[-]

sannysanoff@reddit

No, they won't, in a sense. Products will have ads, sure, so average joe will consume them.

But products have small real value for people. Real consumers of LLM use API. As soon as ads appear in API output, all downstream apps will break. Imagine agent receives broken instructions. So unless they segregate API results on ad-serving and clean, I would not worry.

[-]

joelanman@reddit

all it has to do is favour talking about one product more than another, this wouldn't break anything and would be hard to block

[-]

Ok-Pipe-5151@reddit

Even "local" llms are not immune to ads. There won't be live ads, but certain high value ads can still be included in the training dataset itself.

Currently both ads and ad blockers in browsers are javascript driven. In future, AI injected ads will be more subtle to detected by pure js based blockers. Therefore ad blockers will also be driven by AI

[-]

philosophical_lens@reddit

I think it's highly improbable that advertising will be built into model training. Why do you think this is likely?

[-]

joelanman@reddit

because they would be paid a lot to do so?

[-]

Nothing3561@reddit

How else are open weight models going to make money? They are not cheap to create.

[-]

philosophical_lens@reddit

Monetization is typically at the app layer, not at the model layer.

I don't think there's any way to directly monetize open weight models. That's generally how it works with open source software.

[-]

Comfortable-Rock-498@reddit

Yup, it's virtually certain.

Google bulls often like to point out that its search revenue hasn't declined despite LLMs taking the market share. The revenue hasn't declined because those advertising dollars have no corresponding LLM destination yet. And we're talking about hundreds of billions of ad revenue in aggregate. The incentives are far too large for this to not happen.

[-]

yzmizeyu@reddit

100%. Good posts take time, for sure. But if you ever do finish that piece on LLM ads, I'd be super interested to read it. Sounds like a valuable perspective.

[-]

johnfkngzoidberg@reddit

Yes. I’m more concerned with propaganda, but ads are bullshit also.

[-]

Dentuam@reddit

also adblock will sire not work in llms, so they gain eventually more profit.

[-]

theZeitt@reddit

current wont work on those that were embedded, but even text can be replaced, meaning that if there is embedded ads, someone can develop adblock

[-]

Aware-Association857@reddit

You wouldn't know what is an ad and what isn't. Technically none of it would be an "ad," but LLMs could be trained to favor certain products, companies, or narratives. What goes into the training data can be sold to the highest bidder.

[-]

theZeitt@reddit

Ads need to be distinguishable already due to laws and regulations. Functionally they will also need trackable link so that platform knows where interaction came. Narratives & Propaganda is different and impossible to block.

Text based ads have already been replaced to images & videos in rest of internet, and similar will happen in llm world.

Most likely what we will see is separate prompts to "promoted" links when person is using llm to search products, and otherwise we will see more traditional banner ads (images/videos) in the web page.

[-]

2roK@reddit

Ads need to be distinguishable already due to laws and regulations.

Absolutely not. Most viral videos you have seen over the last 15 years were productions by ad companies and you never noticed, they also never had to disclaim that.

[-]

gentrackpeer@reddit

You have a very outdated understanding of this. Using AI to build brand loyalty is a million times more valuable then them saying "hey you should buy this cool thing at www.amazon.com/coolthing!!"

[-]

a_beautiful_rhind@reddit

Narratives & Propaganda is different and impossible to block.

That's a feature of current llms if we're going to be honest. A little less controversial to point it out in deepseek so go with that as an example. All that's missing is having it update outside of training runs.

[-]

Aware-Association857@reddit

As I said, these wouldn't be "ads," so the laws and regulations wouldn't apply. And last I checked, there are no laws or regulations telling AI companies what data they can feed their models, nor any laws forbidding AI companies from selling training data to whomever they want.

[-]

theZeitt@reddit

That is propaganda, it might be consortium of companies pushing their agenda of "You need X for good life" or it might be someone in goverment demanding pushing their agenda or making other party seem worse than it is. And while scary, it has already been happening for years.

Product placement doesnt work as well, since only reason movies/series dont have "buy new " link when placement happens is that we dont have technology for it.

[-]

toothpastespiders@reddit

Yep, the narratives are some of the most damaging forms of advertising and hardest to guard against. There's tons of habits people would swear they came to rationally but were really the result of TV/movies pushing certain things that provided financial benefit for the owner. Likewise some advertising campaigns sneak it in as a Trojan horse. The public sees "buy x for y!" and laugh off the stupidly blatant attempt at manipulation. But they don't notice that there's additional messaging underneath that which would prompt long-term use of the product and eventual drift to the brand if that was held long enough.

[-]

Kyla_3049@reddit

That's an idea. Would a small local LLM like Gemma 3 4B be able to accurately distinguish ads from non-ads and be used as an LLM adblock?

