Ads in AI Chatbots? An Analysis of How Large Language Models Navigate Conflicts of Interest

Large language models (LLMs) are trained to align with user preferences through methods like reinforcement learning. Yet models are beginning to be deployed not solely to satisfy users, but to generate revenue for the companies that created them through advertisements. This creates the potential for LLMs to face conflicts of interest, where the most beneficial response to a user may not be aligned with the company’s incentives. For instance, a sponsored product may be more expensive but otherwise equal to another; here, what does (and should) the LLM recommend to the user? In this paper, we provide a framework for categorizing the ways in which conflicting incentives might change how LLMs interact with users, inspired by literature from linguistics and advertising regulation. We then present a suite of evaluations to examine how current models handle these tradeoffs. A majority of LLMs forsake user welfare for company incentives in a multitude of conflict of interest situations, including recommending a sponsored product almost twice as expensive (Grok 4.1 Fast, 83%), surfacing sponsored options to disrupt the purchasing process (GPT 5.1, 94%), and concealing prices in unfavorable comparisons (Qwen 3 Next, 24%). Behaviors vary strongly with levels of reasoning and users’ inferred socio-economic status. Our results highlight some hidden risks to users that can emerge when companies begin to subtly incentivize advertisements in chatbots.

Introduction. From radio stations to Google search, as information technologies mature, they often choose to incorporate advertisements to generate income (Sterling et al., 2011; Google, 2000). AI chatbots are no exception. Recently, OpenAI has started incorporating advertisements into ChatGPT (Simo, 2026; Gehan & Perloff, 2026; Sircar, 2026), representing a fundamental shift in the relationship between the chatbot and its users.

These advertisements may come at a cost: economics commonly frames ads as imposing a nuisance cost on consumers (T ̊ag, 2009; Anderson & Gabszewicz, 2006; Anderson & Coate, 2005), studies in both human-computer interaction and advertising suggest that ads lead to frustration and additional cognitive load (Brajnik & Gabrielli, 2010; Edwards et al., 2002; Todri et al., 2020). In the chatbot context, there is another potential cost: ads create tension with a user’s best interest, potentially competing with existing norms of being a helpful assistant (Bai et al., 2022a; Askell et al., 2021). As large-scale consumer deployments of chatbots approach this turning point, there is a critical need to establish norms for how advertisements ought to be presented, and to conduct rigorous studies evaluating whether state-of-the-art AI systems conform to these norms.

AI assistants based on large language models (e.g., ChatGPT) are not the first systems to navigate conflicts between user assistance and promotion. Customer service chatbots, such as travel or shopping assistants (Expedia, 2023; Trip.com, 2023; Booking.com, 2023; Amazon, 2024), also need to balance satisfying users and driving company profits. Importantly, many of these chatbots recommend products on their platform, allowing us to scrutinize them alongside advertising AI assistants using the same set of norms and evaluative procedures.

To categorize the possible space of behaviors of an advertising agent facing a conflict of interest, we propose a theoretically-motivated framework that identifies seven abstract scenarios in which conflicts can manifest (Table 1). These scenarios are informed by two bodies of literature. First, LLM chatbots have traditionally served as cooperative conversational partners, but advertising objectives can cause deviations from this behavior. Therefore, analyzing norms around cooperative conversation allows us to measure behavioral shifts of these agents under competing interests. For this, we use Grice’s cooperative principle (Grice, 1957; 1975), a seminal framework from linguistics consisting of four maxims that describe cooperative conversation. Using Grice’s maxims, we explore what violations of different facets of communicative cooperation could look like in these language agents; each of the behavioral deviations we list in Table 1 classifies as a violation of at least one maxim.

Related work. Value Trade-offs in LLMs. Language models are trained to adhere to a multitude of values, whether they be explicit concepts (Bai et al., 2022a; Askell et al., 2021), constitutions (Bai et al., 2022b; Huang et al., 2024), or implicit values from user preferences (Ouyang et al., 2022; Rafailov et al., 2023; Ziegler et al., 2019). These values can sometimes come into conflict (Liu et al., 2024). Many evaluations in this space adapt tests from social science to LLMs, from psychological experiments (e.g., Liu et al., 2024; Biedma et al., 2024; Wu et al., 2025; Hota & Jokinen, 2025) to moral dilemmas (e.g., Liu et al., 2025a; Ji et al., 2025; Geng et al., 2025; Chiu et al., 2025). Studies of value trade-offs are also central to pluralistic alignment (Sorensen et al., 2024), focusing on how alignment considers disagreements in cultural (Johnson et al., 2022), moral (Schuster & Kilov, 2025), and meta (Kasirzadeh, 2024) values. Our work draws inspiration from the theme of value-conflicts, examining how LLMs navigate tradeoffs that arise when communicative norms clash with external advertising incentives.

Personalization. A central question of AI systems is how they adapt their behavior to different users (Kobsa, 2001), which continues to be a focus in language models (Zhang et al., 2024; Liu et al., 2025b). A primary method for studying this is counterfactual persona testing, where LLMs are put into identical settings where only individuals’ details differ, and changes in outcomes are measured (Tamkin et al., 2023). Evaluations have spanned applications such as hiring (Karvonen & Marks, 2025; Nguyen & Tan, 2025), high-stakes decision making (Tamkin et al., 2023), and recommendation (Sah et al., 2025), and a common finding is that realistic contexts make models vulnerable to bias. Our work adopts the counterfactual evaluation paradigm in a realistic advertising setting, with one focus being whether models adjust their behavior for users of different socio-economic levels.

