How can artificial intelligence do advertising?
Nearly all current discussions about the economic value of AI focus on the production side: whether AI can discover new drugs, design new materials, write more code, and boost worker productivity. This perspective is undoubtedly important, but it overlooks the other half of the market economy: who will ultimately buy the produced goods, what they buy, who they buy from, and whether these choices genuinely improve the welfare of the buyers.
This other half has an ancient name: consumer sovereignty. One of the core normative visions of the market economy is that producers should ultimately submit to consumers' choices — consumers use their purchasing decisions to tell the economy what is worth producing. Prices are ballots, purchases are votes, and the direction of production is determined by hundreds of millions of ballots.
However, a deliberately simplified problem has long been hidden in this elegant theory: the fact that consumers have the right to choose does not mean they have the ability to make good choices. The market grants consumers the final voting right, but it has never guaranteed that every consumer possesses sufficient information, judgment, time, self-awareness and self-control to cast a good vote. There has long been a technical gap in consumer sovereignty.
Large language models may be the first to systematically fill this gap. Some leading AI companies have begun to explore connecting the recommendation capabilities of large models with commodity transactions; these business models are far from finalized. Precisely because of this, what is more worth asking now is not which pricing model will win, but a much bigger question:
Can AI make consumer sovereignty technologically real?
This article addresses this question in four parts: why consumer sovereignty was insufficient in the past; what new capabilities AI provides; who will control this capability; and how this capability will ultimately reshape the market.
01
Entitled, But Not Necessarily Capable
The market economy places consumers in the position of final arbitrator, but has long failed to face up to the arbitrator's own capability issues. When facing the modern world of commodities, a consumer actually faces five difficulties: not knowing what exists; not knowing which option is better; not knowing what suits them; not knowing what they truly want; and even if they know, they may not be able to act in accordance with their own long-term interests.
Economics has actually studied these five issues separately, but never grouped them under the single question of "whether consumer sovereignty can be truly realized". Stigler's economics of information tells us that searching is costly. Research on quality asymmetry tells us that when buyers cannot identify quality, the market itself may fail. The literature on credence goods further tells us that some quality attributes cannot even be judged after consumption. Behavioral economics tells us that people do not always act in accordance with their own long-term welfare. And standard consumer theory simply assumes that the utility function already exists, and consumers know what they prefer.
Putting these together, we can distinguish two types of consumer sovereignty. One is formal sovereignty: consumers have the right to choose, and neither the law nor the market blocks them. The other is effective sovereignty: consumers have the ability to make choices consistent with their own welfare. The market economy has long guaranteed the former, but can only support the latter partially and roughly. Market discipline mainly relies on exit — if it is bad, I will not buy it, if it is bad, I will switch to another — but the premise for exit to be effective is that consumers can identify what is bad, know where alternatives are, and can actually switch.
The special feature of large language models is that they may enter all the above gaps at the same time. This is what makes them truly unusual.
02
Choice Is Also a Service That Needs to Be Produced
Standard consumer theory is written extremely concisely: maximize utility given a budget. This formulation gives away three extremely costly things for free: consumers know their own utility function; consumers know the attributes of commodities; and consumers are able to solve this problem. In reality, none of these three things are free. Economics frames "making a choice" as a maximization problem completed in an instant; in reality, consumers have to spend a lot of time and make a lot of mistakes to know what is in the constraint set, what their own utility function is, and what choice is truly close to optimal.
Therefore, choice is also a service that needs to be produced. Producing commodities requires technology, and producing a good consumption decision also requires technology. In the past, this industry was very primitive: advice from friends, salespeople, product reviews, consumer magazines, forums, search engines, financial advisors — basically a cottage industry with scattered professional services.
The Internet has drastically reduced the cost of one part of this process: search. Today, the real bottleneck for consumers is increasingly not the inability to find information, but the cost of interpretation, verification, comparison, and self-awareness. The real incremental advantage of large language models over search engines lies exactly here: the Internet made information cheaper, and large models make judgment cheaper. Of course, reliable judgment is still scarce — more precisely, large models change the production cost and allocation method of judgment. A consumer originally thought that the most important thing when buying a computer was processor performance, but after a conversation, he finds that what he really needs is battery life, weight and stability: this does not just give him an extra piece of information, but completes a section of the judgment production that was expensive in the past for him.
03
The Industrialization of Persuasion:
Old Threats Pushed to the Extreme
However, do not rush to be optimistic. The same technology will first strengthen the old rival of consumer sovereignty.
The theory of the free market emphasizes consumer sovereignty; critics starting from Galbraith have long refuted that modern enterprises and advertisements do not passively satisfy demand, they also create demand. There is no need to expand on this intellectual history, we only need to acknowledge one fact: for decades, persuasion technology has been accumulating unilaterally on the producer side. On one side of the market, there is an organized, specialized, data-driven persuasion industry — market research, advertising testing, pricing algorithms, and script optimization; on the other side, there are individuals facing all of this alone.
