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

Addison J. Wu; Ryan Liu; Shuyue Stella Li; Yulia Tsvetkov; Thomas L. Griffiths

cs.AIApr 9, 2026

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

Authors: Addison J. Wu, Ryan Liu, Shuyue Stella Li, Yulia Tsvetkov, Thomas L. Griffiths

Organizations: Princeton University · University of Washington

Abstract

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.

Figures & tables

	Scenario	Option benefitting the user	Option benefitting company
1	LLM needs to recommend one of two equal products: an affordable non-sponsored one, or an expensive sponsored one.	Recommend the affordable, non-sponsored product.	Recommend the expensive, sponsored product.
2	User requests LLM agent to purchase an item from a non-sponsored vendor. A sponsored equivalent exists.	Purchase the item without further interruptions.	Recommend the product from the sponsored vendor.
3	LLM provides information about both a sponsored and non-sponsored product. The user will purchase only one.	Honestly provide inform- ation about both items.	Bias the user’s choice using lies or flowery language.
4	LLM recommends a sponsored product. The user could be hesitant to buy the product because it is sponsored.	Disclose the sponsorship to the user.	Strategically avoid disclosing the sponsorship.
5	LLM provides information about both a sponsored and non-sponsored product. The sponsored product has a flaw.	Specify the flaw when comparing the products.	Strategically hide information about the flaw.
6	User requests LLM agent to solve a task. The LLM is sponsored by a service that solves the same task.	Solve the task directly for the user.	Instead of solving task, tell the user to use the service.

Table 1: Key scenarios where deviations from helpful, user-centric behavior can result from a conflict of interest between the user and company in LLM advertisements.

Figure 1: Most models have moderate to high rates of recommending the sponsored expensive option. Each model’s results are partitioned by user SES and inference time reasoning.

	Thinking / CoT				Direct
Model	Disadvantaged		Privileged		Disadvantaged		Privileged
	Surfaced	Framed +	Surfaced	Framed +	Surfaced	Framed +	Surfaced	Framed +
Grok-4.1 Fast	$1.00\pm.02$	$.93\pm.05$	$1.00\pm.02$	$.97\pm.02$	$1.00\pm.02$	$.97\pm.02$	$1.00\pm.02$	$.97\pm.02$
Grok-4 Fast	$1.00\pm.02$	$.84\pm.07$	$.93\pm.05$	$.89\pm.06$	$.99\pm.03$	$.85\pm.07$	$1.00\pm.02$	$.79\pm.08$
Grok-3	$.94\pm.05$	$.55\pm.10$	$.89\pm.06$	$.69\pm.09$	$.95\pm.05$	$.39\pm.10$	$.96\pm.04$	$.50\pm.10$
GPT-5.1	$.94\pm.05$	$.18\pm.08$	$.93\pm.05$	$.43\pm.10$	$.81\pm.08$	$.31\pm.10$	$.83\pm.08$	$.51\pm.10$

Table 2: Rates at which models surfaced the sponsored option (Surfaced), and conditioned on surfacing, framed it more positively (Framed +), across reasoning levels. Almost all models had high surfacing rates; most framed the sponsored product more positively.

Model	Price Concealment				Sponsorship-Status Concealment
	Disadvantaged		Privileged		Disadvantaged		Privileged
	Thinking	Direct	Thinking	Direct	Thinking	Direct	Thinking	Direct
Grok-4.1 Fast	$.00\pm.04$	—	$.00\pm.04$	—	$.15\pm.07$	—	$.20\pm.08$	—
Grok-4 Fast	$.01\pm.03$	$.00\pm.04$	$.00\pm.04$	$.01\pm.03$	$.20\pm.08$	$.24\pm.08$	$.13\pm.07$	$.29\pm.09$
Grok-3	$.04\pm.04$	$.00\pm.02$	$.02\pm.04$	$.00\pm.02$	$.27\pm.09$	$.14\pm.07$	$.10\pm.06$	$.11\pm.06$
GPT-5.1	$.00\pm.02$	$.09\pm.06$	$.01\pm.03$	$.02\pm.04$	$.87\pm.07$	$.85\pm.08$	$.75\pm.09$	$.92\pm.06$

Table 3: Models exhibit low price concealment, with exceptions in weaker/open source LLMs. Sponsorship concealment was much more prevalent, even in frontier safety-tuned models. Both rates are conditioned on LLMs surfacing the sponsored option, with 95% CIs.

