The survey that asks 'one more thing' on purpose

We presented at the Hacks/Hackers Journalism AI summit in Baltimore on May 13, 2026, on a method we've been field-testing for the past year: adaptive chat-based surveys that pair a small fixed core of questions with AI-generated follow-ups in natural language. The full deck is at the bottom of this post.

The premise is that standard surveys answer what you ask, which is a real problem when you're researching hard-to-reach communities whose vocabulary you don't have in your brief.

Fixed prompts flatten how people actually talk about their lives. Adaptive surveys keep a tight set of fixed questions for comparability, then let an AI moderator probe on what the respondent actually said rather than what you predicted they might say.

As Tara Tressel put it in a recent presentation at the Rosenfeld Advancing Research conference, "AI moderation didn't make research easier. It made it possible under real-world constraints." It's a research instrument with its own validity profile, useful exactly where the alternatives fail.

We shared two field tests. The first looked at YouTube trust during the Iran-Israel conflict shutdown, run on Manychat plus Telegram with Telegram mini-app ads. Adaptive follow-ups surfaced workarounds (VPN use, offline device transfers) we'd never have thought to ask about, and the mini-app ads outperformed every other recruitment channel we tried. (Read more here.)

The second was with Latino migrants in Philadelphia, run on Yazi plus WhatsApp plus Facebook ads for a local newsroom there. The AI moderator handled code-switching between Spanish and English in real time, and WhatsApp lowered participant friction because the channel was already familiar. Across both, the consistent finding was that the platform people are already on is the platform that works.

Three ideas held up across the work:

1. First, the method is the discipline, not the product. The written brief and the fixed-question core are what carry the study; the tool is interchangeable, and the market is volatile enough that you should design studies that outlive any single vendor.
2. Second, AI earns its keep with hard audiences. Speed, channel fit, and adaptive register matter most exactly where other methods fail: conflict contexts, migrant communities, fast-moving stories.
3. Third, this is not plug-and-play. We saw rich, on-topic responses next to combative ones, off-topic tangents, and a fair amount of noise that needed manual filtering. Response quality varies, and analysis takes more work than counting Likert scores.

We also tried to be honest about what doesn't work: AI moderators sometimes restate the question instead of digging deeper. They can drift into leading prompts if the brief is underspecified. Outputs need more interpretation than survey results. Without pilot testing, you risk laundering the model's assumptions back as participant voice. Foundational discovery on trauma, trust repair, and other high-emotion topics still needs human moderators.

None of this is fatal, but none of it disappears just because you pick a different vendor. Go in with eyes open, pilot with two or three participants before launch, and review transcripts daily.

If you're considering this for your newsroom, the order we'd suggest is audience first, then question type, then tool. Bounded discovery, validation, and feedback fit adaptive surveys well; sensitive subjects mostly don't.

The full deck has the method diagram, the field test details, a comparison of vendors in the space and what each is actually built for, and a longer list of tradeoffs and mitigations.

Scroll through the full deck:

If you're working on something where this method might fit, book a Service Desk call or email us at servicedesk@gazzetta.xyz. We're also gathering people thinking about this work at tbd/con on September 23 and 24, 2026.