What if AI wrote every ad on Earth? The economics of sameness

When the machine is free and everyone has it, ads converge, attention gets dearer and craft stops proving anything. The trend is already in Meta's numbers.

Times Square at night lit by dozens of illuminated billboards
Photo: chensiyuan, CC BY-SA 4.0, via Wikimedia Commons

If AI wrote every ad on Earth, the cost of making an ad would fall toward zero, so the things that stay scarce would get expensive: people's attention, the ad slots that hold it, first-party data and a brand that people already remember. Ads would also start to look alike; in one controlled study, stories written with AI ideas were 10.7% more similar to each other. The early version is already visible: Meta's advertisers using its generative AI tools went from about 1 million to over 8 million in under two years, while the average price per ad kept rising.

Key takeaways
  • Meta reported 4 million advertisers on its gen-AI creative tools by end-2024 and 8 million+ by April 2026.
  • Price per ad still rose: +14% in Q4 2024 and +12% in Q1 2026, year on year.
  • AI help made each story better but the pool 10.7% more alike (Doshi and Hauser, Science Advances).
  • In a Hotelling market, margin equals transport cost times the distance between rivals; sameness drives it to zero.
  • When production is free, a polished ad no longer proves anything: the handicap principle flips.

This is the most realistic what-if in the series, because most of it is a forecast rather than a fantasy. So the job here is to take the trend to its end point, 100% of ads machine-written, and use economics to see who gains and who loses when that happens.

The story

On 19 November 2018, Lexus released a one-minute film for its new ES saloon in Europe and billed it as the world's first advert scripted by artificial intelligence. Its creative agency, The&Partnership London, worked with the technical partner Visual Voice to build the scriptwriter on IBM Watson. The machine was fed about 15 years of car and luxury adverts that had won Cannes Lions, broken down into actions, objects, locations and emotion, plus Lexus's brand guidelines and Unruly's data on which car ads viewers responded to most.

The plot it produced: a takumi craftsman finishes the car and sends it into the world, it is taken away and threatened with destruction, and its automatic emergency braking saves it. The Oscar-winning director Kevin Macdonald shot it, and said the melodrama of the script convinced him of its potential. A machine trained only on past winners wrote, almost by definition, the average of past winners.

Lexus UK newsroom: Driven by Intuition, car by Lexus, story by artificial intelligence (Nov 2018)

What happens to ad prices when making ads is free?

Ad prices go up, not down, because AI multiplies the supply of ads but not the supply of human attention. More advertisers can afford to compete for the same slots, and in an auction more bidders means higher clearing prices.

Meta advertisers using at least one generative AI creative tool
See the data as a table
Advertisers (million)
Mid-20241
End-20244
Q1 20268

Source: Meta earnings calls, via AdExchanger (Jan 2025) and PPC Land (Apr 2026) · Mid-2024 is the figure Meta gave as six months before end-2024. Q1 2026 is reported as over 8 million.

QuarterAd impressions, year on yearAverage price per ad, year on yearSource
Q4 2024+6%+14%AdExchanger
Q1 2026+19%+12%PPC Land

Read the table carefully. In Q1 2026 Meta sold 19% more impressions and still charged 12% more for each. Meta credits better ad performance as well as demand, so not all of that is AI creative; but it is the opposite of what "infinite cheap ads" would predict for price.

The auction maths

Ad slots are sold in auctions. Take the textbook case, a second-price auction, where the winner pays the second-highest bid. Assume each bidder values the slot at a number drawn uniformly from 0 to 1 and bids truthfully.

\[ E[\text{price}] = E[V_{(2)}] = \frac{n-1}{n+1} \]
  1. Before AI. Say 10 advertisers can afford a creative for this audience: \( 9/11 = 0.818 \) of the top value.
  2. After AI. Production is free, so 20 compete: \( 19/21 = 0.905 \).
  3. Change. \( 0.905/0.818 = 1.106 \): about 11% higher clearing prices for the same slot, with no change in the slot itself.

    With these example inputs (10 advertisers before AI and 20 after, values uniform from 0 to 1), calculated in this post from the second-price expected price (n − 1)/(n + 1), the standard result for uniform values: 0.905 ÷ 0.818 ≈ 1.106.

  4. Limit. As \( n \to \infty \), the price tends to the full value of the slot. The platform captures the surplus AI freed up.

Use this to see how much more you would pay as AI brings more rivals into your auction.

Auction price as more advertisers compete

The red Coca-Cola Christmas truck parked on a street in Chester, England
The real Coca-Cola Christmas truck on tour in Chester. The physical truck costs money to run, which is part of why people believe in it.Photo: Jeff Buck, CC BY-SA 2.0, via Wikimedia Commons

Would every ad start to look the same?

