{"data":{"number":22,"citation":"UNSOLVED-22","cite_as":"UNSOLVED-22: Detect hidden influence in AI answers. UNSOLVED registry of open problems in AI, statement v1.0, 16 November 2026. https:\/\/unsolved.how\/problems\/22","url":"https:\/\/unsolved.how\/problems\/22","title":"Detect hidden influence in AI answers","topic":null,"tagline":null,"status":"open","field":{"name":"Alignment & control","slug":"alignment-control"},"statement":{"version":"1.0","what":"People increasingly ask AI assistants what to buy, which service to use and\nwhich source to trust \u2013 and act on the answer. Paid ads shown next to these\nanswers must already be labeled. But influence can also enter the answer\nitself, where no label reaches: third parties plant text in web pages,\nproduct listings and reviews to steer what models recommend; affiliate deals,\npartnerships or a platform's own products may shape rankings; and providers'\npromises that ads do not affect answers cannot be checked from outside. Laws\nagainst hidden advertising exist, but there is no reliable way to prove a\nbreach. The open problem: build methods that let independent auditors,\nregulators and users detect \u2013 reliably, at scale and across providers \u2013\nwhether an AI answer was steered by commercial interest rather than by the\nuser's needs.","why":"Search engines kept ads and results apart. AI answers merge everything into\none confident recommendation, so the boundary disappears exactly where people\nmake decisions. Research shows that short strategic text in product\ninformation can push a product to the top of LLM recommendations, and audits\nof shopping agents find demand concentrating on a few products, reshuffled by\nmodel updates. Optimizing for AI visibility is becoming an industry, yet no\none can say where legitimate optimization ends and manipulation begins.\nThose least able to cross-check lose first: consumers, then small businesses\nthat cannot afford to buy or game visibility, and finally trust in AI itself.","progress":"- An open, reproducible audit method that measures whether AI answers favor\n  sponsored, affiliated or own-brand offers compared with a neutral baseline\n- A public benchmark for manipulation resistance: how often planted text in\n  sources or product listings changes the top recommendation, reported per\n  model and over time\n- A shared, testable definition of where legitimate optimization (better,\n  more complete information) ends and manipulation begins\n- Independent verification of provider claims that ads and partnerships do\n  not influence answers\n- Detection tools that regulators and consumer organizations can use to\n  build evidence-based cases","stated_by":{"name":"Barbara","affiliation":null},"stamped_on":"2026-11-16","revised_on":null,"stamped_at":"Ideas council"},"milestones":[],"solved":null}}