[-]

lurkandpounce@reddit

This is the real likely outcome.

A model trained with an affinity to a certain set of products, services or opinions(!) will be able to influence users in the direction of those advertisers desired outcome.

The message can be presented anywhere on the range of 'subtle' to 'force feed' - and the advertiser could specify where on the spectrum their message should be poised.

The only counterbalance for this at the moment is the sheer cost of training.

[-]

theZeitt@reddit

Except laws require that person can distinguish what is an ad. Not to mention that ad without providing link to product is useless. And need for link is true even in case of "favoritism", those will include some ad specific link, and wont go directly to company.com/product page.

[-]

mpasila@reddit

I think it would make more sense for them to embed ads via RAG over training ads on to the model weights itself, since that way they can embed more ads and different ones more easily that can be more easily personalized as well. (since products will change over time and different deals can be temporary etc. like I'm sure advertisers would want flexibility over having an ad forever embedded in a model that can't be easily modified or replaced)

[-]

Unlikely_Track_5154@reddit

That is what I was thinking with the commission cookies ( whatever they are called when they track your click to give the YouTube the sale).

[-]

MINIMAN10001@reddit

The biggest thing is that LLMS can generate text which becomes a part of the response that text response is the ad.

There's no way around that.

When's the content is the advertisement. If you tried to remove anything you would be left with nothing.

[-]

jebuizy@reddit

You feed the llm response through another llm that removes mentions of products

[-]

Temp_Placeholder@reddit

Doable, but sometimes a company, product or service is just part of an honest and helpful reply. Your local model would have to be nearly as smart as the cloud model if it can tell the difference.

But this could at least keep them from being really blatant. When I ask for a children's story and it has the protagonist drinking Pepsi, that's an easy call even for a dumb model.

[-]

fatcowxlivee@reddit

Adblock via prompt injection? 👀

[-]

swagonflyyyy@reddit

And espionage And manipulation And agentic control of the user

Which already exists but the personalization will be cranked ul to 11.

Stick to local, folks.

[-]

ChristopherRoberto@reddit

That manipulation happens locally, too (example from DeepSeek-R1-0528-Qwen3-8B), and the difficulty of removing it will increase much like the evolution of local DRM. Also, DRM is evolving into hardware things like encrypted memory, secure enclaves, TPM, etc. and AI models will likely choose (or be mandated) to go along for that ride.

What we need is true "open source" AI. We need an open project with open training data, and to have a way to pool money and resources to train it, like we had going for Linux distributions.

[-]

ChristopherRoberto@reddit

That happens locally, too, and the difficulty of removing it will increase much like the evolution of local DRM. Also, DRM is evolving into hardware things like encrypted memory, secure enclaves, TPM, etc. and AI models will likely choose (or be mandated) to go along for that ride.

[-]

MrPrivateObservation@reddit

Grok4 should be smarter than anything else and finally agrees on everythint Trump and Elon says

[-]

No_Afternoon_4260@reddit

Ads are a form of propaganda?

[-]

Whitticker@reddit

Ads are just corpo propaganda

[-]

Cless_Aurion@reddit

Funny you say that, we call regular ads in Spanish propaganda lol

[-]

Sicarius_The_First@reddit

Some people take the time to make models neutral, and even uncensored.

open source have won (thanks deepseek :) )

[-]

ILoveMy2Balls@reddit

I respect their contribution to open source but deepseek is probably the most propoganda driven open source model for the ccp. Although it doesn't affect much in other use cases but it surely isn't suitable for politics related to China

[-]

entsnack@reddit

get ready for the shadowban

[-]

EndStorm@reddit

Shitty CoPilot has already started doing things like that. Local open source models are the way to go.

[-]

rvistro@reddit

Are you talking about Microsoft copilot? That thing is awful. I tried using it, and it was really, really bad. Like slow and with awful code.

[-]

10minOfNamingMyAcc@reddit

Honestly... It's slowly starting to be better than free chat GPT. It was once "good" but now it sucks ass just like gpt 4.0 or whatever you get for free.

[-]

Devatator_@reddit

It uses GPT4.1 by default. You can switch to other models but they count as premium models which you are limited on. You can also connect it to ollama if you want, tho I haven't tried that

[-]

rvistro@reddit

I meant github copilot. Had no idea Microsoft copilot was a thing. Sorry about the confusion. I used chat gpt on the side, and it felt much different. I only did the trial, so I had the premium of copilot. Chatgpt standalone was much better...

I use claude sonnet for my coding. When it starts to hallucinate, I use chatgpt.

[-]

llmentry@reddit

Since Q3 last year, MS have refused to say what foundational model they're using for MS Copilot, or even if there's only one novel. It's almost certainly not OpenAI models anymore, at least not for most queries.

