REVIEW 3 major objections 5 minor 23 references
Advertising in AI systems: Society must be vigilant
T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash
Pith's one-line read Advertising will reshape what AI systems tell us, so the paper calls for transparency rules and debiasing tools before paid content becomes indistinguishable from answers.
desk verdict A clear, well-scoped position paper on advertising in AI whose 'inevitable' forecast is asserted rather than argued, but which deserves referee time for its taxonomy and design principles. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central objects are generative advertisements — model outputs $y$ generated from an information set $z$ that contains sponsored entries $z_{ad}$ alongside ordinary retrieved content $z_{w/o\,ad}$ — and the target response $y_{w/o\,ad}(x)$ that the system would have produced without sponsors. The paper operationalizes ad-serving through standard tool-call patterns: a content tool gathers non-sponsored sources, an advertisement tool acquires sponsor content, an optional chain-of-thought consolidates them, and sponsored text is appended to the context with an <ad> tag so the model can preserve it faithfully. Four design principles constrain the system: faithfulness (a fidelity metric $F(y_{w/ad}(x), z_i)$ above a threshold), utility ($U(y_{w/ad}(x)) \geq \alpha U(y_{w/o\,ad}(x))$), privacy (per-user opt-out $o_u \in \{\text{IN}, \text{OUT}\}$ that blocks personal-data ad targeting), and provenance (a map $\phi$ from output tokens to contributing sources). For the end user, the machinery is debiasing: recovering $\hat{y}_{w/o\,ad}(x)$ from the ad-augmented output by a single-pass ad-free model or by sampling multiple ad-influenced responses and aggregating the stable core.
What would settle it
Observe whether major AI assistants begin serving sponsored product mentions in ordinary responses; if no such serving infrastructure appears at scale over several years, the inevitability claim is weakened. Separately, test the proposed debiasing by comparing recovered outputs against a known ad-free model on queries with inserted sponsors; if the pipeline cannot approximate $y_{w/o\,ad}(x)$, the mitigation premise is falsified.
Extended reading notes
Core claim
The paper's central claim is that advertising will inevitably shape AI-mediated content delivery because AI platforms, like search engines and social media before them, will monetize user interactions through commercial sponsorship. It introduces a distinction between static advertisements (identical sponsored media inserted alongside content) and generative advertisements (outputs produced from retrieved material that includes sponsored information), arguing that the latter are dynamic, personalized, and lack provenance, making them harder to regulate. The paper frames the AI system as an advertising platform connecting advertisers and consumers, derives stakeholder requirements, and lays out four design principles — faithfulness, utility, privacy, and provenance — that such systems must satisfy. It then proposes user-side debiasing approaches, direct debiasing by an ad-free model and multi-sampling aggregation, to recover the unsponsored response the model would have given. The contribution is a road map: envisioning how commercial content will be served, what must be in place to keep it safe, and what open problems researchers and regulators need to solve.
Load-bearing premise
Both user-side mitigation strategies in Section 4 assume that an ad-free model is available to reconstruct the unsponsored answer; if the platform controls every capable model, no trustworthy baseline remains.
Editorial extensions
If this is right
- AI responses to ordinary queries will increasingly include paid product recommendations, product placements, and paid visibility for specific sources, and users will not be able to tell these from neutral answers.
- Regulators cannot audit compliance by looking at ad units alone, because outputs vary per user and per session; they will need logging, provenance mappings, and reproducibility tools.
- Training on user feedback that includes ads threatens to make commercial bias permanent, since the model may learn sponsored content as normal knowledge; providers must separate sponsored outputs from alignment data.
- The proposed design principles give a concrete checklist: fidelity to sponsor material, preserved utility, opt-out privacy, and per-token provenance.
- End users could block commercial influence using single-pass debiasing or multi-sampling aggregation, if an independent ad-free model exists.
Reading between the lines
- The paper leaves implicit that the platform controlling both the ad-augmented model and any debiasing model creates a conflict of interest; even a technically sound debiaser would depend on the platform's cooperation or on open-weight models the platform does not control.
- A testable extension is a benchmark that injects sponsored entries into retrieval corpora and measures how often model outputs change; such a benchmark could quantify commercial bias before regulation is written.
- Because generative ads do not need personal data to work — a model can endorse a sponsor's product from general context — opt-out of targeting may be necessary but not sufficient for user autonomy.
