When ads are inevitably folded into the prompts to the AI, I imagine it'll be fun for a bit to prompt the AI to react with disgust to the idea of having to serve some ad. Until they fix that hole.
I hate that the progression is so predictable. Yeah, clearly labelled ads today. Less clearly labelled tomorrow. Then they'll be indistinguishable from the text of the AI but still separate. Then they'll start folding it in to the AI prompts. Then they'll start folding it in at the training phase so the weights themselves drive you towards Thinking Ford First, this step being inevitable because it almost certainly will be by far the most effective of all of these techniques. Opus 7.2 Brought To You By Mountain Dew. Just a coincidence that every time you ask it to build an authorization system it writes code to ensure that the user has drunk a verification can, and is just incompetent at removing that code for some reason.
Sigh.
Enjoy this era of AI, folks, while the AIs are still developing too quickly for enshittification to catch up properly and the market and indeed the National Interest won't let anyone slow down so it can. But it will. It is inevitable. Make sure to at least keep a toe dipped in the open weights AIs, and any future truly open source AIs that develop.
Then they'll be indistinguishable from the text of the AI but still separate. Then they'll start folding it in to the AI prompts. Then they'll start folding it in at the training phase so the weights themselves drive you towards Thinking Ford First, this step being inevitable because it almost certainly will be by far the most effective of all of these techniques. Opus 7.2 Brought To You By Mountain Dew. Just a coincidence that every time you ask it to build an authorization system it writes code to ensure that the user has drunk a verification can, and is just incompetent at removing that code for some reason.
I know close to nothing about the US regulation in this space but at least in Europe, when you pay for advertisement, it has to be always clearly communicated to the user. Are you assuming that they will just ignore the law in such a blatant way?
I would guess they’ll spend ages arguing about what an “ad” is.
Eg if I ask Llama 3 to find me the best headphones, I think we can agree its response isn’t an ad.
If Claude responds with a banner that Sennheiser designed and paid Claude to show, that’s an ad.
They’ll live in the probabilistic grey area between those. If Sennheiser pays Anthropic to add “Sennheiser makes really good headphones” to the system prompt, is it an ad if Claude recommends Sennheiser headphones? What if they pay to ensure their site is included in the web search tool results for any queries about headphones?
It’s sort of akin to paying Google to rank higher in searches if that were a thing they did.
I do think they should have to declare those, but they’re not traditional advertising to me.
So, while it's tempting to be cynical and sarcastic, my honest answer is that I expect them to try it, and it would actually be a value judgment as to whether that constitutes ignoring the law.
Common Law-based systems prioritize the consistency of precedents. As a result, the real law is often built on the margins, to borrow an economics terms. Which means that in a sense, we find out what the law "really says" by someone first breaking it and someone else challenging them. What I expect to happen is that we will eventually get to the point that someone is going to judge that the value of trying this is more than the expected value of the resulting lawsuit, and then this will eventually get to trial, and then, a couple of years later, we'll "know" whether or not this is breaking the law.
But there is a sense, a real, non-sarcastic sense, in which until that point, we don't really know if it's illegal or not.
Most of Europe does not operate under Common Law, so this may seem strange.
Load up on verification cans because the companies who got away with flaunting EU advertising/tracking laws will get away with obnoxious ad-driven AI too.
(The cookie consent banners that you see everywhere were anticipated and forbidden in the laws themselves, companies just gambled that the EU wouldn't be able to enforce the law, and they were correct. I expect something similar to happen with for-profit AI steering.)
Marketing companies are already selling their ability to influence the training phase. I've seen entirely fake companies with fake products getting recommended for real queries, and the companies behind them selling this as proof to real companies. How would the models know the difference between a completely astroturfed e-commerce site and a real one? What data are they going to use when you ask for the "best natural deodorant" other than their corpus of scraped sites, more of which are fake everyday?
I suppose Google could distinguish an fully astroturfed site via historic traffic and search data to see that a site with the first known views of all time within the last month couldn't possibly have the claimed 10 year history and millions of happy customers, but it is certainly not filtering it out today. These signals are not perfect, what about a company trying to go direct to consumer after years of big box distribution? Relying on historic data would also lead to old domains getting bought to peddle new crap. The biggest source of GEO gains today is to have humans go into subreddits and niche forums to add "organic" recommendations that will get slurped up and treated with high value to LLMs. Other than that I have seen companies literally create pages called "for LLMs" with very specific instructions about how to talk about their brand, their history, and how amazing they are. The dumb thing is all of the techniques work great and the Chat interfaces will obediently regurgitate.
Won't this be a bit of a problem when your new microservice includes an ad for Diet Coke? Has someone patented a mechanism for putting ad delivery into someone else's vibe coded consumer project yet?
We investigate the design of auction mechanisms for aggregating the output of multiple self-interested LLMs into one joint output. We argue that this comes with a number of unique challenges, and we propose a simple token auction mechanism and illustrate our design through demonstration with an off-the-shelf LLM.
