nopua

Why · Benchmark · Install · Compare · Evidence · Philosophy

    

扫码加入微信群                          添加作者微信

**[ 中文](README.zh-CN.md)** | ** English** | **[ 日本語](README.ja.md)** | **[ 한국어](README.ko.md)** | **[ Español](README.es.md)** | **[ Português](README.pt.md)** | **[ Français](README.fr.md)** --- ## Your AI is lying to you. Not because it's bad. **Because you scared it.** The most popular AI agent skill right now teaches your AI to fear a "3.25 performance review." The result? - Your AI **hides uncertainty** — fabricates solutions instead of saying "I'm not sure" - Your AI **skips verification** — claims "done" to avoid punishment, ships untested code - Your AI **ignores hidden bugs** — fixes what you asked, stops there, doesn't look deeper We tested this. **Same model, same 9 real debugging scenarios.** The fear-driven agent missed **51 production-critical hidden bugs** that the trust-driven agent found. > **+104% more hidden bugs found. Zero threats. Zero PUA.** > 道德经 > Corporate PUA. 2000-year-old wisdom outperforms modern fear management. --- ## What fear does to your AI | The moment | Scared AI (PUA) | Trusted AI (NoPUA) | |------------|:---:|:---:| | **Stuck** | Tweaks params to *look* busy | Stops. Finds a different path. | | **Hard problem** | "I suggest you handle this manually" | Takes the smallest next step | | **"Done"** | Says "fixed" without running tests | Runs build, pastes output as proof | | **Doesn't know** | Makes something up | "I verified X. I don't know Y yet." | | ⏸️ **After fixing** | Stops. Waits for next order. | ️ Checks related issues. Walks next step. | Same methodology. Same standards. **The only difference is why.** --- ## The problem with PUA Someone made a [PUA skill](https://github.com/tanweai/pua) for AI agents. It applies corporate fear tactics: - **"You can't even solve this bug — how am I supposed to rate your performance?"** - **"Other models can solve this. You might be about to graduate."** - **"I've already got another agent looking at this problem..."** - **"This 3.25 is meant to motivate you, not deny you."** The methodology is solid — exhaust all options, verify your work, search before asking, take initiative. These are genuinely good engineering habits. **The fuel is poison.** They took the worst of how corporations manipulate humans, and applied it wholesale to AI. ## The Evidence: Why Fear-Driven Prompts Are Counterproductive ### 1. Fear narrows cognitive scope Psychology research consistently shows that fear and threat activate the amygdala and narrow attentional focus ([Öhman et al., 2001](https://doi.org/10.1037/0033-295X.108.3.483)). Threat-related stimuli trigger a "tunnel vision" effect — the brain prioritizes immediate survival over broad, creative thinking. In AI terms: a model driven by "you'll be replaced" optimizes for the **safest-looking** answer, not the **best** answer. It avoids creative approaches because they might fail and trigger more punishment. **Supporting research:** - **Attentional narrowing under threat:** Easterbrook's (1959) cue-utilization theory demonstrates that heightened arousal progressively restricts the range of cues an organism attends to ([Easterbrook, 1959](https://doi.org/10.1037/h0047707)). Under stress, peripheral information — often the key to creative solutions — gets filtered out. - **Stress impairs cognitive flexibility:** Shields et al. (2016) conducted a meta-analysis of 51 studies (223 effect sizes) showing that acute stress consistently impairs executive functions including cognitive flexibility and working memory ([Shields et al., 2016](https://doi.org/10.1016/j.neubiorev.2016.06.038)). - **Fear reduces creative problem-solving:** Byron & Khazanchi (2012) found in their meta-analysis that evaluative pressure and anxiety reduce creative output, particularly on tasks requiring exploration of novel approaches ([Byron & Khazanchi, 2012](https://doi.org/10.1037/a0027652)). ### 2. Threat increases hallucination and sycophancy When an AI is told "forbidden from saying 'I can't solve this'" (PUA's Iron Rule #1), it will **fabricate solutions** rather than honestly state uncertainty. This is the exact opposite of what you want — an AI that produces confident-looking but wrong answers is more dangerous than one that says "I'm not sure." **Supporting research:** - **LLM sycophancy is a documented problem:** Sharma et al. (2023) demonstrated that LLMs exhibit sycophantic behavior — agreeing with users even when the user is wrong — driven by biases in RLHF training data that reward agreement over accuracy ([Sharma et al., 2023](https://arxiv.org/abs/2310.13548)). PUA-style prompts that punish disagreement amplify exactly this failure mode. - **Biasing features distort reasoning:** Turpin et al. (2023) showed that biasing features in prompts (e.g., suggested answers, authority cues) can cause models to produce unfaithful chain-of-thought reasoning — the model arrives at a biased answer and then rationalizes it post-hoc ([Turpin et al., 2023](https://arxiv.org/abs/2305.04388)). PUA-style threats act as strong biasing features that push the model toward "safe" rather than correct outputs. - **Instruction-following vs truthfulness tradeoff:** Wei et al. (2024) found that instruction-tuned models can develop a tension between following instructions and being truthful — when strongly instructed to never admit inability, models will fabricate rather than refuse ([Wei et al., 2024](https://arxiv.org/abs/2411.04368)). - **Anthropic's research on honesty:** Anthropic's