How government teams use narrative intelligence to understand public policy conversations
.png)
Narrative intelligence helps government teams see which stories about a policy are being told in public content, how fast each is spreading, and whether any rests on a factual error. It describes conversations, not beliefs: public posts show what people chose to publish, which is one window into opinion, not the whole of it. Treating the two as the same is the error this piece is written to avoid.
The sections below walk the practical use case: what narrative intelligence is, how AI helps read spoken video alongside captions, where it fits in a public sector decision, and the safeguards a government team should set.
What narrative intelligence does
It turns a large volume of public posts into a short list of stories people are telling about a policy, with the evidence behind each. A communications team can see that "the new charge is a tax on workers" is being discussed in rising volume while another claim has faded, and focus effort where it matters.
Stories here means recurring themes in posted content, not confirmed states of public opinion. A rising theme in posts is evidence that posters are talking about it; whether it reflects the broader population's view is a separate question, usually requiring survey research or qualitative work alongside.
The lineage of this discipline is older than the current wave of platforms. For the history in open-source intelligence, see from the Arab Spring to narrative intelligence.
How AI helps read spoken video
Much public conversation now lives in spoken video, where the point is made out loud rather than written in a post description. AI transcribes that speech, reads it alongside the description, and groups posts that make the same point in different words.
Channel relevance matters here. In the US, 43% of adults under 30 regularly get news on TikTok (up from 9% in 2020). That figure describes a specific group (young US adults) using a specific channel (TikTok) for news; it does not establish that policy discussion on that channel is predominantly spoken rather than written, which is a separate content-modality question. It does say that for one important audience, TikTok is a channel worth including in a policy monitoring scope.
Speech transcription accuracy varies by language, so findings in each language need a local reviewer.
Supporting a public sector decision
Narrative intelligence gives decision-makers evidence, not anecdotes: which stories exist in posts, how fast each is growing in volume, which content signals indicate who is carrying it, and whether its factual claims are correct.
The UK Government Communication Service's RESIST 2 counter-disinformation toolkit (available via GCS's counter-disinformation page) describes the same sequence from Recognise through Early warning, Situational insight ("so what?" briefings for senior officials, key takeaways first), Impact analysis, Strategic communication and Tracking effectiveness. Narrative intelligence supplies the content evidence for those steps; survey research, policy records and local knowledge supply the rest.
Separating opinion from factual verification
A platform can surface a claim. Whether the claim is true is a separate check against authoritative information. In the worked example below, "hospital trips are exempt" is a factual question answered by the city's own policy document and the regulation that implements it, not by how many videos say otherwise. Narrative intelligence tells you the claim is spreading; the policy team tells you whether it is correct.
From narratives to a briefing
Short, key takeaways first. One story per card. Each card carries the content evidence (volume, views, example posts) and the factual check separately. A senior official reading the briefing can click through to either.
Deciding whether to respond
Response priority depends on several factors, weighed together.
- Factual accuracy. Is the claim true, false, or contested?
- Potential harm. Could acting on this claim cause injury, loss, or missed services?
- Spread. Volume now, trajectory over time, views reached.
- Amplification risk. Would a correction amplify the claim beyond the audience currently reached?
- Uncertainty. Is the content signal strong, or ambiguous enough to need more review before action?
A factual error is one consideration, not the top of a fixed priority list. A small-volume false claim that risks harm (encouraging people to skip a medical appointment, say) ranks differently from a large-volume false claim that is self-correcting in the comments.
What it looks like in practice
Illustrative only, not a dig benchmark.
Westmere is a fictional city that announces a congestion charge. Over the next month, its communications team reviews public video about the charge. The analysis groups posts into three stories.
Assignment rule for posts that make more than one claim: a post is counted in each story it clearly makes, and the overlap is reported. Dedup for the overall total.
Deduplicated unique posts across the three stories: 420 + 180 + 60 − 25 − 10 = 625 unique posts.
