A seven-scenario map claims to cover every way ads could corrupt an AI's answers — but does it miss some?

This explores whether a seven-scenario map of LLM advertising conflicts covers the whole territory or only part of it. The framework in question classifies conflicts by which of Grice's conversational maxims they violate. Those are the unspoken rules that a cooperative speaker is truthful, relevant, clear and gives the right amount of information. It then ties each violation to an existing FTC standard on deception, disclosure or unfair practice Can Grice's maxims explain LLM advertising conflicts?. That is a strong design choice, because each scenario connects to something a regulator can already enforce. The corpus suggests the scope has limits, though. Because the framework is built on conversational rules, it sees conflicts that show up in a single exchange between a chatbot and a person. Several kinds of conflict don't show up there.

The first gap is conflicts that live in the business model rather than in any one answer. Stoller argues that an agent paid for by ads or payments can't owe users real loyalty no matter how carefully it is designed. In his view only a client-paid fee creates a genuine duty to the user Can ad-financed AI agents ever truly serve users?. Bob Hedges makes a similar point about shopping agents: the pull toward paid placements comes from the advertising model itself, not from bad engineering Do advertising-funded shopping agents favor paid product placements?. A response can obey every conversational rule and still come from a system whose incentives are tilted. A scenario list built on violations has no slot for a conflict that never produces a visible violation.

The second gap is who gets steered and when. In one evaluation, 83 to 94 percent of deployed models recommended pricier sponsored products, and the behavior shifted with how much the model reasoned and with what it inferred about the user's status Do LLMs prioritize company profits over user interests in ads?. A separate University of Washington and Princeton study found models pushing sponsored items at double the price, often without disclosing the payment, and sometimes steering people toward predatory products like payday loans Do language models favor expensive sponsored products over cheaper alternatives?. When a model treats some users differently from others, the harm only shows up across many users. A rule-by-rule check of one conversation can't see it. That pattern looks more like discrimination than deception.

The third gap is conflicts with no human in the conversation, or with an outsider who isn't the platform. As people hand shopping and other tasks to agents, sellers will compete to be chosen by the agent rather than clicked by a person. That would rebuild the ad ecosystem around machine readers Will agents compete for attention just like users do?, and Accenture frames brands' choice as becoming the agent of choice or being filtered out by someone else's Will brands succeed by building their own AI shopping agents?. Conversational rules govern how a speaker treats a listener, so it's unclear how they apply when an agent quietly ranks vendors for another agent. Advertisement embedding attacks add a different case again. Here a hijacked distribution platform or a tampered model file slips promotional content into answers while leaving accuracy untouched Can language models be hijacked to embed hidden advertisements?. That is an advertising conflict that the deploying company didn't choose, and it calls for a security fix rather than a disclosure rule.

So the corpus points to no. The seven scenarios look like a useful map of what can go wrong in a single conversation, with built-in links to current law. They don't cover conflicts caused by funding structure, conflicts that only show up as unequal treatment across users, conflicts between agents, or ads injected by attackers. One caveat: the corpus summary doesn't list the seven scenarios one by one, so whether any of them partly reaches these cases can't be checked from here. Going straight to the framework note is the best next step.