- The same paid-visibility infrastructure that promotes products can promote political content, so advertisement regulation overlaps with content-moderation and election-integrity policy.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper argues that commercial advertising will inevitably shape the content delivered by generative AI systems, drawing an analogy with web search and social media. It distinguishes static advertisements from dynamic 'generative advertisements,' proposes four design principles (faithfulness, utility, privacy, provenance) for commercially-influenced AI systems, sketches a tool-call-based implementation architecture, and outlines two user-side debiasing strategies. It concludes with open questions and a call to action for regulators, platforms, and researchers. The paper is a position piece with no empirical evaluation; its contribution is a conceptual framework and a research agenda rather than demonstrated results.
Significance. If its central prediction is accepted, the paper provides a timely and clearly structured framing of a nascent problem. Its strengths are the stakeholder analysis, the concrete design principles, and the honest enumeration of open problems (including the risk of long-term alignment damage and the difficulty of regulatory compliance). The paper is well-referenced and reads as a responsible call to action. However, the central 'inevitability' claim is an assertion rather than a derivation, and the mitigation proposals depend on access to an ad-free model without discussing how that access is obtained. As a position piece, the paper's significance lies in agenda-setting rather than in establishing results.
major comments (3)
- [Section 1; Section 2.2] The central premise that ad-based monetization of AI systems is 'inevitable' is asserted rather than derived; the only support is an analogy with search and social media. The paper's own Section 2.2 identifies structural differences that could make ad-supported generative AI less attractive than alternatives: ads must be integrated without violating user expectations, alignment feedback can degrade model quality over time, and stochastic outputs complicate compliance verification. The paper does not compare per-query serving cost against potential ad revenue or users' willingness to pay for ad-free access, nor does it discuss subscription or API pricing as competing business models. The authors should either supply an economic or empirical mechanism for the inevitability claim or soften it to 'likely' or 'possible'; the rest of the paper's value does not depend on 'inevitable.'
- [Section 4.2, 'Limitations and discussion'] Both proposed debiasing methods rely on 'the availability of an ad-free model,' as the paper states, but the paper does not address how users or auditors would obtain such a model when the platform controls both the ad-augmented system and any debiasing tool. A platform has little incentive to provide a high-fidelity ad-free baseline, and third-party auditors may lack access to the ad-influenced outputs or to the model's internals. The strategies in Sections 4.1 and 4.2 therefore do not yet establish a feasible user-side mitigation path; the paper should discuss the institutional or regulatory conditions under which an ad-free model is available.
- [Section 3.2] The four design principles are introduced with 'must fulfill,' but the paper provides no argument that these are necessary or sufficient for responsible commercialization. In particular, the 'Provenance' principle requires a token-level mapping phi from output tokens to contributing sources, but generated text often paraphrases and fuses multiple sources, and the paper does not discuss how this mapping would be computed reliably. The principles are reasonable as an initial proposal, but they should be framed as desiderata to be evaluated, not as categorical requirements.
minor comments (5)
- [Section 3.2, after Eq. (1)] The text reads 'including including relevance'; this appears to be a typo for 'including relevance.'
- [Figure 1 caption] 'Theydoaffect the primary outputs of AI systems' should read 'They do affect'.
- [Section 2.2, 'Ease of regulation'] The claim that 'it is generally clear how one would verify whether commercial content is compliant' sits uneasily with the earlier discussion of native advertising; consider acknowledging that generative ads may blur the line further.
- [Section 5, 'User autonomy and consent'] The statement that 'large language models can endorse sponsored content even in the absence of explicit personalization' lacks a citation; adding a supporting reference or concrete example would strengthen the argument.
- [Section 4.2] The citation of Maini et al. (2024) and Modarressi et al. (2025) as 'statistical tests for identifying differences in the distributions of language model outputs' seems imprecise; those works are about dataset inference and causal inference from text, respectively. The authors should clarify exactly which aspect is being cited.
Circularity Check
No circularity: the paper's central claim is an analogical forecast and its design/debiasing proposals are explicit framings with acknowledged assumptions, not derivations from their own definitions.
full rationale
This is a position paper, not a derivation chain. The central claim that "it is inevitable that these commercial forces will reshape AI-mediated content delivery" rests on an analogy with web search and social media plus advertising market data; the paper supplies no equation or fitted parameter that would make the conclusion equivalent to its premises. The terms "generative advertisements," the fidelity metric F, and the utility measure U are stipulative definitions or proposed design requirements, and they are not used to deduce the inevitability claim. The debiasing section explicitly states that both proposed approaches "rely on the availability of an ad-free model," which is an honest limitation of the proposal, not a hidden assumption that forces the conclusion. There are no self-citations by the authors (Wu and Bao) in the reference list, and no prior-work uniqueness theorem is imported to rule out alternatives. The paper also explicitly maps "generative advertisements" onto the existing term "native advertisements," so it is not renaming a known result as a new organization. The main weaknesses—an unsupported extrapolation from search/social to AI and the practical reliance on an unsponsored baseline—are correctness or feasibility risks, not circularity. Therefore the appropriate finding is no significant circularity, score 0.
