Media Party New York @ Brown - 2026

Small, mission-driven newsrooms face a persistent editorial bottleneck: too many pitches, too few editors, and a set of editorial values around representation, framing, and sourcing that resist easy automation. At Minority Africa, a Uganda based publication covering minorities on the continent, we wanted to know whether AI could help triage our now too many pitches without flattening the judgment behind it.
In 2025, we built Iraka, a custom GPT-based tool trained on a structured dataset of over 400 de-identified historical pitches paired with editorial decisions, reviewer comments, and outcome labels. The system prompt encodes Minority Africa's editorial criteria (including originality, representational depth, feasibility, and alignment with the publication's justice-oriented mission) and the tool returns a decision prediction (accept, develop, or reject), a score out of five, and written feedback designed to mirror the developmental tone of our editorial process.
In phase one, we benchmarked Iraka against real pitches it had not previously seen. It matched editorial decisions about 70% of the time, but feedback alignment was significantly lower at around 30%, showing that while decision replication is achievable, generating the kind of nuanced, contextual feedback editors actually give remains a major gap. We also tried to build it across different models with varied outcomes.
Phase two, now underway, targets the feedback layer directly. We are building synthetic and counterfactual pitch sets — controlled variations where elements like framing, sourcing, or representation are intentionally altered — to stress-test the tool on edge cases and ethical dilemmas absent from our historical data.
This talk shares what we built, what it got right, where it broke down, and the harder question underneath: what does it mean to encode editorial values into AI when those values center justice and representation?