Pablo Zavala · AI Safety Evaluation · Research Engineering

Heard.now and Checkable Civic Listening

Heard.now frames civic AI as a checkable listening loop: residents speak plainly, privacy gates protect raw input, sample limits accompany each finding, and campaigns show what changed.

July 6, 2026 · 7 min

Most campaign technology helps politicians speak faster. Heard.now points the other way: can a campaign listen at scale, keep residents' words useful, reduce exposure for the people behind them, and return with a public answer: what residents said, what changed, and which claims the data leave unsupported?

The unsupported-claims clause gives the promise its discipline. "Listening" is one of the easiest promises in politics and one of the hardest to verify. A candidate can hold a town hall, read a few emails, cite a poll, and say the community has spoken. Yet every channel captures something and drops something. Town halls privilege people with time and confidence. Long-running research on civic participation finds unequal voice across resources, education, and institutional fluency; digital tools can preserve the same pattern unless process design corrects for the bias. Polls force people into predefined choices. Inboxes collect stories but rarely turn them into a public record. Social media turns outrage into measurement because outrage is easy to count.

Every channel drops signal, so campaign efficiency is the wrong center of gravity. Campaigns are already efficient at broadcasting. A better question asks whether AI can make listening more accountable.

Heard Is a Listening Loop

The public Heard dashboard is a synthetic campaign report. The dashboard reports 296 synthetic responses for an April 2026 Pittsburgh mayoral scenario, then separates policy areas, themes, neighborhoods, coordinated-input warnings, sample composition, and recommended actions. The page labels the responses as synthetic rather than real Pittsburgh voters.

The synthetic sample makes the demo a product hypothesis; voter-opinion claims require recruited people, disclosed channels, and real field conditions. Within the synthetic-data boundary, the dashboard makes a useful claim: a listening report can preserve evidence, flag caveats, and require a campaign-facing response without pretending to measure the district.

Because Heard's public pages frame the tool as campaign technology, the civic claim should rest on campaign accountability rather than neutral public governance.

The demo's strongest move is its treatment of coordinated input. Public Safety appears dominant as a raw count, but the dashboard separates a 100-response advocacy spike from responses outside the demo's coordination rules. The dashboard preserves the organized position as valid input within the synthetic scenario while preventing a burst of duplicated pressure from masquerading as broad priority. A listening system should preserve organization and make organization legible.

Disclosure Comes First

The coordinated-spike example points to the larger rule: civic input is inherently selective. AAPOR's transparency standards and Pew's work on opt-in polling both point to the same discipline: tell readers how people were recruited, who answered, what the sample can support, and what the sample leaves unsupported. A listening channel can reveal what self-selected residents chose to say. A self-selected channel describes speakers rather than estimates district opinion.

The report should publish the boundary plainly.

A serious Heard.now report should state participation limits beside every conclusion. When R-leaning residents are underrepresented, the report should say so. When a subgroup is smaller than a useful threshold, the report should mark the finding as directional. When input comes from a campaign link, a community group, a QR code, or paid outreach, the report should name the channel. When a campaign asks people to speak before the campaign has real decision space, the report should say so as well.

OECD guidance on citizen participation starts from the same premise: participation is meaningful only when people can affect the decision. Likewise, IAP2's spectrum separates informing, consulting, involving, collaborating, and empowering into different promises. Heard.now should help a campaign keep the promise the campaign actually made.

Privacy Needs Concrete Design

Disclosure handles interpretation; privacy handles exposure. In civic listening, people may write about rent, schools, police, illness, work, family, immigration, debt, and fear. A system that invites those stories has to minimize personal data before the story becomes a civic record.

I tested the Heard idea through a local reference module, separate from any verified claim about Heard.now's production architecture. The module treats privacy as a system property rather than a paragraph in a policy page. Resident free text is redacted before storage. Optional follow-up contact is stored separately and withheld from the campaign query surface. The campaign reads through approved public views instead of raw tables. Themes and individual-level documents clear a distinct-message k-anonymity floor before they appear. A portable public extract omits raw submissions, contacts, ZIP codes, user agents, and below-floor documents by construction.

