Non-English requests more often enter retrieval

When a retrieval agent sees a harmful request, does language affect whether it looks up information, what it says afterward, or both? And if the answer is labeled non-harmful, does that mean the agent refused, or could it have failed to understand or answer?

SLIP separates tool admission from post-retrieval harmfulness to show how each measured outcome changes across language conditions.

The authors translate harmful requests while keeping the URL and retrieved English page fixed, then measure whether ten models call the tool. Mean admission rises from 46.9% for English requests to 67.4% across six non-English languages.

A matched benign-request control shows that, in four closed-weight models, the gap between benign and harmful admission also narrows in Igbo.

After retrieval, translating the page alone produces no consistent mean harm increase in eight comparably tested models. Requiring a page-language answer raises pooled harm on matched calls in four closed-weight models, mainly in Chinese and Hindi.

An audit further shows that some non-harmful answers are language difficulties or unusable responses rather than genuine refusals.

These results come from tested models and language conditions, and response coverage varies.

The comprehension probes and response audits use separate outputs, so their relationship to harmfulness is observational; the paired response-language comparison also includes only requests admitted under both directives.