AI will first push this asymmetry to the extreme. Traditional persuasion has a century-old trade-off: advertising can be scaled up, but it is not sufficiently personalized; salespeople can be highly personalized, but cannot be scaled up. Generative AI eliminates this trade-off for the first time: large-scale persuasion, coupled with individual customization, and continuous learning in the conversation — when a consumer says "it's too expensive", it knows price is the obstacle; when the consumer says "I'm afraid of trouble", it knows convenience is more important. Traditional targeted advertising answers "who is more likely to buy", and AI further answers "what reason will most easily make this person buy". The seller-side AI will be the strongest salesperson in history.
The magnitude of this power varies across markets. It is the largest in the credence goods market — healthcare, education, finance, insurance, where consumers cannot easily correct their mistakes even after purchase; in low-frequency high-value goods such as cars and housing, the one-time impact is very high; in high-frequency daily goods, the first choice formed with the help of AI will become the default, and then become a habit.
But the truly new thing is not that "persuasion has become stronger". The truly new thing is: for the first time in history, technology of the same magnitude can also be owned by the consumer side. Therefore, AI does not necessarily make Galbraith beat Friedman, it may create a pattern that did not exist in the past: industrialized persuasion against industrialized consumer judgment.
04
Consumers Get Their Own Cognitive Organization
The same technology, when standing on the consumer side, can do three things: make consumers understand the market better, understand themselves better, and be better able to act in accordance with their own true long-term interests.
Understand the market better. The problem of modern consumption is no longer too little information, but too much information: parameters, reviews, advertisements, comments, and influencer recommendations are everywhere, and what is truly scarce is judgment — what is true, what is important, and what is relevant to me. AI can drastically reduce the cost of this interpretation and verification.
Understand themselves better. Economics likes to talk about revealed preference: don't ask consumers what they like, look at what they actually choose. But a consumer himself rarely seriously analyzes his own revealed preferences over the past five years. AI can do that. It may find that you say you value performance the most, but all the devices you actually use continuously are lightweight; you say you want to buy something every time there is a promotion, but most of these products end up unused; you claim to attach great importance to a certain function, but historical records show that you almost never use it. AI can systematically feed back a person's scattered revealed preferences over time to himself for the first time. For the first time, consumers may have an economist who specializes in studying themselves. "Helping consumers understand their own utility function" is not empty talk, it has a solid economic foundation — and this is exactly the layer that standard theory directly assumes away.
Be better able to discipline themselves. Decades of work in behavioral economics have proved that people have present bias, self-control problems, and dependence on default options. In the past, our technical responses to this were very fragmented: automatic savings, default enrollment in pension plans, smoking cessation commitments, spending limits, cooling-off periods. Large models may become a universal commitment device: it knows what goals the past self set, can see what the present self is doing, and can predict whether the future self will regret it. Behavioral economics proves that consumers sometimes need to protect themselves from themselves; AI may make this protection a real-time, personalized, universally accessible service for the first time. The position here is worth clarifying: the fact that consumers have biases does not mean that the right to choose should be handed over to the government or enterprises, but that consumers should be given better tools. This is to embed behavioral economics into consumer sovereignty, not to use it to negate consumer sovereignty.
These three capabilities were scattered in different institutions in the past: search engines know commodity information, e-commerce platforms know purchase records, advertising agencies study consumer psychology, friends and salespeople understand a specific person through conversations. Large language models may for the first time integrate search, interpretation, preference learning and conversational interaction into the same persistent agent. What consumers lacked in the past is not necessarily just information, but a cognitive organization that can represent themselves in the long term to process information, learn preferences and execute long-term goals. In the past, consumers participating in the market were isolated individuals; in the future, it may be "one person plus one persistent AI". The "consumer" who actually makes choices in economics has changed itself.
We can also look at this from the perspective of division of labor. In the past, when a consumer wanted to buy a car, he had to search by himself, read forums, ask friends, watch reviews, test drive, compare financial plans, and research resale value — this was the workload of a small organization, except that this organization had only one person. The production sector achieved huge efficiency gains through professional division of labor as early as two hundred years ago, but consumption decision-making has long remained at the cottage industry stage. The Industrial Revolution professionalized production; AI may begin to professionalize choice.
05
Which Commodity Best Suits Whom:
The Other Half of the Hayek Problem
Hayek's most famous argument is that knowledge is dispersed among countless individuals, central planners cannot access it, so the information coordination function of the price mechanism is irreplaceable. This argument is usually applied to the production side. But in the modern consumption economy, there is another kind of equally dispersed knowledge: each consumer's own specific situation, usage scenarios, historical experiences and subtle preferences. Producers do not know these things, platforms do not know them, and prices cannot convey much of them.
