Using AI to sort and summarize public comments on environmental documents

Reference article · 5 min read

When a comment period closes on a draft environmental impact report or a general plan update, the volume that arrives in the final days often exceeds what the assigned planner can read before the final document is due. Organized campaigns produce hundreds or thousands of near-identical letters. Individual submissions repeat points already raised, arrive as scanned attachments, or come in a language other than English. The obligation to prepare written responses does not shrink because the inbox grew.

Language models are genuinely useful for a narrow set of tasks here: grouping similar text, detecting near-duplicates, drafting internal summaries for triage, and routing comments to the right technical reviewer. They should not decide which comments legally require a response, write responses that leave the building unreviewed, or stand in for the original letter anywhere in the record. Most of the work of doing this well is keeping that line visible.

Freeze a complete intake record before any processing

Before a model touches anything, capture every submission as received. That means the original file or message in its original format, plus metadata: date and time received, channel (portal, email, mail, hearing transcript, voicemail), the name and affiliation as the commenter gave them, and a list of attachments.

Assign a stable comment identifier at intake. Every later artifact, cluster, summary, and response references that identifier. Anything the model produces is a derived working copy; the original stays untouched and goes into the record in full. Close the corpus at the end of the comment period and log late submissions separately rather than mixing them in, since how those are treated is a legal question for your counsel, not a processing question.

Separate substantive comments from campaign duplicates

Deduplication is where automation earns its keep, and where it can quietly cause harm.

• Exact and near duplicates. Hashing catches identical text. Embedding similarity catches reformatted or lightly edited versions. A person, not the tool, sets the similarity threshold, and staff should read the borderline band rather than trusting the cutoff.
• Template plus personalization. Many campaign letters include a few sentences the signer wrote themselves. Diff each letter against the detected template and preserve the unique text as its own item. A personalized insert buried in a form-letter bucket is the most common and most damaging error in this workflow.
• Topic clustering. Group by subject (traffic, water supply, greenhouse gas analysis, alternatives, noise) so each cluster routes to the reviewer who can actually answer it.
• Substantive screening as a flag, not a decision. A model can propose that a comment raises an environmental issue. Staff decide. Under CEQA, a lead agency evaluates comments on environmental issues raised during the review period and prepares written responses; the scope of that duty and the timing requirements before certification are set by statute and the CEQA Guidelines, and details change, so confirm the current requirements with counsel. Federal NEPA practice has its own rules, which have been revised in recent years.

Keep every response traceable to a source letter

Build a comment and response matrix that runs in both directions. Every comment identifier must appear in at least one cluster, and every drafted response must cite at least one verbatim excerpt with its comment identifier and page reference. Run both checks as a hard gate before the final document is assembled.

Responses should be written against the quoted original text, not against the model's paraphrase. Keep the excerpt visible next to the draft so the reviewer can see what is actually being answered. Where a master response covers a campaign, list the identifiers it covers rather than summarizing the count in prose alone.

Attachments, scans, and non-English submissions

Scanned letters need optical character recognition, and the extraction confidence should be visible to reviewers so low-confidence pages get read by a person. Technical reports submitted as exhibits should be routed whole to the technical reviewer, not compressed into a bullet by a summarizer.

For non-English submissions, machine translation is reasonable for initial triage and routing. Anything that will be quoted or answered substantively deserves a qualified human translation, and both the original and the translation belong in the record. Language access obligations may apply under Title VI for recipients of federal financial assistance and under state and local policies; requirements vary, so check yours.

Say what you did in the record

A short methods note in the response-to-comments introduction costs little and prevents an avoidable argument later. Describe how comments were organized, state that automated grouping and deduplication were used as a staff aid, confirm that all comments appear in full, and state that responses were prepared and approved by agency staff.

Retain the configuration: model and version, similarity thresholds, prompts, and run dates. Prompts and outputs may themselves be public records and may fall under a litigation hold, so decide retention deliberately rather than letting them live in a chat history.

How to start and what to ask vendors

Test on a closed project first. Run a past comment set through the workflow and compare the clustering against the responses the agency already published.

Ask any vendor: where data is processed and stored, whether agency inputs are used to train models, whether thresholds are adjustable and logged, how OCR confidence is surfaced, whether the full matrix exports to open formats, and what happens to the data at contract end. In the contract, require no training on agency data, deletion on request, subprocessor disclosure, exportability, and vendor cooperation with public records requests.

HQT works with state and local agencies on data and document workflows of this kind, including intake, records handling, and system integration.

Frequently asked questions