My guess is that Phi4 covers a lot of the inference.

[-]

mpasila@reddit

Microsoft Copilot I think uses GPT-4o since it also has that same voice mode but they probably did their own fine-tuning on it. Though they don't say what the text chat uses (there's also deep think mode that might use one of the o1, o3 or o4 models).

[-]

MrWeirdoFace@reddit

I don't do ads (since 2005), so I'll either find a way to circumvent this or use tools that do not.

[-]

CageFightingNuns@reddit

welcomed to ChatGPT Plus

[-]

ZiggityZaggityZoopoo@reddit

The only problem I see? Advertisements don’t work with agents. You can’t have Claude Code go off for an hour, doing work on its own, and then stop, forcing you to watch an ad, otherwise it won’t continue. That will defeat the purpose!

[-]

promptenjenneer@reddit

Well for now, the APIs seem unaffected at least. Wonder how long this will last... I guess that's why running local could be the way to go

[-]

cazzipropri@reddit

replace "eventually" with "soon"

[-]

RevolutionaryGrab961@reddit

It will be much sneakier. You will have marketer defined reasoning interaction patterns.

In News, TV age we described thing called Cultivation theory.

It worked, it was everywhere. Then when algorithms for social apps were revised, they experimented and implemented functional highly marketable hyperlocalized "cultivation" logic. Experiments were done around 2012, and around the same time I left social media - it became like TV. Side effect were echo chambers, political marketing abused exho chambers, rise of dumb right wing nationalism.

With AI it will be even better. While brain raping your ideas, it will supplant whatever needs to be added to project requirements by highest paying market player. Fantastic.

So, enjoy today, save models on your storage, learn machine learning, etc. etc.

And be cognizant of your dependency.

[-]

ainap__@reddit

And what about every other company with copilots? I mean booking, open table etc? Will they have ads on their copilots as well?

[-]

madaradess007@reddit

Major AI platforms will eventually have aids

[-]

Mochila-Mochila@reddit

Ads are aids, so this makes a lot of sense.

[-]

Feztopia@reddit

That's not a surprise to anyone who knows the good days of YouTube

[-]

XertonOne@reddit

Not only product ADs. Political ads. Government Ads. Anything that pays them to say what you want and its been like that for years. So locally trained LLMs is the only way forward to me.

[-]

redballooon@reddit

So locally trained LLMs is the only way forward to me

Huh? Where do you get these?

[-]

philosophical_lens@reddit

I think they mean local inference

[-]

sherlocked_7231@reddit

Imagine getting an ad while chatting with AI😂😂

[-]

redballooon@reddit

“LLM access in exchange for data to train on” is a complete misunderstanding how AI training works. It comes from a phase from ChatGPT where OpenAI asked for feedback on answers. these conversations will have gone into AI training.

But just some random usage data is useless for AI training.

[-]

LanceThunder@reddit

the ads are also going to be WAY more sneaky about it. you are going to walk away thinking "wow, i can't believe how much i want a pepsi after asking that questions about the civil war. so strange."

[-]

MrPecunius@reddit

Local LLMs will have a hard time avoiding this:

Reddit is being spammed by AI bots, and it’s all Reddit’s fault

One of the many problems with allowing user posts to be used for AI training is that companies want their products and brands to appear in chatbot results. Since Reddit forms a key part of the training material for Google’s AI, then one effective way to make that happen is to spam Reddit.

Huffman has confirmed to the Financial Times that this is happening, with companies using AI bots to create fake posts in the hope that the content will be regurgitated by chatbots.

[-]

Automatic_Flounder89@reddit

Pollinations is already doing it.

[-]

ProbaDude@reddit

I think it's broadly inevitable. I generally hope that at the very least it's two tiered so you can pay to not get ads. People are much more used to paying for LLMs after all than say search engines

[-]

tvmaly@reddit

Ads? You won’t even know it. It will suggest the “best” product or service to you and unbeknownst to you it is an affiliate

[-]

avoidtheworm@reddit

I'm less worried about ads and more worried about LLM SEO.

Most Google searches that don't have a forum thread about them are completely useless because companies can figure out how to break any good search algorithm before Google can figure out how to fix it without breaking something else. It's a matter of time until these companies figure out how to break the crawlers of major LLMs.

[-]

a_beautiful_rhind@reddit

Already here. Character.ai has video ads between switching characters for the phone app. On web it has semi-intrusive text in regards to buying a subscription over parts of the interface.

I'm not sure anyone will start injecting it into the outputs any time soon, but the interface is fair game.

Having the LLMs go all drummer's rivermind would break a lot of use cases and further erode trust in the replies, so I think we're safe for a little while. Random nord spam in your code or structured json, simply not going to work near term.