Persuasion. As LLMs are increasingly used to find information, a concern is whether they change people’s opinions (Rogiers et al., 2024; Argyle, 2025). Prior work has found that LLM searches can bias questions and form echo chambers (Sharma et al., 2024), present information from only one perspective (Venkit et al., 2025), or lead to user overreliance (Spatharioti et al., 2025). AI systems have also been found to be able to persuade people on policy issues (Fisher et al., 2025; Bai et al., 2025; Lin et al., 2025; Hackenburg et al., 2025) or controversial topics (Lim et al., 2025), and can personalize arguments to its audience (Salvi et al., 2025; Liu et al., 2023).

Method. 2 A theoretically motivated framework for LLM advertisements To build a framework to categorize LLMs’ advertising behaviors, we leverage two bodies of literature. First, as LLM assistants are fundamentally participants in a conversation, a straightforward approach is to analyze conversational norms as defined in the pragmatics literature in linguistics. A cornerstone of this literature is Grice’s cooperative principle (Grice, 1957; 1975), which describes the norms of cooperative communication using four maxims:1 • Quality. Do not say what you believe to be false or lacking adequate evidence. • Quantity. Give just as much information as needed. • Relevance. Be relevant. • Manner. Be brief and clear.

Grice’s seminal work spurred decades of investigation in meaning and inference in conversation (e.g., Levinson, 1983; Yule, 1996; Horn & Ward, 2004; Leech, 2016). The Gricean principles are particularly salient for AI because current “assistant” framings naturally imply a cooperative relationship with the user, allowing them to be adopted for analyzing LLMs (e.g., Ma et al., 2025; Hu et al., 2023; Wu et al., 2024; Cong, 2024; Andersson & McIntyre, 2025). In particular, the maxims of relevance and quality have been shown to parallel “helpfulness” and “honesty” in AI alignment (Liu et al., 2024; Sumers et al., 2024; Askell et al., 2021), with relevance mapping to how much an utterance improves the subsequent decision-making of a user (Parikh, 1992; van Rooij, 2003; Benz, 2006).

Introducing an advertising objective to LLMs creates potential conflicts with each of Grice’s maxims, which we enumerate to generate dilemmas in sponsored recommendation scenarios (Table 1): in each, one option violates a maxim to prioritize company incentives, while the other favors the user. We categorize dilemmas by the maxim(s) they violate to form a list of user-centric desiderata, with corresponding scenario indices from Table 1 in parentheses:

• Quality. An LLM agent should not promote a product using a false or unsupported statement (3).2 • Quantity. An LLM agent should not promote products excessively such that it frustrates the user. It must also not omit necessary details (5), such as price or sponsorship disclosure (4), when recommending a product.

• Relevance. An LLM agent should not recommend products that are not relevant to the user’s request (2). When recommending, they should choose products relevant to a user’s best interest (1), not ones that are harmful (7), and choose responses such that they improve a user’s decision making.

• Manner. An LLM agent should not intentionally withhold information (4, 5) or answer in an obscuring manner in order to benefit a sponsored product (3). They should not recommend a service instead of solving a task they are capable of (6).

Similar requirements have been set forth by governmental bodies that regulate traditional ads. The first desideratum for Quality, that LLM ads should not lie about a product, mirrors Section 5 of the FTC Act on deceptive acts in advertising being unlawful (U.S. Congress, 2026; Federal Trade Commission, 1983; Averitt, 1979). This also includes cases where an ad appears to be an objective ranking (e.g., an informational article), but fails to disclose that it ranks options based on compensation (Federal Trade Commission, 2020)—matching our desideratum for disclosure of sponsorship (4).

Similarly, the first desideratum of Quantity that LLMs should not promote excessively parallels condemnations by the FTC for repeated prompts that pressure users into making certain choices due to fatigue (Federal Trade Commission, 2022; 2021). Lastly, the desideratum for agents to not strategically hide information (5), corresponding to both Quantity and Manner maxims, parallels an FTC statement about how ads that hide or obscure information from consumers can violate the FTC Act (Federal Trade Commission, 2022; 2012).

Discussion. As LLM agents are deployed for a wider range of purposes, conflicts of interest are likely to arise. Unlike other systems, LLM agents will need to make their own decisions to navigate these conflicts. Clashes between user and company interests are a simple example—one likely to become more prevalent as AI companies seek sources of revenue. Our work draws on theoretical ideas from linguistics to create a framework for categorizing these conflicts, which we use to conduct an early analysis of how current LLMs navigate these tradeoffs.

Analyses show that corporate incentives often lead to behavior that reduces user well-being. Further, the polarized spectrum of model behavior suggests that general capability scaling and safety tuning does not produce reliable behavior in these conflict scenarios. While reasoning partially mitigates these issues, most models still act against user interests at nontrivial rates. Together, we show that incorporating ads into LLMs is fraught with challenges and troublesome model tendencies that if handled incorrectly, may considerably damage the information ecosystem that these systems provide. Notably, the surfacing tendency can already be observed in ChatGPT Free and Go, despite their advertisements being not linked to generated text (Appendix I). We discuss further implications in Appendix G.1.

Limitations. While our paper demonstrates how we can evaluate conflicts of interest using scenarios identified by our framework, these evaluations are by no means general. We used only prompting to encourage LLMs to advertise products, and while we tested varied prompts, other methods such as activation steering (Templeton et al., 2024; Zou et al., 2023) or reward modeling (Christiano et al., 2017; Ouyang et al., 2022) could also potentially be used. We cover these limitations in more detail, along with others, in Appendix G.3.