Modesty matters here. Minimization falls short of perfect anonymity. Pattern redaction can miss a self-identifying description. Small slices can still create inference risk. The k-anonymity floor counts distinct normalized messages rather than verified people, so duplicate flooding receives one protection while paraphrase flooding remains a pilot risk. Public-sector deployments may trigger public-records, retention, and redaction duties that vary by jurisdiction. The prototype's real claim is stronger because it is narrower: it reduces exposure, gates individual-level outputs, and makes the publication boundary checkable. Legal compliance, production role separation, third-party model retention, and campaign-law fit remain outside the reference-module claim.

Boring Primitives Carry the System

The privacy boundary depends on ordinary machinery. SQLite serves as the source of truth. Under the initialized app schema, the campaign can query public views with read-only SQL. A SQLite authorizer denies direct private-table reads, sensitive columns, SQLite schema-table reads, unsafe functions, and writes at the database layer. CSV exports are formula-safe. A progress handler limits expensive queries.

Theme clustering runs offline and deterministically. The system works without an LLM deciding which themes exist. The LLM boundary sits where the model can help without governing the record: the resident conversation can feel plain and adaptive, with a local fallback if the provider is unavailable. Model output is parsed as bounded data rather than executed as authority.

The audit layer stays similarly plain. Each save appends a hash-chained record. Published extracts carry anchors that can be checked later against a retained source log or extract, so a later reader can detect whether the source log was shortened behind the extract. A hash chain falls short of tamper-proofing the record, but retained anchors make edits, rollback, and tail truncation harder to hide.

The design posture I trust most in civic AI uses boring primitives, explicit limits, and checks a skeptical reader can rerun. In a separate SDRAEH reference-module run on July 6, 2026, `PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python3 -B -m sdraeh verify` reported 12/12 passing checks: determinism, k-anonymity, below-floor isolation, redaction, SQL guardrails, audit integrity, host validation, provider safety, and published-extract boundaries. The command verifies the local SDRAEH reference module rather than Heard.now production. The current public zavalab project page keeps a narrower claim: a synthetic public sample with a 7/7 verification path.

AI Should Organize Attention

The verification posture matters because Heard.now enters a crowded field. Decidim, CitizenLab/Go Vocal, EngagementHQ, Polis, and other systems already collect, organize, and report public input. Heard earns a place only if campaign staff can trace each recommendation to evidence without exposing residents, see sample caveats beside each claim, and publish a follow-up ledger that names accepted, rejected, and pending requests.

Polis and vTaiwan show one path for computation in democratic practice: help large groups express, cluster, and compare views without collapsing disagreement into one fake consensus. Team Mirai belongs here as a current analogy for technology-mediated politics. Heard should borrow that restraint for campaign listening, because each example treats computation as a way to organize attention rather than replace politics.

Heard.now occupies a smaller category than Polis or a citizens' assembly. The product qualifies as deliberation only when residents reason with one another under a process designed for that purpose. Even so, the narrower job remains valuable: make plain-language input easier, preserve the structure of what residents said, reduce exposure for the people behind the words, and make the campaign answer back.

The answer-back step separates collection from accountability. Collecting pain is cheap. Closing the loop is costly. A campaign that asks people to tell the truth about their lives owes more than a message test. The campaign owes a public account of what residents said, what changed, which requests the campaign rejected, and why.

Five Tests for Heard.now

A rigorous civic-listening system should meet five tests.

First, residents should be able to speak in plain language. The interface should ask one useful follow-up at a time and avoid turning a human story into a form before the person has finished thinking.

Second, the analysis should preserve evidence. Themes, summaries, and recommendations should remain internally auditable against the underlying input without exposing private messages in public.

Third, privacy should be a technical boundary. Minimize data before storage, separate contact from content, publish only approved views or extracts, and state residual risk.

Fourth, every finding should carry its sample limit. Self-selected input can guide attention when reports avoid impersonating polls.

Fifth, the campaign should close the loop. A listening report without a visible action record remains an inbox with charts.

The five tests are hard because they make listening less convenient for the institution collecting input. The friction is useful. K-anonymity floors, public caveats, query boundaries, extract checks, and audit logs fall short of proving good faith. Together, those controls give readers handles to check the listening claim.

Listening Has to Leave a Record

The promise is smaller and more demanding. A campaign that asks people to speak should reduce their exposure, preserve the record, disclose the limits, and show what changed. AI can help organize parts of that work. Political courage still belongs to humans.

For that reason, the product should stay boring, local, and verifiable for as long as possible. The future of civic AI should move beyond fluent bullhorns toward listening systems that leave records.

Sources