Stacked Podcast discusses a Connecticut self-represented litigant who embedded text in a court filing instructing any reviewing AI system to agree with the filing. According to the episode, the reviewing tool detected the instruction and warned its user, so the discussion does not establish that the prompt changed the court's decision.
The episode uses the incident to examine prompt injectionPrompt injection occurs when untrusted content attempts to redirect an AI system's behaviour beyond the authority of that content. wherever people pass untrusted documents into AI systems. It argues that legal, hiring and other high-stakes workflows become especially vulnerable when automated summaries or recommendations are accepted without a human checking the original material and the model's reasoning.
The hosts compare hidden prompt instructions with early black-hat search tactics that concealed repeated keywords from human readers. They also connect the problem to fabricated legal citationsA hallucination is an AI output that presents false, unsupported, or invented information as though it were reliable., weak oversight and the need for systems that surface suspicious instructions instead of silently following them.
A separate model-news section reviews reported Gemini 3.7 Flash benchmark and pricing claims, framing the release as part of a trend toward models optimized for narrower workloads. The episode does not independently test those claims and treats paper benchmarksA benchmark is a defined test or collection of tasks used to compare a system's performance under specified conditions. as incomplete evidence of practical quality.
Watch on YouTube
