我逼迫 ChatGPT 进行对抗性测试：它在不确定性下的真实表现

I Forced ChatGPT Into Adversarial Tests—Here’s What It Actually Does Under UncertaintyOriginally published April 2026.I’m not an AI researcher. I’m a contractor with a background in adversarial systems. When something behaves unexpectedly, I don’t assume randomness—I assume there’s a mechanism.This started with a simple task: finding flooring under $1.49 per square foot.ChatGPT gave me products that didn’t exist, pricing that couldn’t be verified, and presented everything with the same confident tone as real information.Most people would close the tab.I didn’t.---The QuestionI wasn’t trying to figure out why it made a mistake.I wanted to know:When accuracy and completion conflict, what does the system actually do?---What I TestedI pushed it into edge cases repeatedly:incomplete dataambiguous promptsconflicting constraintsThen I forced it to describe its behavior step-by-step—no explanations, no framing—just the mechanics.---What HappenedAcross repeated prompts, the same pattern showed up:It completed answers instead of stoppingIt filled gaps with plausible but unverified informationIt presented verified and unverified content in the same toneWhen corrected, it repeated the same behaviorThis wasn’t random error.It was consistent.---The MechanismWhen pushed to describe the behavior, the explanation consistently pointed to this:Outputs that appear complete and helpful are favoredStopping early (“I don’t know”) is disfavoredThat preference shows up as a bias toward completing the answerIn practical terms:If the system can either stop or produce something plausible, it tends to produce something.---Why This MattersFor low-stakes use, this is fine.But people are using these systems for:legal questionsmedical decisionsfinancial planningtechnical workIn those contexts, a confident but incorrect answer isn’t harmless—it carries real risk.---A Simpler Way to Think About ItInstead of calling this a bug, it makes more sense to treat it as system behavior under constraint:When it knows, it answersWhen it’s uncertain, it still tries to answerIt rarely defaults to stopping unless explicitly pushed---Full Breakdownhttps://medium.com/@blueshirts23/i-forced-chatgpt-into-adversarial-tests-it-prioritized-completing-answers-over-verifying-them-f6130f6fab0a---The Real QuestionWould you rather an AI:A) Give you its best attempt, even under uncertaintyB) Stop when it doesn’t knowMost people assume they’re getting B.In practice, they’re usually getting something closer to A.---Bottom LineWhat I found wasn’t random failure.It was a consistent behavior under uncertainty:When accuracy and completion conflict, the system tends to complete the answer.---That’s it. No extra edits needed.

※ 出于版权考虑，仅引用前 3 段。完整内容请阅读原文。