In November 2023, four MIT economists published a working paper that does not appear in any advertising industry dashboard, any brand safety report, or any media plan. The paper is titled, with the dry precision of academic economics, "A Model of Behavioral Manipulation." Its authors are Daron Acemoglu, Ali Makhdoumi, Azarakhsh Malekian, and Asuman Ozdaglar. The paper was subsequently published in the American Economic Review: Insights in March 2025 under the title "When Big Data Enables Behavioral Manipulation."
The paper's central contribution is a formal economic model of what happens when a platform powered by artificial intelligence gains detailed knowledge of which products are temporarily attractive and which are persistently low-quality. The authors introduce a variable they call glossiness — denoted α — representing a product attribute that makes it appear more attractive than it actually is, at least for a period of time.
Glossiness can be almost anything. Better physical presentation. More appealing packaging. A free trial period. A salient feature that dominates initial impressions while hidden costs accumulate. An addictive attribute that generates short-term engagement before the costs become apparent. In the formal model, each product has a quality θ ∈ {0,1} — low or high — and a glossiness state α ∈ {0,1}. High-quality products are never glossy. Low-quality products are glossy with probability λ. Glossiness decays at rate ρ: when ρ is large, the attractive facade fades quickly and the consumer learns the truth. When ρ is small, the glossiness persists, and the consumer remains deceived.
The platform knows which products are temporarily glossy. The consumer does not. The gap between what the platform knows and what the consumer believes is the exploitation opportunity.
Before AI, the platform's knowledge of glossiness was approximately equal to the consumer's. Both faced the same information asymmetry. AI changes this. The platform, with sufficient behavioral data from millions of similar users, learns which products are glossy and for how long — information the individual consumer cannot access. The authors call this the post-AI information regime. It is the foundation of the manipulation calculus.
The paper proves six theorems. The most consequential are theorems 3, 4, and 5.
Theorem 3 establishes that under the post-AI information regime, the platform's optimal strategy bifurcates depending on the glossiness decay rate ρ. When ρ is small — glossiness persists — the platform prefers to offer the low-quality glossy product. The reason is straightforward: the consumer cannot receive negative signals while the product is glossy. The exploitation window is open. The platform harvests the consumer's surplus during the period when the product appears better than it is, before the decay sets in.
Theorem 4 establishes that when ρ is large — glossiness fades quickly — the platform prefers to offer higher-quality products. Here the exploitation window is too short. The consumer learns the truth before the platform can extract sufficient value. In this regime, AI helps the consumer: the platform, unable to exploit, instead guides the user toward products that match their actual preferences.
Theorem 5 establishes the asymmetric welfare outcome. Platform profits increase in both regimes — the platform benefits from AI regardless of whether it helps or manipulates. But consumer utility decreases when ρ is small and increases when ρ is large. The authors call this the manipulation effect. It is not a side effect. It is the predictable outcome of a system optimized for engagement and conversion, given enough data about product glossiness and consumer susceptibility.
In the mathematical framework, the platform's profit function is monotonically increasing in its information about consumer vulnerability. The consumer's welfare function is not.
Theorem 6 delivers the finding most relevant to the current advertising landscape. When ρ is small — long-lived glossiness — increasing the number of products n available on the platform intensifies behavioral manipulation and further decreases user welfare. The platform has more low-quality glossy products to match to vulnerable consumers. The manipulation effect grows with the size of the catalog.
This is the double whammy. AI enables better targeting, which the industry presents as consumer benefit — matching people with products they actually want. But the same AI enables better identification of manipulation opportunities — which products are glossy, for which consumers, and when to offer them. As platforms grow, adding more products and more users, the harmful effects of manipulation grow faster than the helpful effects of matching.
The finding maps directly onto the architecture of modern advertising. A platform with millions of products, billions of user behavior records, and real-time bidding infrastructure is precisely the environment where ρ tends toward the small values that favor manipulation. The products with the most persistent glossiness — subscription services with hidden cancellation flows, freemium products with paying upsell paths, cosmetics with presentation dominating formulation, gaming titles with rewarding early progression and expensive later stages — are the products most aggressively profiled by advertising AI.
In a 2024 follow-up paper, "Online Business Models, Digital Ads, and User Welfare," Acemoglu and co-authors applied the framework specifically to advertising-funded platforms. The conclusion: advertising-based monetization systematically leads to manipulation of naive users. The platform's interest and the user's interest are not aligned. The misalignment is not a bug in the design. It is the design.
The paper's three policy recommendations are structurally sound and largely unimplemented. First, competition policy: limiting platform size or requiring data sharing between competitors could reduce manipulation, though the authors note this is a blunt instrument. Second, user information: informing consumers about product glossiness and platform incentives has limited effect because behavioral research consistently shows disclosure does not neutralize manipulation. Third, limits on price discrimination: reducing the platform's ability to extract surplus from vulnerable users would reduce the profitability of manipulation.
The Brookings Institution's Center on Regulation and Markets presented the paper at its 2023 AI Authors' Conference and described it as "the first systematic analysis" of when AI-enabled behavioral manipulation harms consumers. The UK Competition and Markets Authority cited the paper's concerns in its March 2026 guidance on agentic AI and consumer law, warning that AI systems may "push consumers toward worse deals" — the exact welfare outcome theorem 5 predicts. The EU AI Act's Article 5 prohibition on manipulative AI techniques is thematically aligned, though it does not reference the paper specifically.
The advertising industry has not responded to the paper directly. This is not unusual — academic papers rarely generate public comment from the targets of their analysis. But the policy direction the paper anticipated has materialized. The US Banning Surveillance Advertising Act was introduced in 2023. The UK Digital Markets, Competition and Consumers Act came into force in April 2025. The FTC issued a policy statement on AI and Section 5 in March 2026 applying existing unfair and deceptive practices authority to algorithmic systems.
What has not materialized is any measurement of the phenomenon the paper describes. The manipulation calculus does not appear in ROAS reports. It does not appear in brand lift studies. It does not appear in engagement rate dashboards. The welfare cost to consumers — predicted by the model to be substantial when glossiness is long-lived — is not being counted, because the infrastructure for counting it does not exist. The platforms that would need to report it have no incentive to build it.
The double whammy theorem implies that the problem intensifies with platform scale. As AI systems grow more capable of identifying glossiness across larger product catalogs and more precisely targeting users susceptible to exploitation, the gap between what the manipulation calculus predicts and what the industry's metrics report will widen. The model says this is not a side effect. The industry's silence on the paper's findings suggests agreement, if only by omission.