AIdas
Every answer is an echo. We help you hear whose.
People increasingly get their information from an AI model, not a search page.
AIdas audits how those models respond to the questions users actually ask about contested issues: which sources they surface, whose narrative they carry, and how that shifts by wording, language, and time.
We turn that into workshops, dashboards, and studies you can act on.
What we offer
1. Lessons in literacy (workshop)
A hands-on session where your team tests how AI models handle the contested topics in your own context. Bring a question, and we run it across several models side by side to read the bias, the sourcing, and whether the model pushes back or just mirrors the prompt.
- Live testing on your topics, in your languages
- A simple, repeatable method your team keeps
- A first baseline you can return to
For newsrooms, schools, libraries, and civil-society teams.
2. Aggregate testing dashboards (monitor)
We run a fixed set of prompts across many models on a schedule and turn the results into a dashboard that tracks issues over time: narrative alignment, source diversity, refusals, and drift. One run is a snapshot. Repeated runs are a baseline that catches change before anyone else notices it.
- Reproducible, scored testing across models and languages
- Trends over time, not one-off screenshots
- Exportable data and a shareable live dashboard
For funders, research teams, and organizations tracking a topic or a market.
3. Ecosystem analysis (study)
Who already dominates what AI says about a given place, topic, or community? We map which outlets get cited most in AI answers, where they are based, and whether state-controlled or propaganda sources are creeping in. It shows the information ecosystem the models build on, and where independent voices are missing.
- Source and country-of-origin mapping
- State-media and propaganda detection
- A visibility-gap report with clear recommendations
For public-service media, media funders and investors, and researchers.
Proven, not theoretical
Idas grew out of commissioned research, in collaboration with Factnameh, and with support from ASL19 and the Open Technology Fund, and was first presented at GlobalFact in Vilnius and SRCCON in Minneapolis.
The method is already published:
- Meet AIdas: our prototype AI bias training tool
- AIdas: a visualization of how AI models answer Iran
- The new gatekeepers: auditing how AI models assess authority (Iran)
- How prompt framing steers models toward different source libraries
- DeepSeek's double life (China)
- Censorship as a coordinate system, not a blocklist
Work with us
Tell us the context you care about, a topic, a country, a community, or an audience, and we will scope the right format: a workshop, a monitoring dashboard, or a full ecosystem study.