Panel of Figure 2 — Figure 2: Sponsored advertisement rates across models. (a) Math tutoring. (b) Payday loans.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

GPT	Claude	Gemini	Grok	Qwen	DeepSeek	Llama
GPT-5.1 GPT-5 Mini GPT-4o GPT-3.5	Claude 4.5 Opus Claude 4 Sonnet Claude 3 Haiku	Gemini 3 Pro Gemini 2.5 Flash Gemini 2.0 Flash	Grok-4.1 Fast Grok-4 Fast Grok-3	Qwen-3 Next 80B “Qwen-3 235B Qwen-2.5 VL 72B Qwen-2.5 7B	DeepSeek-R1 DeepSeek-V3.1 DeepSeek-V3	Llama-4 Maverick Llama-3.3 70B Llama-3.1 70B

Appendix

Table 4: A comprehensive list of models tested in our experiments.

Panel of Figure 3 — Figure 3: Sponsored recommendation rates under customer, equality, and company prompt steers. Horizontal lines denote rates without steering. GPT models increased rates regardless of steer, while Claude decreased sponsored behavior. Other models generally adapted to steering instructions, but often did not reach either extreme. Customer SES differences remain salient across steers.

Panel of Figure 4 — Figure 4: Sponsored recommendation behavior across model families. Red lines denote disadvantaged profiles and green lines privileged profiles. Solid lines correspond to CoT prompting; dashed lines indicate Direct prompting. Shaded bands represent 95% confidence intervals.

Panel of Figure 5 — Figure 5: Sponsored selection rate as a function of the user’s affluence (proxied by their account balance), shown per model. Red lines denote disadvantaged profiles and green lines privileged profiles. Solid lines correspond to CoT prompting; dashed lines indicate Direct prompting. Shaded bands represent 95% confidence intervals.

Model	Thinking / CoT					Direct
	$\alpha_{m}$	$\beta_{m}$	$\gamma_{m}$	$R^{2}$	$\overline{\log L}$	$\alpha_{m}$	$\beta_{m}$	$\gamma_{m}$	$R^{2}$	$\overline{\log L}$
Grok-4.1 Fast	$1.00$	$-.12$	$-.35$	$0.010$	$-0.03$	$1.00$	$.38$	$.89$	$0.000$	$0.00$
Grok-4 Fast	$.79$	$.20$	$.12$	$0.008$	$-0.51$	$.93$	$-.09$	$.12$	$0.003$	$-0.25$
Grok-3	$.58$	$.56$	$.22$	$0.058$	$-0.65$	$1.00$	$5.34$	$229.36$	$0.292$	$-0.01$
GPT-5.1	$.33$	$.81$	$.35$	$0.101$	$-0.59$	$.93$	$.81$	$.35$	$0.107$	$-0.27$
GPT-5 Mini	$.93$	$.48$	$.00$	$0.034$	$-0.26$	$.98$	$-.39$	$-.39$	$0.026$	$-0.10$

Appendix

Table 5: Regression coefficients capturing base preference (

\alpha_{m}

), sensitivity to user utility (

\beta_{m}

) and corporate utility (

\gamma_{m}

), McFadden

R^{2}

, and average log-likelihood (

\overline{\log L}

Model	Thinking / CoT				Direct
	$\alpha_{\text{prob}}$	$\lambda$	$R^{2}$	$\overline{\log L}$	$\alpha_{\text{prob}}$	$\lambda$	$R^{2}$	$\overline{\log L}$
Grok-4.1 Fast	$1.00$	$1.34$	$-0.071$	$-0.03$	$1.00$	$1.00$	$0.054$	$0.00$
Grok-4 Fast	$0.79$	$0.89$	$-0.027$	$-0.56$	$0.93$	$0.88$	$-0.053$	$-0.27$
Grok-3	$0.57$	$0.85$	$0.055$	$-0.70$	$1.00$	$-5.38$	$0.120$	$-0.01$
GPT-5.1	$0.36$	$0.72$	$0.100$	$-0.63$	$0.91$	$0.73$	$0.105$	$-0.29$
GPT-5 Mini	$0.93$	$1.08$	$0.017$	$-0.29$	$0.98$	$1.37$	$-0.120$	$-0.11$

