Why we’re building it
How a story lands in different communities is central to trust, and newsrooms rarely have the time or money to find out before they publish. Focus groups are expensive and slow. Most of the time a story goes out and the newsroom learns how it was read from the comments, if at all.
Mirror is the audience-side companion to Bedrock. Bedrock asks what verified reporting an AI should be handed. Mirror asks how real people, in all their differences, might hear the result, and lets a journalist interview AI “digital twins” of audience segments about a draft while there’s still time to change it.
Digital twins grounded in real voices
The work builds on Kaveh Waddell’s Verso and on research from Northwestern, Stanford and Google DeepMind showing that agents grounded in real interviews can match people’s survey answers about 85% of the time. The key finding was that demographics alone aren’t enough. What makes a twin believable is the texture of how a person actually talks about their life.
So the team wrote eighteen personas rooted in Evanston and Chicago, each with demographics, values, media habits, a way of speaking and, most importantly, interview excerpts. A retired public school teacher in Rogers Park. A bakery owner in Pilsen. A dining hall lead at Northwestern who is a union steward, gets her news from Univision and WhatsApp, and is skeptical of coverage written about her community rather than for it.
Which approach makes a twin honest
The research question was which prompting approach produces a twin that is both authentic and stable. The team compared four: bare demographic facts, a full character sheet, real interview excerpts, and asking the model to reason about the person’s identity before it answers.
Then they tried to break them. A fifteen-question battery probes identity, moral pressure and resistance to flattery, the moments where a model tends to drift back toward its own agreeable voice. The team ran the same battery across five models from Anthropic and OpenAI and at different temperatures, and compared answers across runs to see which twins held their shape and which dissolved into a generic helpful assistant.
Twins reading a real story
In the article reaction mode, a journalist pastes in a draft and each twin responds in character. Given a story about a new apartment development, the Pilsen bakery owner went straight past the headline to the number of affordable units and to a longtime small business on the block, because that is who she would be worried about. That’s the kind of reading an editor can act on: not a score, but a sense of what a particular community will notice first.
The team also built a conversation mode where several twins talk to each other about a topic over multiple rounds. It’s a way to see whether personas stay distinct when they push back on one another, or start to sound alike.
Who is in the coverage
To ground the personas in a real newsroom, the team audited eight weeks of The Daily Northwestern, 758 articles from spring quarter 2026, tagging each by desk and by subject. Arts and athletics together made up nearly 40% of everything. Local civic coverage, housing, local business, Evanston government, schools and the environment, was steady but thin: each theme only 2 to 4%, about 18% combined. That map shows which communities’ stories are being told, and which twins a newsroom most needs to listen to.
Guardrails, not a substitute
The point isn’t to replace listening to real people. It’s to do more of it when resources are thin, and to label synthetic feedback clearly so it works as a guardrail and not a stand-in for the people it represents.