97% of AI safety papers using OpenRouter skip safety measure for results

OpenRouter is a third-party API broker that forwards model requests to multiple backend providers. When a researcher sends a request through OpenRouter without explicitly pinning a provider, the system routes that request to a random available provider. Because these providers operate with different infrastructure, update schedules, and quality levels, the same request can produce different outputs depending on which provider handles it on a given day. A comprehensive review of 32 influential AI safety research codebases that report results obtained through OpenRouter found that 31 of them (97%) do not pin their provider, leaving their reported results vulnerable to this variability. The stakes of this practice are high: the article cites precedent for an AI safety paper that was accepted to NeurIPS only to have its core results entirely overturned when provider-quality issues were identified. To protect research integrity, the article recommends that researchers using OpenRouter or similar third-party providers take precautions to ensure they pin to a specific, vetted provider rather than accepting random routing. However, the article notes that the current selection of providers is insufficient for ideal scientific reliability, meaning that pinning alone may not fully resolve the underlying risk. Replicators of published results are advised to test whether findings hold up once identified provider bugs are fixed, suggesting that some existing results may not replicate until this gap is addressed.

The review identifies a widespread methodological vulnerability in AI safety research: 31 of 32 influential codebases that rely on OpenRouter for their reported results fail to pin their provider choice. This matters because OpenRouter acts as a broker, routing each request to one of several backend providers without determinism. Since these providers exhibit variable quality, a research result obtained one day against Provider A may not replicate against Provider B, even if the model and prompt are nominally identical. The article notes that this risk is not theoretical — a prior NeurIPS-accepted paper in AI safety saw its core findings completely overturned when provider-quality issues were identified. The widespread nature of the gap (97% of codebases) suggests that many researchers may not be aware of the hazard or may underestimate its severity. The article acknowledges that while precautions are available, the current selection of providers is insufficient for ideal scientific reliability, indicating that even pinning a provider may not fully solve the underlying problem.

No comments yet. Be the first to share your thoughts!