Summary
This March 2024 newsroom experiment remains one of the clearest concrete cautionary records about AI in reporting work. Jon Keegan tested ChatGPT as an assistant on an East Palestine derailment scenario and found that the tool sounded confident while failing on sourcing discipline, coordinates, and evidentiary traceability. The useful nuance is that the failure was not universal: Keegan also found the model more promising as a coding helper than as a reporting brain.
Why It Matters
This is a durable direct journalists story because it does not stop at abstract warnings.
- it shows what goes wrong when a reporting workflow depends on a model that cannot produce reliable receipts for where it got information
- it captures the hidden labor cost of trying to force a chatbot into a rigorous reporting workflow through repeated prompts and corrections
- it distinguishes between bad and better use cases, warning against source-heavy reporting assistance while preserving bounded uses such as code generation
- it documents how one newsroom translated the test into policy by banning AI-created stories or artwork and requiring disclosure, checking, and tool review
What the Source Says
Keegan says he spent substantial time trying to use ChatGPT as an assistant for a hypothetical East Palestine reporting workflow and that the process "didn't go so well." He describes the model giving poorly sourced information, imprecise locations, and answers based on vague "general knowledge." He says the sessions were labor-intensive because he had to keep figuring out where the system got its information and redirect it with precise instructions. At the same time, he says the model was more useful when asked to generate simple Python code. The article ends by describing The Markup's updated AI policy: no publishing of AI-created stories or artwork, mandatory labeling or disclosure, rigorous checking, and case-by-case review of security, privacy, and ethics.