The combined-views column is overlapping. Views for the 25 posts that carry both A and B are counted in both A's 2.6M and B's 700K; the 10 posts carrying both A and C are counted in both A's and C's totals. The three view numbers cannot be summed into a single "4.4M total views" without deduplicating the overlapping posts' views, which the team did separately before any single aggregate figure went into the briefing. Current cumulative views reflect what the posts have reached by the time of collection, not what they had reached at any earlier point.
Location signals: aggregate-level regional breakdowns rely on declared or consented geolocation (platform signals that posters have agreed to share, or self-declared location in bio or post). Content alone does not identify where a poster is from. Where declared signals are sparse, the briefing reports it as such and does not infer.
What the team did
Story C is a factual error (hospital trips are exempt, per the city's published scheme). Harm risk is real (people skipping hospital trips to avoid a fine that does not apply). The team published a short spoken explainer on the exemption and briefed hospital front desks. Story A is a policy disagreement, not an error; it went to the policy team as input, not to a rebuttal. Story B was tracked without immediate action.
Response by narrative. A: assess and route to policy. B: monitor. C: correct and brief front-line services. Not purely in volume order, and not purely in "factual error first" order either; a combination of accuracy, harm risk and amplification trade-off.
What they did not do
No profiling of individual creators. No inference of individual identities. The team worked with aggregate public content only. That constraint is non-negotiable in a public sector setting, and the method note in every output states it.
What to look for in a platform
Ask five questions.
- Does it transcribe speech and read it with the post description, in the languages your public uses?
- Does it group posts into stories and link each back to the source?
- Can it show aggregate volume by region when posters have declared or consented to location signals, and clearly report when they have not?
- Does it tell you what is outside its coverage?
- Can your team test it on examples you already know?
For how a monitoring pipeline's visibility can be measured responsibly, see our four-stage detection audit.
Safeguards a government team should set
Work with aggregate public content only, never individual profiling. Treat content signals as evidence, not proof, and send uncertain cases to a human reviewer. Separate meaningful discussion from incidental mentions. Record what is out of scope by platform and language. Keep opinion-from-content and factual verification in separate columns of the briefing. Publish the purpose of the work internally so it is clear this is about understanding public policy conversations, not surveilling citizens.
How dig helps
dig collects public social video, transcribes speech in 100+ languages and analyses posts, descriptions and transcripts in their original language; accuracy varies by language. The Narrative Intelligence app groups public content into stories and links each to the videos behind it. See the RESPOND framework for the sequencing of signal to decision.
Pro tip
Start with the narrative built on a factual error only when the harm and amplification trade-off backs the call. Volume alone is not the trigger.
Key takeaways
- Narrative intelligence describes public policy conversations in posted content; survey or qualitative work is needed to establish broader opinion.
- Keep factual verification separate from content signal; the platform surfaces claims, authoritative information decides whether they are true.
- Response priority weighs accuracy, harm, spread, amplification risk and uncertainty together.
- Stories that overlap in a post need an assignment rule and a dedup for the overall total; overlapping view totals cannot be summed without the same dedup.
- Audience location is reported from declared or consented signals, not inferred from content.
- Aggregate public content only; no individual profiling; human review for uncertain cases.
Frequently asked questions
How does narrative intelligence help governments?
It shows which stories about a policy are being told in public content (including spoken video), their volume and growth, and whether a story's claims are factually accurate (checked against authoritative information). Communications teams use this to decide when to correct, explain or pass to policy.
What is public sector narrative intelligence?
The use of open-source public content to understand public policy conversations. It groups posts into stories, links each to its sources, and works at aggregate level.
What role do narrative intelligence platforms play in public sector decision-making?
They supply the content evidence: which stories exist and how fast each is growing. Decision-makers still decide. Pair with survey research, policy records, local knowledge and (for factual claims) authoritative verification.
Related stories

.png)

.png)