Sparring, not answers

Seeing an example early pulls your ideas toward it. In a study of designers, those shown AI-generated images while brainstorming fixated more on them and produced fewer, less varied ideas (Wadinambiarachchi et al., CHI 2024). In another, writers given AI ideas wrote stories rated as more creative one by one, but the stories became more alike as a group (Doshi and Hauser 2024). If the model writes the first draft, your design converges on everyone else's.

The alternative is to keep the pen and hand the model a different job: opponent.

Prompts that spar

• "Here is my design. Give me five ways it fails in production, most likely first."
• "Assume this shipped and caused an incident six months later. Write the incident report."
• "What change request from the business would hurt this model most?"
• "Make the strongest case for the option I rejected."

Then do the part the model cannot: decide which attacks are real, and turn each real one into a check.

Sparring sessions in this module are studio work, not graded: there is no right list of attacks. The variants lessons are graded: your checks run against designs with faults injected on purpose.