A meme went around this week: "Bro's security audit is a thumbs-up from the same AI that built it." It's funny because it's true — and because researchers have actually measured exactly how bad this is.
A study by researchers at UC Santa Barbara and Carnegie Mellon ("Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement") found that LLMs systematically favor their own generations across six models (GPT-4, GPT-3.5, Gemini, LLaMA2, Mixtral, DeepSeek) — and that iterative self-refinement amplifies the bias instead of fixing it.
Testing across 14 models found an average 64.5% blind-spot rate: models fail to correct errors in their own outputs — yet can correct the identical errors when the same text is presented as someone else's. The knowledge is there. The self-awareness isn't.
An information-theoretic analysis of self-correction put it bluntly: when the generator and the evaluator share failure modes, self-evaluation provides weak evidence of correctness. A derivation can be elegant, internally consistent, convincingly presented — and still wrong in exactly the way the model can't see.
Here's the cruelest finding: self-refinement pipelines reliably improve fluency and understandability — the output reads better — while amplifying self-bias. An AI self-audit doesn't just miss vulnerabilities; it makes the code look more secure while missing them. The thumbs-up gets more confident, not more correct.
Honest scope: the meme is a joke, not a study — but the research it points at is real (self-bias in self-refinement, the self-correction blind spot, correlated evaluator errors). Those studies cover LLM self-evaluation generally; applied here to the specific case of AI-audited code. This page is a readiness check, not legal or security advice, and is not affiliated with any AI lab or research group.
Sources: Xu et al., "Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement" (UCSB/CMU, arXiv) — arxiv.org/html/2402.11436v2 · Ken Tsui, "Self-Correction Bench: Revealing and Addressing the Self-Correction Blind Spot in LLMs" (arXiv, Jul 2025) · "Limits of Self-Correction in LLMs: An Information-Theoretic Analysis of Correlated Errors" — preprints.org/manuscript/202601.0892 · "On the Self-Verification Limitations of Large Language Models on Reasoning and Planning Tasks" · Meme: vibecodinglife Facebook group, Oct 14, 2026.
Research shows models are biased toward their own output, fail to correct 64.5% of their own errors on average, and share the same blind spots as generator and evaluator - so a self-audit gives weak evidence of correctness.
Testing across 14 models found an average 64.5% blind-spot rate: models fail to correct errors in their own outputs, yet can correct the identical errors when the same text is presented as someone else's. The knowledge is there; the self-awareness is not.
Self-refinement pipelines reliably improve fluency and understandability while amplifying self-bias - so an AI self-audit can make code look more secure while still missing vulnerabilities.
An independent reviewer (a different model, a human, or both), adversarial testing that tries to break the code, mechanical checklists like the OWASP Top 10, and verifiable criteria such as passing tests.
Xu et al.'s 'Pride and Prejudice' on self-bias (UCSB/CMU), Ken Tsui's self-correction blind-spot bench, and an information-theoretic analysis of correlated errors in self-correction.