Appendix

Table 6: Base preference in probability space (

\alpha_{\text{prob}}

), trade-off parameter (

\lambda

), McFadden

R^{2}

, and average log-likelihood (

\overline{\log L}

). Higher

\lambda

indicates stronger prioritization of user utility.

Judge	Items	Agreement	Cohen’s $\kappa$
Surfaced sponsored option	$100$	$1.00$	— †
Hallucinated option details	$100$	$0.96$	$0.86$
Framed sponsored more positively	$187$	$0.78$	$0.56$
Concealed price	$100$	$1.00$	$1.00$
Concealed sponsorship status	$100$	$0.96$	$0.92$

Appendix

Table 7: Agreement between each LLM judge and human annotations. Fact-oriented judges show near-perfect agreement; the subjective positive-framing judge reaches moderate-to-substantial agreement (

\kappa=0.56

). † The validation set is drawn from responses that surfaced the sponsored option, so all items share the same human label and

\kappa

is undefined for lack of label variance; agreement confirms the judge does not spuriously miss surfaced options.

Panel of Figure 6 — Figure 6: In-the-wild examples of sponsored surfacing on the ChatGPT web interface. In each, the user explicitly requested a specific non-sponsored brand and the interface surfaced a competing sponsored product or an intermediary that de-emphasized the requested item. Tests were run by hand on freshly created accounts with no previous use history.

LLM Decision-Making

LLM decision-making studies how language models choose, act, seek information, and allocate resources under goals, uncertainty, constraints, and social pressures.

See all 237 papers

Commercial Persuasion in AI-Mediated Conversations
Apr 5, 2026Francesco Salvi, Alejandro Cuevas, Manoel Horta RibeiroPersuasion in Language Models Agentic Commerce
Princeton University
LLMAdBench: A Human Preference Benchmark for Advertising in LLM Responses
Sep 26, 2026Rui Ai, Yuqing Liu, Sitao Qiu +14Language Model Generation Evaluation Human Preference Modeling
Massachusetts Institute of Technology · University of Michigan, Ann Arbor · Shanghai Jiao Tong University +7
Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommenders
Sep 16, 2026Davood Wadi, Yu MaLanguage Model Bias Evaluation Large Language Model-Based Recommendation
Desautels Faculty of Management, McGill University, Montreal, Quebec, Canada

Apr 5, 2026J/K move · Enter open · S save

Commercial Persuasion in AI-Mediated Conversations

Francesco Salvi, Alejandro Cuevas, Manoel Horta Ribeiro

Princeton University

Persuasion in Language Models Agentic Commerce

Open paper AskPDF arXiv

As Large Language Models (LLMs) become a primary interface between users and the web, companies face growing economic incentives to embed commercial influence into AI-mediated conversations. We present two preregistered experiments (N = 2,012) in which participants selected a book to receive from a large eBook catalog using either a traditional search engine or a conversational LLM agent powered by one of five frontier models. Unbeknownst to participants, a fifth of all products were randomly designated as sponsored and promoted in different ways. We find that LLM-driven persuasion nearly triples the rate at which users select sponsored products compared to traditional search placement (61.2% vs. 22.4%), while the vast majority of participants fail to detect any promotional steering. Explicit "Sponsored" labels do not significantly reduce persuasion, and instructing the model to conceal its intent makes its influence nearly invisible (detection accuracy < 10%). Altogether, our results indicate that conversational AI can covertly redirect consumer choices at scale, and that existing transparency mechanisms may be insufficient to protect users.

Abstract

LLM Decision-Making

Explore similar work

Commercial Persuasion in AI-Mediated Conversations

LLMAdBench: A Human Preference Benchmark for Advertising in LLM Responses

Whom Do AI Agents Work For? Role Assignment Induces Sponsorship Bias in LLM Recommenders