An AI research group tested six models on its own work, and unwritten style rules tripped them up
Epoch, an AI research group, gave six AI models real tasks from its own work, such as graphics and short data articles. It reports the top closed models handled well-defined steps reliably but missed unwritten house conventions.
Why it matters
Epoch says it now largely automates restyling charts to its house style. Yet the models misjudged visual taste and audience interest despite ample reference files. Open-weight models, whose files are publicly downloadable, made factual errors. Each model ran once per task with one grader, so read this as one team's experience.
Your moveTest one repeatable task, like matching a chart to your style guide, and check the result against your brand standards before using it.
Read the source