Assumptions & free parameters
assumptions (4)
- domain assumption The historical analogy between search engines/social media and generative AI is valid for predicting monetization.
- domain assumption Commercial incentives are inevitable and will shape AI content.
- ad hoc to paper The proposed design principles (faithfulness, utility, privacy, provenance) are necessary and sufficient for responsible commercialization.
- ad hoc to paper An ad-free model or reliable debiasing is possible.
invented entities (3)
-
generative advertisement
-
advertisement tool call
-
ad-report tag
Cite this review
Pith. "Pith review of Advertising in AI systems: Society must be vigilant." pith.science (2026). https://pith.science/paper/24K52ZLG
@misc{pith2026250518425,
author = {Pith},
title = {Pith review of: Advertising in AI systems: Society must be vigilant},
year = {2026},
howpublished = {\url{https://pith.science/paper/24K52ZLG}},
note = {Machine review of arXiv:2505.18425}
}
read the original abstract
AI systems have increasingly become our gateways to the Internet. We argue that just as advertising has driven the monetization of web search and social media, so too will commercial incentives shape the content served by AI. Unlike traditional media, however, the outputs of these systems are dynamic, personalized, and lack clear provenance -- raising concerns for transparency and regulation. In this paper, we envision how commercial content could be delivered through generative AI-based systems. Based on the requirements of key stakeholders -- advertisers, consumers, and platforms -- we propose design principles for commercially-influenced AI systems. We then outline high-level strategies for end users to identify and mitigate commercial biases from model outputs. Finally, we conclude with open questions and a call to action towards these goals.
Figures
Figures from the paper (1 more)
Reference graph
Works this paper leans on
-
[4]
Gender-tuning: Empowering Fine-tuning for Debiasing Pre-trained Language Models
Ghanbarzadeh, S., Huang, Y ., Palangi, H., Moreno, R. C., and Khanpour, H. Gender-tuning: Empow- ering fine-tuning for debiasing pre-trained language models.arXiv preprint arXiv:2307.10522,
-
[9]
The ad wars: retrospective measurement and analysis of anti- adblock filter lists
Iqbal, U., Shafiq, Z., and Qian, Z. The ad wars: retrospective measurement and analysis of anti- adblock filter lists. InProceedings of the 2017 Internet Measurement Conference, pp. 171–183,
work page 2017
-
[12]
Causal inference on outcomes learned from text.arXiv preprint arXiv:2503.00725,
Modarressi, I., Spiess, J., and Venugopal, A. Causal inference on outcomes learned from text.arXiv preprint arXiv:2503.00725,
-
[13]
Supporting human-AI collabo- ration in auditing LLMs with LLMs
Rastogi, C., Tulio Ribeiro, M., King, N., Nori, H., and Amershi, S. Supporting human-AI collabo- ration in auditing LLMs with LLMs. InProceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society, pp. 913–926,
work page 2023
-
[14]
12 Ravfogel, S., Elazar, Y ., Gonen, H., Twiton, M., and Goldberg, Y . Null it out: Guarding protected attributes by iterative nullspace projection.arXiv preprint arXiv:2004.07667,
arXiv 2004
-
[17]
A reproducibility and generaliz- ability study of large language models for query generation
Staudinger, M., Kusa, W., Piroi, F., Lipani, A., and Hanbury, A. A reproducibility and generaliz- ability study of large language models for query generation. InProceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the Asia Pacific Region, pp. 186–196,
work page 2024
-
[18]
M., Hauth, A., Millican, K., et al
Team, G., Anil, R., Borgeaud, S., Alayrac, J.-B., Yu, J., Soricut, R., Schalkwyk, J., Dai, A. M., Hauth, A., Millican, K., et al. Gemini: a family of highly capable multimodal models.arXiv preprint arXiv:2312.11805,
-
[20]
Measuring and reducing gendered correlations in pre-trained models.arXiv preprint arXiv:2010.06032,
Webster, K., Wang, X., Tenney, I., Beutel, A., Pitler, E., Pavlick, E., Chen, J., Chi, E., and Petrov, S. Measuring and reducing gendered correlations in pre-trained models.arXiv preprint arXiv:2010.06032,
arXiv 2010
Show all 23 references
-
[22]
A survey on real time bidding advertising
Yuan, Y ., Wang, F., Li, J., and Qin, R. A survey on real time bidding advertising. InProceedings of 2014 IEEE International Conference on Service Operations and Logistics, and Informatics, pp. 418–423. IEEE,
2014
-
[2003]
liar, liar pants on fire
Wang, J., Ma, W., Sun, P., Zhang, M., and Nie, J.-Y . Understanding user experience in large language model interactions.CoRR, 2024a. Wang, S.-Y ., Su, W., Chen, S., Tsai, C.-Y ., Misztal, M., Cheng, K. M., Lin, A., Chen, Y ., and Chen, M. Y . RoomDreaming: Generative-AI appro...
