks. But important technical and contractual choices determine risk: whether the Copilot deployment is a tenant‑isolated enterprise instance, whether vendor terms forbid downstream model training on council data, how logs and transcripts are stored, and whether inputs are retained beyond project needs.

---

Risks, limitations and real harms to watch for​

• Hallucination and mis‑summarisation.
• Generative models can produce plausible but incorrect summaries or conflate disparate points; an AI might omit a material technical objection or incorrectly attribute a point to “supporters” rather than “objectors,” with downstream consequences for decision‑making. This is not theoretical — it’s a documented limitation of current LLMs and summarisation systems.
• Frequency bias and marginalisation of technical submissions.
• If the algorithm ranks themes by frequency without weighting technical credibility, mass template messages can drown out sparse but critical evidence from statutory consultees or technical experts. Councillors must be made aware of such weighting rules.
• Opacity and auditability shortfall.
• Without publication of prompts, raw AI outputs and human edits, the public cannot readily verify that the AI summarisation was accurate, neutral and complete. The council’s planning portal contains the raw comments, but the interpolation performed by Copilot must itself be auditable.
• Data protection and model‑training risk.
• Public representations often include names, addresses and personal data. Councils must confirm contractual safeguards (explicit non‑training clauses, data residency, deletion rights) and undertake Data Protection Impact Assessments (DPIAs) before sending text into third‑party LLMs. Several councils that have piloted Copilot emphasise DPIAs and governance boards as essential controls. (local.gov.uk)
• Procedural fairness and public trust.
• When an emotive local dispute intersects with innovative technology, even a well‑intentioned administrative shortcut can be perceived as opaque or mechanistic. Opponents may frame AI usage as outsourcing democratic judgment to a black box, which will increase political friction and fuel FOI requests and legal challenges.

---

What good practice looks like — a pragmatic checklist for councils​

The Bath case provides a useful stress test for municipal AI governance. A robust approach should include the following minimum elements:

• Human‑in‑the‑loop as mandatory. Every AI‑derived summary that informs councillors must be reviewed, annotated and signed off by a named planning officer who certifies that material planning points were correctly captured.
• Publish the audit trail. Release (as an appendix to committee papers or via the planning portal) the prompts/instructions used, the AI’s raw summarisation output, and the final, human‑edited summaries with a short note explaining the sampling methodology. This enables councillors, journalists and campaigners to verify fidelity.
• DPIA and procurement protections. Conduct and publish a DPIA specific to the use of public representations; ensure contracts explicitly prohibit third‑party model training on council data, define retention/deletion periods, and lock data‑residency terms.
• Weighting and flagging rules. Distinguish and flag technical submissions from statutory consultees, statutory bodies (Environment Agency, Historic England, ICOMOS‑UK), and expert evidence; present them separately from simple expressions of support or opposition.
• Representative sampling at committee. During the meeting, require officers to present representative original submissions for the principal themes, rather than relying exclusively on tables of summary counts.
• Public communication. Clearly state in consultation materials and on the planning portal if AI will be used and explain the safeguards in plain English; proactive disclosure reduces surprise and builds civic literacy.

These are best practice steps other councils have implemented when piloting Copilot and similar tools. Pilots that paired enterprise Copilot deployments with governance frameworks reported time savings while keeping human oversight central. (microsoft.com)

---

Legal and political context in the Bath case​

The Bath stadium proposal has navigated several high‑profile legal and political crossroads. The development sits within a World Heritage City and has attracted inputs from ICOMOS‑UK, the Environment Agency (raising ecology and flood risk flags), and high‑profile local campaigners. The Secretary of State’s intervention request — directing the council not to grant approval without "specific authorisation" — makes the committee decision potentially subject to central government referral, amplifying the need for defensible, transparent processes at the local level. (feeds.bbci.co.uk)
If a decision is challenged, an appellant may scrutinise whether the council had properly evaluated material planning considerations. In that context, the provenance, accuracy and auditability of any AI‑derived summary may become an evidential issue in appeals and judicial review proceedings. That risk further underlines the need for an auditable trail and explicit human sign‑off.

---

How journalists, councillors and campaigners should respond​

• Inspect the committee papers directly; request the planning officer’s appendices and any AI‑related appendices explaining exactly what Copilot did, how it was prompted, and which officer(s) reviewed the outputs.
• Read the raw submissions on the council’s online planning portal and sample those that appear to underpin each reported theme.
• If the audit trail is not published proactively, seek disclosure (via FOI if necessary) of the prompts, the unedited AI output, and a record of officer edits.
• For councillors: demand transparency at the committee — ask officers to present verbatim representative comments for each major theme during the meeting and to explain their sampling methodology.
• For councils: adopt the checklist above before repeating the approach; publish plain‑English notes for the public.

These steps are practical and proportionate: they don’t ban AI, but they insist that any AI assistance be accompanied by traceable human stewardship.

---

Strengths and legitimate benefits​

• Efficiency gains for routine work. When used correctly, Copilot can reduce the time spent on repetitive triage and drafting, freeing officers to concentrate on statutory assessments and technical inquiries. Case studies from UK councils report measurable productivity improvements when governance is applied. (microsoft.com)
• Consistent initial coding. AI can apply consistent labelling and surface frequent themes quickly, which is valuable when hundreds or thousands of near‑duplicate responses create noise.
• A pragmatic response to capacity constraints. Councils often face episodic surges in consultation volume; a well‑governed AI‑assisted workflow can be an operationally sensible part of a modern planning service.

These are real, demonstrable benefits — but they depend on robust safeguards and transparent human oversight.

---

Where claims remain uncertain and what to verify​

• The planning officer’s report is the primary source for the claim that Microsoft Copilot was used to summarise representations. Local reporting has quoted the officer’s wording, but the definitive verification is the committee report and its appendices on the council’s website; readers should consult the actual committee papers and the planning portal to inspect the precise phrasing and any appended methodological note. Until those documents are examined directly, aspects of the process (exact prompts used, configuration of Copilot, contractual protections) remain partially unverifiable.
• Contractual protections with Microsoft (for instance, whether the council used an enterprise‑grade deployment with explicit non‑training clauses and controlled telemetry) are procurement matters; the council should be asked to disclose the procurement route, DPIA and any contract clauses relevant to data use. If those artefacts are not published proactively, they are appropriate FOI targets.

Flagging these uncertainties is not scepticism for its own sake: it is a demand for procedural clarity where democratic decisions and legal rights are in play.

---

Conclusion — a measured path forward​

The Bath case is a live demonstration of a broader municipal reality: councils will increasingly face surges of public input and will turn to AI to manage scale. That shift is neither conceptually wrong nor inevitably dangerous. But the legitimacy of planning decisions depends as much on transparent process and human judgement as on technical efficiency.
Using Microsoft Copilot to summarise thousands of representations can be entirely defensible — provided the council publishes the audit trail, conducts a DPIA, insists on human sign‑off, and separates out and highlights technical, statutory or expert submissions rather than relying on raw frequency counts. In politically charged planning decisions — especially those that may be called in by ministers and are exposed to legal challenge — these safeguards are not optional niceties: they are essential elements of good governance.
Bath’s planning committee, local campaigners and journalists should therefore treat the AI element as a subject of inquiry, not an argument for or against the stadium itself: ask to see the committee appendices, demand the Copilot audit artifacts, and verify that officer verification was robust. If councils want to harness the practical benefits of AI while preserving public trust, the lesson from Bath is simple and urgent — use AI to scale human judgment, not to substitute for it. (feeds.bbci.co.uk)

---

Source: Bath Echo Council uses AI to summarise thousands of stadium comments