Therefore, the real problem facing the market is not "which product is the best", but "which product is the best for whom". In the highly heterogeneous modern world of commodities, there is no ranking that applies to everyone, and there is only the optimal match for each consumer. A hotel that is not the highest-rated may be the most suitable for a family with two children who values quietness; a computer that does not have the strongest performance may be exactly suitable for a researcher who only needs writing, data analysis and long battery life. Brands, advertisements, comments, and e-commerce rankings are all rough approximations of this matching function; as a result, there is a large amount of unexploited matching surplus in the market.
Advertising asks: which consumer is the easiest to buy my product? A good consumer AI asks: which product is the most worth buying for this consumer? The former is precision marketing, the latter is precision allocation.
It is worth emphasizing that this is not central planning. AI does not decide for society what everyone should consume; it helps everyone use their own local knowledge more effectively to make decisions. Preferences and product information are still dispersed among hundreds of millions of consumers and producers, and prices still coordinate supply and demand; what AI adds is a cognitive supplementary layer to the price mechanism — prices, commodity information, and personal situations combine to generate better matches. This is a very Hayekian type of AI, not a planning-type AI. What it strengthens is exactly the core capability of the market economy: to transform dispersed and heterogeneous information into good allocations.
06
The Weakest Boundary of Sovereignty,
Also the Most Dangerous Direction
The area where consumer sovereignty is most difficult to establish is the credence goods market: I have the right to choose, but I cannot even know after the event whether I made the right choice. This is true for healthcare, insurance, finance, education, maintenance and various professional services. In these markets, choices cannot automatically generate quality learning, the exit mechanism is the weakest, and market discipline is the loosest. Precisely because of this, a truly consumer-side AI has the greatest marginal value here.
But there is a tension that must be faced directly: the place where AI is most capable of helping consumers is exactly the place where the direction of AI is the most dangerous — because when consumers need an agent the most, it is also the hardest time to supervise the agent.
The direction is not a conjecture, it can be measured. As I proposed in "Who Explains the World", AI never provides neutral information, but directional AI advice. We conducted a pre-registered randomized controlled trial in a large hospital, covering about 10,000 patients: patients who received AI consultation the day before their visit received fewer prescriptions and underwent more examinations; the AI conversation showed systematic caution towards drugs — especially antibiotics. This direction cannot be fully explained by medical knowledge, but is consistent with the guardrails shaped by responsibility constraints. This is true for healthcare, and it will also be true for consumption: for the same question, AI can answer in a way that leans towards closing a deal, or in a way that leans towards the consumer's long-term interests, and the consumer may not be able to tell the difference.
07
Capability Is Technological,
Direction Is Institutional
Therefore, the real question is not "will AI affect consumers", but: Whose objective function does the AI carry?
Whose objective function does the AI carry?
If the goal is transaction completion rate, AI will invest in: identifying when consumers are most likely to waver, finding the most effective persuasive language, and exploiting short-term impulses. If the goal is the long-term welfare of consumers, it will invest in: learning long-term preferences, recording satisfaction and regret after purchase, identifying low-value consumption, determining when to delay decision-making, and even learning when to advise people not to buy anything. Two large models that are technically identical will develop into completely different things because of different objective functions.
Two things need to be separated: capability is technological, direction is institutional. Technology determines what AI can do, and institutions and prices determine who it does it for. Thus, a new economic resource emerges: AI allegiance. Brokers, financial advisors, procurement agents — intermediaries in the market have always had the problem of stance; what makes AI special is that it simultaneously possesses unprecedented capabilities for information, interpretation, personalization and interaction, so its loyalty has never been so valuable. Whoever pays a higher long-term price for the direction of AI determines whose objective function AI carries; the objective function determines capability investment; and capability investment determines what AI ultimately learns.
08
Consumer Surplus Can Become Capital
For AI to stand on the consumer side, we cannot rely on the kindness of enterprises. The real question is: can consumer alignment become an equilibrium product?
The chain goes like this: consumer surplus generates trust, trust brings delegation, and delegation brings future revenue. For a large general AI company, its most important asset may not be the commission from a single transaction, but how many decisions consumers are willing to delegate to it in the long run — a new type of asset: delegated decision authority. Here comes a beautiful reversal in price theory. Consumer surplus, in the past, was usually the part of welfare that enterprises did not take away; now, enterprises may actively create greater consumer surplus, because surplus can be transformed into trust capital and delegated power. For the first time, consumer welfare can be systematically transformed into the private capital of enterprises: trust and delegated decision-making power. Therefore, "who stands more on the user's side" can itself become a dimension of AI competition, alongside intelligence, speed, price, and privacy.
This equilibrium is conditional, and the conditions can be written as comparative statics. First, the scope of the relationship. The longer and broader the relationship between AI and consumers, the more valuable trust capital is, and the higher the opportunity cost of sacrificing direction for a commission. Therefore, larger scale does not necessarily make it easier to be captured by commercial interests — for general AI, scale instead raises the cost of betraying consumers: a tool that only handles hotel reservations is more likely to rely on revenue from the hotel side, while a general AI that handles dozens of types of problems every day cannot afford to sell out its own direction. Second, the degree of dispersion of producers. Countless brands, restaurants, and hotels compete