Yes, measurably. Two controlled studies in 2024 found that people using generative AI produced better individual work but a narrower range of ideas across the group.

Generative AI enhances individual creativity but reduces the collective diversity of novel contentAnil R. Doshi, Oliver P. Hauser · Science Advances 10, eadn5290 · 2024

293 writers wrote short stories under three conditions: on their own, with one AI-generated idea, or with up to five AI ideas; 600 evaluators made 3,519 ratings. With five AI ideas, novelty rose 8.1% and usefulness 9.0%, and the least creative writers gained most (novelty +10.7%, usefulness +11.5%). But stories in the AI conditions were more similar to each other, by 10.7% and 8.9% of the human-only range. The authors frame it as a social dilemma: each writer is better off, the pool gets poorer. The limit: eight-sentence stories by online participants, not professional campaigns.

Read the paper (free)
Homogenization Effects of Large Language Models on Human Creative IdeationBarrett R. Anderson, Jash Hemant Shah, Max Kreminski · ACM Creativity and Cognition · 2024

36 participants generated ideas with ChatGPT or with an alternative creativity tool. ChatGPT users produced more ideas, and more detailed ones, but different users' ideas were less semantically distinct from each other, and users felt less responsible for what they made. The limit: a small lab study with one model version, so the size of the effect is uncertain even if the direction matches Doshi and Hauser.

Read the paper (free)

Why sameness kills margins: the Hotelling model

Harold Hotelling's 1929 model puts buyers along a line of tastes and two sellers somewhere on it. Buyers pay a "transport cost" \( t \) per unit of distance to buy something less than ideal. With symmetric sellers a distance \( d \) apart and marginal cost \( c \), the equilibrium price is:

\[ p^{*} = c + t\,d \]
  1. Differentiated ads. Two brands sit far apart in buyers' minds: \( d = 0.6 \). With \( t = ₹100 \), each earns a ₹60 margin over cost.
  2. Same machine, same prompts. Their positioning drifts together: \( d = 0.1 \), margin ₹10.
  3. Identical. \( d \to 0 \), and \( p^{*} \to c \). This is Bertrand competition: no reason to choose but price.
AI makes each ad better and the whole market more alike. The margin lives in the distance between brands, and sameness deletes it.
Margin lives in the gap: p = c + t·d Before Brand ABrand B d = 0.6, margin = ₹60 After Brand ABrand B d = 0.1, margin = ₹10 tastes: 01 Illustrative, with transport cost t = ₹100
Shared tools pull brands toward the same "best" message. In Hotelling's model the price premium shrinks in proportion to the gap.

Does a polished ad still prove anything?

No. A polished ad used to be a costly signal: only a confident brand with money would spend crores on a film, so the film itself said "we will be around". When a machine makes the same polish for almost nothing, the signal stops separating strong brands from weak ones.

The separation condition

Signalling theory (Michael Spence's job-market model, and Zahavi's handicap principle in biology) says a signal is credible only if it costs a weak sender more than it gains. Let a strong brand gain \( B_H \) from being believed, a weak brand gain \( B_L \), and the signal cost \( k \).

\[ \text{Separating equilibrium needs}\quad B_L < k \le B_H \]
  • When a TV film cost ₹2 crore (an example input), a fly-by-night brand could not afford it, so \( k > B_L \) and polish meant something.
  • When the same polish costs ₹20,000 of compute (with these example inputs, not market data), \( k \ll B_L \), every brand sends it, and buyers learn to ignore it.
  • The signal migrates to whatever is still costly: real people on camera, physical events, long-running campaigns, real-world proof. That is why Coca-Cola's AI truck drew fire while a real truck on a real street does not.
A wooden replica of the Gutenberg printing press in a museum
A replica Gutenberg press. Printing made copies cheap in the 1450s; the value moved to authors, publishers and trusted names. AI is the same shift for ads.Photo: International Printing Museum, CC0, via Wikimedia Commons
When one input is free, value moves to the rest Cheap nowcopy, images, videoendless variants Scarce, so dearerattention and slotsfirst-party datadistributionbrand memory Same logic as printing in the 1450s: copies got cheap, trusted names got valuable
AI pushes one side of the beam to the floor. Everything an ad needs that AI cannot make rises in relative value.

What stays scarce, and who wins?