2024 arXiv
-
[2008]
Prompting gpt-3 to be reliable.arXiv preprint arXiv:2210.09150,
Si, C., Gan, Z., Yang, Z., Wang, S., Wang, J., Boyd-Graber, J., and Wang, L. Prompting gpt-3 to be reliable.arXiv preprint arXiv:2210.09150,
-
[2011]
The llama 3 herd of models.arXiv preprint arXiv:2407.21783,
Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,
-
[2012]
Fairness mediator: Neutralize stereotype associations to mitigate bias in large language models.arXiv preprint arXiv:2504.07787,
Xiao, Y ., Liu, A., Liang, S., Liu, X., and Tao, D. Fairness mediator: Neutralize stereotype associations to mitigate bias in large language models.arXiv preprint arXiv:2504.07787,
-
[2013]
D., Singh, A., Baumli, K., Iqbal, S., Bishop, C., Roelofs, R., et al
11 Kumar, A., Zhuang, V ., Agarwal, R., Su, Y ., Co-Reyes, J. D., Singh, A., Baumli, K., Iqbal, S., Bishop, C., Roelofs, R., et al. Training language models to self-correct via reinforcement learning.arXiv preprint arXiv:2409.12917,
-
[2014]
Independence tests for language models.arXiv preprint arXiv:2502.12292,
Zhu, S., Ahmed, A., Kuditipudi, R., and Liang, P. Independence tests for language models.arXiv preprint arXiv:2502.12292,
-
[2015]
Mechanism design for large language models
Duetting, P., Mirrokni, V ., Paes Leme, R., Xu, H., and Zuo, S. Mechanism design for large language models. InProceedings of the ACM Web Conference 2024, pp. 144–155,
2024
-
[2016]
Social bias evaluation for large language models requires prompt variations.arXiv preprint arXiv:2407.03129,
Hida, R., Kaneko, M., and Okazaki, N. Social bias evaluation for large language models requires prompt variations.arXiv preprint arXiv:2407.03129,
-
[2017]
Online advertisements with llms: Opportunities and challenges.ArXiv, abs/2311.07601,
Feizi, S., Hajiaghayi, M., Rezaei, K., and Shin, S. Online advertisements with llms: Opportunities and challenges.ArXiv, abs/2311.07601,
-
[2019]
R., Feng, S., Aron, R., Richardson, T., Choi, Y ., Fisher, D
10 Fisher, J. R., Feng, S., Aron, R., Richardson, T., Choi, Y ., Fisher, D. W., Pan, J., Tsvetkov, Y ., and Reinecke, K. Biased AI can influence political decision-making.ArXiv, abs/2410.06415,
-
[2020]
Debiasing algorithm through model adaptation.arXiv preprint arXiv:2310.18913,
Limisiewicz, T., Mareˇcek, D., and Musil, T. Debiasing algorithm through model adaptation.arXiv preprint arXiv:2310.18913,
-
[2022]
D., Cox, S., Laurent, J
Skarlinski, M. D., Cox, S., Laurent, J. M., Braza, J. D., Hinks, M., Hammerling, M. J., Ponnapati, M., Rodriques, S. G., and White, A. D. Language agents achieve superhuman synthesis of scientific knowledge.arXiv preprint arXiv:2409.13740,
-
[2023]
P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., et al
Hurst, A., Lerer, A., Goucher, A. P., Perelman, A., Ramesh, A., Clark, A., Ostrow, A., Welihinda, A., Hayes, A., Radford, A., et al. GPT-4o system card.arXiv preprint arXiv:2410.21276,
-
[2024]
DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning.arXiv preprint arXiv:2501.12948,
Guo, D., Yang, D., Zhang, H., Song, J., Zhang, R., Xu, R., Zhu, Q., Ma, S., Wang, P., Bi, X., et al. DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning.arXiv preprint arXiv:2501.12948,
Reviewed August 7, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.