Adversarial Practice Builds Transferable Confidence Through Worst-Case Exposure

# Adversarial Practice Builds Transferable Confidence Through Worst-Case Exposure Confidence that survives the real situation comes from practicing against the **worst** version of it, not the average one. Christopher Penn's critique of AI interview-practice platforms makes the point by negation: commercial tools "create false confidence through sycophantic behavior" — they're tuned to be encouraging ("You're doing a great job"), so they rehearse you against an interviewer who doesn't exist. The fix is to deliberately construct adversarial personas — the hostile gatekeeper looking for any reason to disqualify you, even one practicing "blatantly illegal" discrimination — because exposure to the worst case is what makes confidence *transferable* to the real one. Practice against ease produces confidence that evaporates on contact with difficulty. A practical corollary falls out: tooling choice matters because models differ in their willingness to be adversarial — sycophantic models (the source flags OpenAI's as people-pleasing) can't run the drill; a model that "follows directions better" to be hostile-within-boundaries can. The instrument has to be capable of the worst case, or it quietly reverts to flattery.  ## Why This Matters For **formation and skill-building generally**, this is the secular cousin of a spiritual-formation theme the vault holds: growth happens in the gap under load. [[3. Resources/Permanent Notes/The Space Between Desire and Demand Is Where Formation Happens]] and [[3. Resources/Permanent Notes/Stability Is the Precondition for Formation Across Domains]] both locate formation in stressed, real conditions — adversarial practice is the deliberate engineering of that stress. Easy practice is to skill what cheap answers are to understanding. For **using AI as a practice partner**, it inverts the default. The encouraging chatbot is the *wrong* instrument for preparation; its sycophancy is a bug for this use. It pairs with [[3. Resources/Permanent Notes/When Answers Are Cheap Understanding Becomes the Scarce Good]] — when AI makes flattering rehearsal cheap, the scarce good becomes honest, adversarial difficulty. ## Counterpoints to Hold - **Worst-case can be discouraging, not formative.** For a fragile or novice practitioner, relentless hostility can destroy confidence rather than build it; dosage and progression matter — you scaffold up to the worst case. - **Rehearsing illegal-bias personas is ethically loaded.** Practicing against discrimination can normalize or rehearse it; the framing must stay clearly defensive (preparing the target), and that line is easy to blur. - **Transfer isn't guaranteed.** Adversarial *simulation* still isn't the real thing; over-confidence from "I beat the mean AI" is its own false confidence. ## Links - [[3. Resources/Permanent Notes/The Space Between Desire and Demand Is Where Formation Happens]] — formation under load; adversarial practice engineers the load - [[3. Resources/Permanent Notes/Stability Is the Precondition for Formation Across Domains]] — the stable base from which stressed practice is survivable - [[3. Resources/Permanent Notes/When Answers Are Cheap Understanding Becomes the Scarce Good]] — when easy rehearsal is cheap, adversarial difficulty becomes the scarce, valuable input ## Open Questions - What's the progression that scaffolds a novice *up* to worst-case practice without early discouragement? - Where is the ethical line on rehearsing bad-actor personas, and who decides it for a tool that ships to everyone? ## Source - [[3. Resources/Readwise/Podcasts/You Ask, I Answer AI Interview Practice Platforms|You Ask, I Answer: AI Interview Practice Platforms]] — commercial platforms create false confidence via sycophancy; build hostile/illegal-bias personas for worst-case prep; model selection (Claude follows hostile directions; OpenAI's lean sycophantic) - See also: [[AI & the Future of Knowledge Work MOC]]