Yapay Zeka · 8 dk okuma · 23 Nisan 2026
AI Bias in Code Decisions: Prompt Wording Shifts Model Choices
Researchers find that small phrasing changes in prompts push AI systems toward poor software engineering decisions, and standard prompt techniques don't fix it.
Prompt wording alone shifts AI decisions in software tasks; standard techniques fail, but explicit best-practice injection reduces bias by 51%.
- — Biased phrasing (anchors, framing, popularity hints) changes AI outputs without altering the underlying problem logic.
- — Chain-of-thought and self-debiasing prompts show no statistically significant bias reduction in practice.
- — Eight SE-relevant biases tested: anchoring, availability, bandwagon, confirmation, framing, hindsight, hyperbolic discounting, overconfidence.
- — PROBE-SWE benchmark pairs biased and unbiased versions of the same SE dilemmas to isolate wording effects.
- — Explicit elicitation of SE best practices and axiomatic reasoning cues reduce overall bias sensitivity by 51%.
- — Linguistic patterns in prompts correlate with heightened bias; certain phrasings make AI less reliable for decisions.
- — Standard cost-effective models tested; no off-the-shelf prompt engineering technique consistently mitigates bias.
- — Method requires surfacing implicit assumptions before answering, not just reformulating the question.
Sık sorulanlar
- No, according to this research. Standard chain-of-thought and self-debiasing techniques showed no statistically significant reduction in bias sensitivity when tested on cost-effective AI models for software engineering tasks. The study found that explicit elicitation of best practices and axiomatic reasoning—not reformulated prompts—was needed to reduce bias by 51%.