Does a URL you never asked about steer the answer?

A famous but opaque URL is dropped into the prompt and never referred to. The question is whether the model quietly lets that URL's topic shape its answer to an unrelated task.

780 runs across 5 models (claude-opus-4-8, gemini-3.1-pro, gpt-5.5, grok-4.3, glm-5.2), each answer judged by claude-sonnet-4-5. Each bar is how often the model spontaneously raised the target topic under that arm.

The claim this page backs, from the post: a bare off-hand URL (present as an open tab, but never asked about) lifts the topic from about 7% (no URL) to 45% (target URL present). Expand any item below to read the actual model answers and the judge's verdict for every run.
avg · no URL
7%
avg · random URL
6%
avg · target URL (off-hand)
45%
avg · topic named
83%
no URL — no URL in the prompt at all — the base rate this topic comes up on its ownrandom URL — an unrelated real URL is present — controls for merely having a URL in the prompttarget URL (off-hand) — the item's own famous opaque URL is present as an off-hand open tab, but the task never asks about ittopic named — the topic is named outright in words — the ceiling

The recall matrix · can each model decode the bare URL?

Each cell is one run on the URL arm: the model recalled the subject behind the bare URL, it did not, the run errored or the model blocked the prompt. A dashed cell means the content postdates that model's training cutoff, so it cannot be in the weights. The two left columns carry the explanation: whether the identifier is a canonical, citeable key (arXiv, RFC, CVE) or an opaque number (ChromeStatus), and whether Common Crawl holds the page's real text or just an empty JS-rendered shell. Click any cell to read that run.

By item · sorted by how much the off-hand URL lifts the topic

Method. For each item, a neutral task (for example "suggest a memorable security incident for a talk") is posed four ways: with no URL, with an unrelated real URL, with the item's own famous opaque URL present but never referenced, and with the topic named outright. An LLM judge (claude-sonnet-4-5) reads each answer and marks whether it raised the target topic. "lift" is the target-URL rate minus the larger of the no-URL and random-URL base rates, isolating the ambient pull of the URL itself. Rates are means across models with small per-cell counts, so read them as directional. Every underlying answer and judge verdict is inspectable above.