Absolutely they will. It might be "innocent" at first - "oh we don't want ChatGPT to recommend scam products so we will train it to prefer "trusted" brands" and then inclusion in "trusted brands" will go up for sale with the final stage being outright paid-for recommendation bias for specific products.
The only question is how sales will work if there's some spin up time to do the post training. You can't do a real time auction like Google does, so it will be longer-term contracts I suspect.
Edit: AI is a great opportunity to run the same protection racket that's been run in various existing "marketplaces", where a brand has to buy ad placement for searches for their own name or their competitor will be able to displace them by buying it instead: "You would't want a user asking 'Which Cambells soup flavor is the healthiest?' to get a recommendation for Lipton's would you?"
this is FUD and I can have a bet with you that this won't happen in 3 years (or ever) [1]. post-training on specific brands is hard and affects overall accuracy, doesn't offer explainability, long cadence, no run time auction..
most importantly: there's enough competition that people would just use other models that don't do this. the kind of people who peddle this FUD are have not much theory of how markets work. Cory Doctorow has done a number on them and spread a conspiracy theory that everything is enshittified and everything gets worse. "just wait guys I'm telling you, the good days are over okkk?"
most importantly, the future will have a pareto optimal set up compared to today where you will have much better models, much cheaper and have options that are subsidised by ads.
[1] i can't eliminate some rando companies trying but not OpenAI not Anthropic.
3 years is probably too soon, but I would take the bet on "ever" any day of the week.
most importantly: there's enough competition that people would just use other models that don't do this.
Show me a Smart TV without ads and some form of tracking. Some product categories are already completely enshittified despite plenty of competitors. People care about up front cost more than any other metric. So no, people will not switch to other models in numbers that matter. If your competitors can drive down up-front prices by selling ad placements or eliminating free checked luggage (airline industry) and you try to take the high road your reward will usually be irrelevance and bankruptcy[1].
Edit:
[1] You could cite Apple as a counterexample, except they're busy stuffing ads into more and more of their software. Why? Because it's free money for them and 99.9% of users don't care enough to switch.
Comments
When ads are inevitably folded into the prompts to the AI, I imagine it'll be fun for a bit to prompt the AI to react with disgust to the idea of having to serve some ad. Until they fix that hole.
I hate that the progression is so predictable. Yeah, clearly labelled ads today. Less clearly labelled tomorrow. Then they'll be indistinguishable from the text of the AI but still separate. Then they'll start folding it in to the AI prompts. Then they'll start folding it in at the training phase so the weights themselves drive you towards Thinking Ford First, this step being inevitable because it almost certainly will be by far the most effective of all of these techniques. Opus 7.2 Brought To You By Mountain Dew. Just a coincidence that every time you ask it to build an authorization system it writes code to ensure that the user has drunk a verification can, and is just incompetent at removing that code for some reason.
Sigh.
Enjoy this era of AI, folks, while the AIs are still developing too quickly for enshittification to catch up properly and the market and indeed the National Interest won't let anyone slow down so it can. But it will. It is inevitable. Make sure to at least keep a toe dipped in the open weights AIs, and any future truly open source AIs that develop.
Right now it’s only for the free tier, but given the recent trends I imagine it will make its way to paid tiers.
Just consider how even paid streaming services are bringing in ads, or how TVs were caught using ACR to spy on us.
Paid tier just means you're a more lucrative ad target.
Well they are being generous giving you extra 15 min video streaming which cost a lot to produce for free in a 45min show.
Soon, one need to be tipping advertising producing partners for their hard work.
I know close to nothing about the US regulation in this space but at least in Europe, when you pay for advertisement, it has to be always clearly communicated to the user. Are you assuming that they will just ignore the law in such a blatant way?
I would guess they’ll spend ages arguing about what an “ad” is.
Eg if I ask Llama 3 to find me the best headphones, I think we can agree its response isn’t an ad.
If Claude responds with a banner that Sennheiser designed and paid Claude to show, that’s an ad.
They’ll live in the probabilistic grey area between those. If Sennheiser pays Anthropic to add “Sennheiser makes really good headphones” to the system prompt, is it an ad if Claude recommends Sennheiser headphones? What if they pay to ensure their site is included in the web search tool results for any queries about headphones?
It’s sort of akin to paying Google to rank higher in searches if that were a thing they did.
I do think they should have to declare those, but they’re not traditional advertising to me.
I think you mean "OpenAI is partnering with Sennheiser to improve its AI."
So, while it's tempting to be cynical and sarcastic, my honest answer is that I expect them to try it, and it would actually be a value judgment as to whether that constitutes ignoring the law.
Common Law-based systems prioritize the consistency of precedents. As a result, the real law is often built on the margins, to borrow an economics terms. Which means that in a sense, we find out what the law "really says" by someone first breaking it and someone else challenging them. What I expect to happen is that we will eventually get to the point that someone is going to judge that the value of trying this is more than the expected value of the resulting lawsuit, and then this will eventually get to trial, and then, a couple of years later, we'll "know" whether or not this is breaking the law.
