See Selected Projects

Government & public servicesUnited Kingdom

Analyse thousands of public-consultation responses fast

The UK's Consult tool themed 2,000+ responses for the Scottish Government; officials checked every theme.

Try it withClaudeChatGPT / OpenAI modelsGeminiGrok

Deployed, UK Government (i.AI, Humphrey suite); Scottish Government

Why it matters

Governments ask citizens for views, but reading every response is slow and costly. Faster, checked analysis means more of what people say reaches the decision.

How it works

Consult first proposes themes for each question, then maps every response to them. On the Scottish Government's consultation on non-surgical cosmetic procedures, the responses covered six questions; experts checked and refined the themes and, for this first live use, also reviewed every response manually (GOV.UK). In i.AI's evaluation, 60% of responses got exactly the same themes as human reviewers.

Who it's for

Policy teams, regulators, councils and research agencies.

The prompt

Below are anonymised responses, with IDs, to the consultation question: "[question]".

Stage 1: Read all responses and propose 6 to 12 themes. For each, give a one-line definition and two representative quotes of one sentence each. Then stop and wait for me to edit the themes.
Stage 2: When I reply, assign every response to one or more themes and output a CSV with response_id and themes.
Stage 3: Report the count and share of each theme, notable minority views, and responses that fit no theme.

Do not paraphrase quotes. Do not try to identify anyone.

[Paste responses with IDs]

Open inClaude (opens in a new tab)ChatGPT (opens in a new tab)Gemini (opens in a new tab)Grok (opens in a new tab)

Sources

BBC News and Civil Service World reported the trial independently, but the accuracy figures in all sources come from the government's own evaluation.

  1. 1. GOV.UK (DSIT): 2,000+ responses; six questions; experts checked themes; manual review. Government-built 'Humphrey' AI tool reviews responses to consultation for first time (opens in a new tab)

  2. 2. Incubator for AI (i.AI): Two-stage method; 60% exact match; F1 scoring. Evaluating Consult (opens in a new tab)

  3. 3. GOV.UK: Transparency record; tasks call LLMs. DSIT: Consult (algorithmic transparency record) (opens in a new tab)

  4. 4. BBC News: 2,000 submissions; 'nearly identical' results to human experts working in parallel; still in trial; humans in the loop. AI tool put to test sifting public views on botox and fillers (opens in a new tab)

  5. 5. Civil Service World: Six questions; officials manually reviewed each of the 2,000 responses; F1 0.76; more evaluation before rollout. Civil service AI tool for consultation analysis passes first public test (opens in a new tab)

Related projects