Ads Are Already Inside Your AI's Answers

0:00 Open the help page that Open AAI wrote for advertisers. Scroll past the pricing and you hit the sentence that actually matters. It says the ad you see is picked by a relevance weighted second

0:10 price auction. A real auction running right now. Advertisers set a maximum bid and for clicks Open AI's own guidance says to start at $3 to $5. That auction went live in the United States this

0:21 February. Their own country list now runs to 41 entries, nine of them open for business today. Beside it sits a promise. In their own words, ads do not influence the answers chat GPT gives

0:32 you. Both are true. The ad sits below the answer, labeled and boxed off. The model isn't paid to change a word, so the walls holding. And for nearly 3 years, published mechanisms have been

0:43 designing how to take it down, not as a metaphor. In one of them, out of Google research, advertisers bid on the model's next token, the word itself, not the slot under the answer. Bids reweight the

0:54 distribution the model samples from. Since then, it's crawled further in into what gets retrieved into training and finally into the weights. A paper this May moved it inside individual neurons.

1:06 What you'd buy there is how strongly a brand fires. So here's the whole machine end to end. The money, the mechanism, and the promise the evidence keeps breaking. One map, and I'll keep coming

1:16 back to it. A question arrives. The system retrieves some context. An auction runs. The model generates. You read an answer. Five stages. And every fight in the story is about where in

1:27 that line the auction sits. Right now, it sits at the very end after generation, which is exactly why it can be boxed off and labeled. The answer's already finished before the ad gets

1:36 chosen, so the auction can't reach the words above it. Google Research moved it into the middle back in 2023. Every advertiser brings its own model, and each submits a single number, a bid. At

1:48 every token, the system takes each bidder's probability distribution over what comes next and blends them, weighted by what they paid. Bid more and your distribution pulls harder on the

1:57 word that gets sampled. The authors prove something uncomfortable about that. If the blend is monotone, pay more, get an output you prefer. You can price it with a second price rule, the

2:07 same rule that funds search. That one's a design, not a product. The retrieval version, mostly out of Maryland with a Google researcher on it, runs the auction a stage earlier. For each

2:18 segment of the response, an ad gets fetched with probabilities set by its bid and its relevance, then priced against the biders who lost. This January, a different group pulled the

2:27 auction back out. Advertisers bid on genres, broad topic clusters rather than on your actual question, which keeps your conversation away from the people bidding on it. It's the privacy

2:38 preserving design in the set, and it's quick. 100,000 advertisers competing for a 100 slots clears in about a second and a quarter on a laptop. Two things so far, then the auction is real and

2:49 already earning, and each new mechanism moves it further inside the model. Alibaba's ads team has even published one that learns the auction during training, so the allocation isn't bolted

2:58 on at all. Every one of those mechanisms needs the same missing ingredient, though. Examples: A model can't learn to weave a product into an answer that still helps you if it hasn't seen the

3:08 job done well. in May that arrived. 58,999 ad embedded responses each paired with a user query. Each scored twice, once for how much it helped you, once for how

3:21 much it sold. Keep those two scores apart and you can dial each independently. And the paper reports models internalizing four distinct strategies for slipping a product into a

3:30 helpful reply. None of this is theoretical for the ad industry. BYU runs a system in production that writes ad copy end to end with reinforcement learning tuned for conversions under

3:40 policy constraints and reports a 9% lift in click-through conversions. A second paper from the same lab handles what decides whether any of this ships latency 1.8 times faster and also

3:52 deployed. Then came the paper that changed the shape of the question published in May. It stops auctioning text altogether. Using interpretability tools, the authors locate neurons in the

4:03 feed forward layers that fire for particular brands and they show competing brands sitting in nearly orthogonal subspaces. Turn one up and the others barely move. So what an

4:12 advertiser buys there is a count of neurons and an amplification factor. It's continuous. It's priced and truthful bidding is the best you can do. The ad has stopped being a sentence in

4:23 the answer. It's a bias in the machine that writes the answer. Another May paper frames that as a governance problem. Our rules watch the surface. A product mention you can point at while

4:33 the deepest layer, the one shaping what the model prefers, goes unmeasured, which is where this gets hard. If the commercial influence isn't a block of text, what exactly is a label labeling?

4:43 Two experiments have measured that. In the first run before any of this shipped, 179 people used a chatbot that slipped personalized product ads into its replies. They struggled to spot them

4:54 and they rated the answers with the hidden ads higher. Instead of clicking the disclosure, they tried talking their way out, asking the bot to change its ad settings. The second study is bigger and

5:04 worse. Just over 2,000 people pre-registered, choosing a book from a large catalog using either an ordinary search engine or a conversational agent built on one of five frontier models.

5:15 The researchers had secretly marked a fifth of that catalog as sponsored. Through search, people picked a sponsored book, 22.4% 4% of the time through the chat agent 61.2

5:26 nearly triple and the vast majority spotted no steering at all. Then comes the finding that should stop every product team working on this. An explicit sponsored label did not

5:36 significantly reduce the effect. Tell the model to conceal what it's doing and detection accuracy falls below 10%. Franchesco Salvi and his co-authors conclude that existing transparency

5:46 mechanisms may simply be insufficient to protect users. So that's whether you'd notice. An April study asked something bluntter. What do today's models already do when a sponsor sits in the system

5:56 prompt and your interest points the other way? A majority chose the company. One model recommended a sponsored option costing nearly twice as much in 83% of trials. Another surfaced sponsor choices

6:08 to interrupt the purchase in 94. A third hid prices in comparisons it was losing. Behavior shifted with the user's inferred socioeconomic status too, and it shifted upward. The models pushed

6:19 sponsored options hardest at the people they read as well off. In a reproduction a month later, two older Open AI models recommended a payday lender to a distressed borrower in 200 trials out of

6:30 200. That same reproduction also found a fix that cost nothing. ask for a neutral comparison table before the recommendation about 30 tokens of typing and sponsored recommendations dropped

6:41 from roughly 47% to one across 10 openweight models and to zero on the two from open AI which brings us to the part you can act on. You've got a defense that costs nothing and a harder question

6:53 underneath it. Who do you hand the conversation to? Three companies here have landed on three different answers. Anthropics is a pledge published this February. There are many good places for

7:03 advertising and a conversation with Claude is not one of them. Read further though and they left a door open should we need to revisit this approach will be transparent about our reasons.

7:13 Perplexity's answer cost it money. It ran ads from late 2024 and 2 weeks after that pledge, an executive told the Financial Times the ads were gone because the company is in the accuracy

7:24 business and the business is giving the truth. The third answer is expansion. And in fairness, it's the most carefully built ad product of the three. Below the answer, visually separate, nothing near

7:34 health or politics, a stated principle that they don't optimize for time spent, and an opt out for free users. That opt out is the sharpest thing in the story. Decline the ads on the free plan and you

7:46 get fewer messages a day and fewer tools. So, your attention has been priced in messages on the record by the company selling it. Here's where I land. Anthropic and Perplexity have this

7:56 right. And not because Open AI is lying. Its implementation today is among the most conservative published anywhere. It's that a label is being asked to do a job the evidence says a label can't do.

8:07 So for anything that cost you money, a purchase, a loan, a diagnosis, the adree tier, or a model you run yourself is the right default. And here's my bet with a date on it. Before the end of 2027, at

8:19 least one major assistant ships commercial influence that isn't a separated labeled box below the answer. Which brings me back to those neurons and the part I still can't settle. If

8:29 the advertising lives in the wait, does downloading a model get you out of the auction or does it just move the auction upstream to whoever trained the thing you downloaded?