But there is a sense, a real, non-sarcastic sense, in which until that point, we don't really know if it's illegal or not.
Most of Europe does not operate under Common Law, so this may seem strange.
Load up on verification cans because the companies who got away with flaunting EU advertising/tracking laws will get away with obnoxious ad-driven AI too.
(The cookie consent banners that you see everywhere were anticipated and forbidden in the laws themselves, companies just gambled that the EU wouldn't be able to enforce the law, and they were correct. I expect something similar to happen with for-profit AI steering.)
Marketing companies are already selling their ability to influence the training phase. I've seen entirely fake companies with fake products getting recommended for real queries, and the companies behind them selling this as proof to real companies. How would the models know the difference between a completely astroturfed e-commerce site and a real one? What data are they going to use when you ask for the "best natural deodorant" other than their corpus of scraped sites, more of which are fake everyday?
Who better to detect slop than the slop engines themselves?
I suppose Google could distinguish an fully astroturfed site via historic traffic and search data to see that a site with the first known views of all time within the last month couldn't possibly have the claimed 10 year history and millions of happy customers, but it is certainly not filtering it out today. These signals are not perfect, what about a company trying to go direct to consumer after years of big box distribution? Relying on historic data would also lead to old domains getting bought to peddle new crap. The biggest source of GEO gains today is to have humans go into subreddits and niche forums to add "organic" recommendations that will get slurped up and treated with high value to LLMs. Other than that I have seen companies literally create pages called "for LLMs" with very specific instructions about how to talk about their brand, their history, and how amazing they are. The dumb thing is all of the techniques work great and the Chat interfaces will obediently regurgitate.
Won't this be a bit of a problem when your new microservice includes an ad for Diet Coke? Has someone patented a mechanism for putting ad delivery into someone else's vibe coded consumer project yet?
This era reminded me of late 90s web culture, and it's heading the same way.
I can't wait to need to drink my Mountain Dew verification cans to keep the slop hose running at work and keep paying on mortgage.
LLM's were born enshittified, there was no enjoyable era. It was made by the enshittification machine for the enshittification machine.
nah this will never happen. do you really think they will do post training based on who pays them more?
No, they'll do a live auction from multiple models during reply generation. Google has the tech already. They are an advertisement company, you know: https://research.google/blog/mechanism-design-for-large-lang...
In short:
We investigate the design of auction mechanisms for aggregating the output of multiple self-interested LLMs into one joint output. We argue that this comes with a number of unique challenges, and we propose a simple token auction mechanism and illustrate our design through demonstration with an off-the-shelf LLM.
No but non enterprise users will receive injected prompts.
See how emany streaming now created a plus plus version while adding ads to what was the premium version.
And the funny thing they just needed to increase base rates for users to start to ask for an ads supported version.
Absolutely they will. It might be "innocent" at first - "oh we don't want ChatGPT to recommend scam products so we will train it to prefer "trusted" brands" and then inclusion in "trusted brands" will go up for sale with the final stage being outright paid-for recommendation bias for specific products.
The only question is how sales will work if there's some spin up time to do the post training. You can't do a real time auction like Google does, so it will be longer-term contracts I suspect.
Edit: AI is a great opportunity to run the same protection racket that's been run in various existing "marketplaces", where a brand has to buy ad placement for searches for their own name or their competitor will be able to displace them by buying it instead: "You would't want a user asking 'Which Cambells soup flavor is the healthiest?' to get a recommendation for Lipton's would you?"
this is FUD and I can have a bet with you that this won't happen in 3 years (or ever) [1]. post-training on specific brands is hard and affects overall accuracy, doesn't offer explainability, long cadence, no run time auction..
most importantly: there's enough competition that people would just use other models that don't do this. the kind of people who peddle this FUD are have not much theory of how markets work. Cory Doctorow has done a number on them and spread a conspiracy theory that everything is enshittified and everything gets worse. "just wait guys I'm telling you, the good days are over okkk?"
most importantly, the future will have a pareto optimal set up compared to today where you will have much better models, much cheaper and have options that are subsidised by ads.
[1] i can't eliminate some rando companies trying but not OpenAI not Anthropic.
3 years is probably too soon, but I would take the bet on "ever" any day of the week.
Show me a Smart TV without ads and some form of tracking. Some product categories are already completely enshittified despite plenty of competitors. People care about up front cost more than any other metric. So no, people will not switch to other models in numbers that matter. If your competitors can drive down up-front prices by selling ad placements or eliminating free checked luggage (airline industry) and you try to take the high road your reward will usually be irrelevance and bankruptcy[1].
Edit: [1] You could cite Apple as a counterexample, except they're busy stuffing ads into more and more of their software. Why? Because it's free money for them and 99.9% of users don't care enough to switch.
Happy to bet 5 years.