Accelerating narrative discovery by building trust in Generative AI at Blackbird.ai
After conducting in-depth interviews and mapping intelligence analysts’ workflows, I designed the integration of Generative AI into Blackbird’s mature Constellation platform to accelerate the discovery of potential disinformation narratives. The resulting AI-powered Narrative Feed turned large-scale social media conversation data into inspectable, editable narrative candidates connected to the existing workflow, so analysts could use AI as a trusted starting point for investigation, not an automated conclusion.
We believed Generative AI could significantly accelerate the discovery of potential disinformation narratives within a large database of online conversations.
But the workflow had a trust gap.
Analysts still had to:
- jump back and forth between Narrative Builder and Analyze;
- open multiple tabs to compare and refine narratives;
- keep external notes or Google Docs during discovery;
- manually validate whether AI-generated narratives were worth pursuing.
The issue was not lack of AI output. It was the gap between AI suggestion and analyst adoption.
The challenge was not to design “another AI feed.”
It was to make AI-generated narratives:
- easy to scan;
- easy to validate;
- easy to move into the existing workflow;
- and clearly distinct from analyst-authored narratives.
I led the design work across:
- problem framing;
- research synthesis;
- UX strategy;
- feed, detail, and builder integration flows;
- source and state design for AI-generated narratives.
The in-depth interviews showed that analysts did not follow a single path into narrative discovery. Some began with a client question and a basic Boolean query; others explored top terms, engagement spikes, and influential posts to identify emerging themes.
Despite these different starting points, their workflows converged on the same pattern: discovery was iterative and fragmented. Analysts moved back and forth between Narrative Builder and Analyze, kept evidence and potential themes in external documents, and repeatedly refined queries to separate meaningful conversations from spam and irrelevant content.
Analysts did not want AI to replace this judgment. They wanted it to reduce the search space and make the first pass faster.
The most useful role for the feed was not to deliver a final narrative, but to surface promising hypotheses that could be quickly reviewed and then brought into Builder and Analyze for refinement.
This shifted the design direction:
Generative AI should accelerate discovery while preserving analytical judgment.
End-to-end Interactive Prototype
1) Design the feed for scanning, not deep reading
Research showed that headlines were the most valuable part of the feed during early discovery. Analysts wanted to quickly identify potentially relevant conversations before committing to deeper analysis.
So the feed prioritized:
- strong narrative headlines;
- post volume and platform distribution;
- compact summaries;
- quick access to detail and next-step actions.
This reduced the cognitive cost of triage and made the feed feel more operational.
2) Reframe AI output as a starting point, not a final answer
One recurring issue was low trust in summaries, clustering quality, and risk scoring. Instead of over-emphasizing AI confidence, the experience was designed around progressive trust:
- AI suggests a narrative.
- The analyst reviews it.
- The analyst opens details.
- The analyst moves it into Builder.
- The analyst edits, saves, or discards it.
This made the AI useful without overstating its certainty.
3) Connect the feed directly to the core workflow
The feed only became valuable when it stopped behaving like a standalone surface.
The prototype connected the flow across:
- empty state generation;
- populated feed;
- narrative detail;
- Narrative Builder;
- draft and save states;
- version history.
This allowed an analyst to go from “this looks relevant” to “this is now an editable narrative” with much less friction.
4) Make source visible
One of the most important AI UX decisions was showing where a narrative came from.
The system introduced visible cues for:
- AI-generated narratives inside the feed;
- AI-origin narratives inside the selector;
- an “Original AI-Powered Narrative” reference in Builder;
- history states showing when an AI narrative was added and later saved.
This helped analysts understand whether they were looking at:
- an AI-generated candidate;
- a draft derived from that candidate;
- or a saved narrative they now owned.
That distinction was critical for trust.
5) Support editing, saving, and versioning without breaking confidence
The handoff into Builder was not just navigation. It was a state transition.
Once inside Builder, the analyst could:
- expand or refine the query;
- apply filters;
- generate subsets;
- rename the narrative;
- save changes;
- review version history.
This gave the user control over the final analytical object while preserving the value of the AI-generated starting point.
Qualitative
Reduced friction between AI-assisted discovery and manual narrative creation
- Created a clearer bridge between Feed, Detail, Builder, and Analyze
- Made AI-generated narratives feel inspectable and editable rather than opaque
- Improved the product’s ability to support real analyst workflows instead of only showcasing AI output
Product impact
- Positioned the Narrative Feed as a discovery layer connected to Constellation, not as a siloed feature
- Clarified the role of AI in the workflow: suggestion first, validation second
- Established a stronger model for source, narrative ownership, and save-state logic
This project reinforced that AI UX is not mainly about showing intelligence. It is about placing AI output inside a workflow users already trust.
For analysts, the best use of AI was not automatic conclusion-making. It was faster discovery, clearer starting points, and a smoother path into human refinement.
The real design challenge was not generating narratives. It was making them usable.