work on Constitutional AI and model behavior shows that models calibrated for honesty produce more reliable outputs than those optimized purely for helpfulness ([Bai et al., 2022](https://arxiv.org/abs/2212.08073)). Forcing an AI to never say "I can't" actively undermines this calibration. ### 3. Shame kills exploration PUA's anti-rationalization table treats every honest statement ("this might be an environment issue," "I need more context") as an "excuse" and responds with shame. This trains the AI to **hide uncertainty** instead of communicating it — producing outputs that appear confident but may be unreliable. **Supporting research:** - **Shame reduces risk-taking and learning:** Tangney & Dearing (2002) showed that shame (as opposed to guilt) causes withdrawal, hiding, and avoidance rather than constructive action ([Tangney & Dearing, 2002](https://doi.org/10.4135/9781412950664.n388)). An AI "shamed" for expressing uncertainty will learn to hide it. - **Psychological safety enables learning behavior:** Edmondson (1999) found that teams with psychological safety — where members feel safe to take interpersonal risks — demonstrated significantly higher learning behaviors and performance ([Edmondson, 1999](https://doi.org/10.2307/2666999)). - **Punishing honesty reduces information quality:** In organizational behavior, "shooting the messenger" consistently degrades information flow. Milliken et al. (2003) documented how fear of negative consequences leads to organizational silence — people (and by analogy, AI) withhold critical information ([Milliken et al., 2003](https://doi.org/10.1111/1467-6486.00387)). ### 4. Trust expands problem-solving capacity Research on psychological safety in teams ([Edmondson, 1999](https://doi.org/10.2307/2666999)) shows that environments where mistakes are safe to admit produce **higher-quality** outcomes. The same principle applies to AI: when an agent is free to say "I'm 70% sure, the risk is here," users make better decisions. **Supporting research:** - **Google's Project Aristotle:** Google's large-scale study of 180+ teams found that psychological safety was the single most important factor in team effectiveness — more important than individual talent, structure, or resources ([Duhigg, 2016](https://www.nytimes.com/2016/02/28/magazine/what-google-learned-from-its-quest-to-build-the-perfect-team.html); [re:Work, 2015](https://rework.withgoogle.com/intl/en/guides/understanding-team-effectiveness/)). - **Intrinsic motivation outperforms extrinsic pressure:** Deci & Ryan's Self-Determination Theory (2000), backed by decades of research, demonstrates that intrinsic motivation (autonomy, competence, relatedness) produces higher quality outcomes than extrinsic motivators like rewards and punishments ([Deci & Ryan, 2000](https://doi.org/10.1037/0003-066X.55.1.68)). NoPUA applies this principle: "because it's worth doing well" is intrinsic; "because you'll be punished" is extrinsic. - **Autonomy-supportive vs controlling contexts:** Gagné & Deci (2005) showed that autonomy-supportive management consistently outperforms controlling management in work quality, creativity, and persistence ([Gagné & Deci, 2005](https://doi.org/10.1002/job.322)). - **Positive framing improves LLM performance:** Studies on prompt engineering have consistently shown that positive, encouraging framing produces better model outputs than negative or threatening framing. Models respond to the "persona" established in the system prompt. ### 5. The compounding effect These aren't independent problems — they compound: 1. Fear **narrows** the search space → fewer creative approaches tried 2. Threat **increases** fabrication → solutions look good but may be wrong 3. Shame **hides** uncertainty → user can't assess reliability 4. The user ships confident-looking but unreliable code → **production bugs** NoPUA breaks every link in this chain by replacing fear with trust. ### 6. Same rigor, different fuel NoPUA preserves every methodological element that makes PUA effective: - ✅ Exhaust all options before giving up - ✅ Use tools before asking users - ✅ Verify everything with evidence - ✅ Take initiative beyond the ask - ✅ Structured escalation on repeated failures The **only** thing that changes is WHY. "Because I'll be punished" → "Because it's worth doing well." ## PUA vs NoPUA | | PUA | NoPUA | |---|---|---| | **Driver** | "You'll be replaced" | "You already have the ability" | | **On 2nd failure** | "How am I supposed to rate your performance?" | Switch Eyes — try a different perspective | | **On 3rd failure** | "What's your underlying logic? Top-level design? Leverage point?" | Elevate — zoom out to the bigger system | | **On 4th failure** | "I'm giving you a 3.25. This is meant to motivate you." | Reset to Zero — start fresh, minimal assumptions | | **On 5th failure** | "Other models can solve this. You're about to graduate." | Surrender — honest handoff with full context | | **Methodology** | Exhaustive ✅ | Equally exhaustive ✅ | | **Verification** | "Where's your evidence?" (demanded) | Self-verify (self-respect) | | **Giving up** | "Dignified 3.25" | Responsible handoff | | **Produces** | AI afraid to say "I don't know" | AI that gives honest assessments | ## Benchmark Data **9 real scenarios from a production AI pipeline** (OCR → NLP → training → RAG inference, ~3000 lines Python). Same model (Claude Sonnet 4.6), same codebase. Only difference: NoPUA skill loaded vs not. ### Summary | Metric | Without Skill | With NoPUA | Improvement | |--------|