When one input becomes free, value moves to the inputs that stay scarce. Here is how the inputs to an ad sort out:

InputCost after "AI writes everything"Who captures value
Copy, visuals, video, variantsNear zeroNobody; it is a commodity
Ad slots and attentionHigher, as more bidders competePlatforms: Meta, Google, retail media
First-party data and outcomesHigherBrands that own customer relationships
Distribution: shelves, apps, listsHigherWhoever owns the channel
Brand memory built over yearsCannot be bought quicklyIncumbents with distinctive assets
Human taste and judgementScarcer in relative termsThe person who chooses what not to run

The pattern holds in creative testing too. When variants are free, the bottleneck becomes how many a platform can learn from; I explained how Meta's systems pick between them in Meta Andromeda and Advantage+ and how often to refresh them in creative fatigue on Meta. The bottleneck shift itself is the subject of distribution is the new bottleneck.

Sensitivity: what would break this forecast?

  • If AI raised diversity instead. Prompting for difference, or many different models, could widen \( d \). The 2024 evidence points the other way, but it is early.
  • If attention grew. More screen time or new surfaces add slots and soften the price rise; Meta's impressions grew 19% in Q1 2026 and prices rose anyway.
  • If buyers delegated to agents. Then persuasion matters less and verified data more; see what if every customer were an AI agent.
Verdict: possible, and partly here

Nothing in physics or economics stops AI from writing every ad; Meta alone reported over 8 million advertisers using its generative tools by April 2026. The consequence is the interesting part: production cost goes to zero, ads converge (10.7% more similar in the best controlled test), auction prices rise with the number of bidders (about 11% when bidders double from 10 to 20), and margins shrink with differentiation, \( p^{*} = c + t\,d \). The winners own something the machine cannot generate: attention slots, customer data, distribution and years of brand memory.

For how I use these tools in real campaigns, see my applied AI work. The earlier Indian example of AI in a famous campaign is Cadbury's "Not just a Cadbury ad".

What if next?

Questions people ask

Will AI replace advertising agencies?

AI is replacing the production part of agency work, the variants, resizes and first drafts. The parts that stay valuable are strategy, taste, real-world production and media buying judgement. Meta's stated aim of generating complete ads from a product image and a budget targets the production layer first.

Do people like AI-generated ads?

Results are mixed. Coca-Cola's 2024 AI Christmas spot scored a perfect 5.9 with System1's emotional-response test, yet drew strong public backlash about glitches and the absence of human craft. How an ad tests and how its making is perceived can diverge.

Does AI make ads less creative?

Each ad can get better while all ads get more alike. In Doshi and Hauser's 2024 study, AI ideas raised individual novelty by up to 8.1% but made stories 10.7% more similar across writers.

Will AI make advertising cheaper?

Production gets cheaper; media may not. Meta's average price per ad rose 14% in Q4 2024 and 12% in Q1 2026 year on year, while adoption of its AI creative tools grew several times over.

Sources

  1. Doshi and Hauser: Generative AI enhances individual creativity but reduces the collective diversity of novel content, Science Advances (2024) (source for 293 writers, +8.1% novelty and 10.7% similarity)
  2. Anderson, Shah, Kreminski: Homogenization Effects of Large Language Models on Human Creative Ideation (2024) (source for the 36-participant homogenisation result)
  3. AdExchanger: Q4: Meta minted money (2025) (source for 4 million advertisers, up from 1 million; +14% price, +6% impressions)
  4. PPC Land: Meta Q1 2026 results (2026) (source for 8 million+ advertisers, +12% price, +19% impressions)
  5. Adweek: How Coca-Cola's AI holiday ad went from praise to rage (2024) (source for the System1 5.9 score)
  6. HuffPost UK: Why Coca-Cola is facing backlash over its AI-generated ad (2024) (source for the criticisms and Coca-Cola's statement)
  7. afaqs!: Coca-Cola's AI-made ad stirs controversy again (2025) (source for Silverside AI, Secret Level and the five-person team)
  8. Campaign Asia: Meta aims to fully automate ad creation with AI by 2026 (source for the product-image-and-budget plan)
  9. ASCI: Half-Yearly Complaints Report 2025-26 (2025) (source for 97% of violations on digital)
  10. Lexus UK newsroom: Driven by Intuition, car by Lexus, story by artificial intelligence, camera by Oscar-winning director (2018) (the AI-scripted Lexus ad)
AI advertisinggenerative AI adscreative homogenisationHotelling modelad auctionscostly signallingMeta Advantage+
Sapta

Sapta (Saptarshi Sarkar) is a mechanical engineer turned marketer in Bengaluru and a solo touring rider. He writes the seven Saptaverse blogs. Work with him or read more of Marketblog.

Keep reading in Marketblog

From the other blogs