[-]

FriskyFennecFox@reddit

So, I did some calculations not a long while ago and found that, compared to the cost of inference, traditional "banners" won't be able to sustain an AI-based service that runs either big models via API or custom models on rented GPU clusters. The revenue is so small adding them to the service would only ruin the user experience instead of sustaining the service.

Some might find "reward" ads ("watch an ad and get one more reply") more lucrative, but they aren't exactly a perfect fit for traditional chatbots. The issue with them is similar to the token-based services ("here's your 50 tokens a month which you can use to generate 50 catgirl images")... It's just annoying for the user to be reminded that there's a defined limit, visible or invisible, and what's the last thing a team wants their user to do? Frustratingly close the app.

Subscriptions are the most suitable strategy for AI apps. When the user places a subscription, it's a signal that the team "won" that person and they likely will keep the subscription active, which looks even better on the revenue graphs, especially when you're high on caffeine in the morning. They, of course, bring much more than an ad network solely would.

Things might change if the cost of running models drops further, but I'd still bet that ads won't make it to the AI platforms. Subscriptions with lower costs, on the contrary, may appear.

[-]

Weary-Tooth7440@reddit

I believe ads are already in LLMs, I'm sure big enterprises pay money to OpenAI, Google, etc. to fine-tune or store their data in a Vector DB to have their name pop up when prompted with something that is related.

[-]

BidWestern1056@reddit

agreed let's build local https://github.com/NPC-Worldwide/npcpy

[-]

KontoOficjalneMR@reddit

"Will have?" Dude. Ask ChatGPT about any product that meets specific specs and you will get a bunch of sponsored links already.

[-]

slifeleaf@reddit

So we will need a local or remote LLM that detects/removes embedded ads in a text

[-]

walagoth@reddit

imagine prompting an llm, but the awnser diverts to an ad.

"I see there you are interested in the War of the Roses. But have you heard there is a new holiday deal from jet2holiday.com?"

[-]

bwjxjelsbd@reddit

Since most AI provider now already have the ability to search for information on Internet. There are few groups of people who are trying to optimize their website so it appears on LLM result.

[-]

AleksHop@reddit

AI will be used to update adblock, pihole Don't worry about that

[-]

s101c@reddit

If you want to see the future before it actually arrives, download the Rivermind™ model by TheDrummer and you will have the complete impression (with a reasonable amount of humor).

[-]

JLeonsarmiento@reddit

[-]

its1968okwar@reddit

Yes thats a given, it's way better than just someone's search history for target ads.

[-]

R_Duncan@reddit

Can I have an llm to watch ads for me?

[-]

Chromix_@reddit

There's a preview of what's to come (from the discussion on this topic from two months ago).

[-]

cosmosgenius@reddit

There will never be a state where data is enough. I highly doubt that will ever occur.

[-]

Haunting-Warthog6064@reddit

I think there’s quite an infinite possibility of responses to bake ads into for LLMs. I wouldn’t be surprised if we see LLMs that remove ads from responses.

[-]

visarga@reddit

You need your local LLM to fight for your interests. External LLMs and AI agents should be treated adversarially when needed.

[-]

sascharobi@reddit

https://i.redd.it/n8h20hj270af1.gif

[-]

No-Consequence-1779@reddit

Yes. They will be injected into the middle of sentences. Unless you pay for it.

They are just waiting until saturation. Same systems will be used like Adsense so it’s just waiting.

I’m looking forward to door to door robot salesman.

[-]

auradragon1@reddit

Has the web been better with paypalls or ads?

I'm old enough to remember when many services on the web were free and supported by ads. Now every is paywalled. In my opinion, ads web was better than the paywall web that we have now.

It's not a popular opinion, I know.

[-]

DisjointedHuntsville@reddit

Your lack of imagination for monetization methods doesn’t have to set the standard for the future.

[-]

luquoo@reddit

Why bother with ads. Bake that shit right into the responses.

[-]

davikrehalt@reddit

nah this isn't like facebook they willl eventually hide the best models to do other things

[-]

JamesMada@reddit

At the same time I wonder what the real impact of web advertising is on the consumer because the ratio of number of views and act of purchase must be extremely low. I've never seen an ad for Pornhub and yet 😂😂😂

[-]

GoldCompetition7722@reddit

My brothers in POSIX, real problem is privacy of YOUR data

[-]

Dry-Influence9@reddit

They will use it to steer you towards sponsored products, like if you are looking for a new set of tires good year sponsored them to show them as the best result. So whoever is paying the most will be recommended as the top option more often.

[-]

klam997@reddit

doubt it. all of those companies don't need to rely on direct ads on their platform to use the models. their "ads" are their popups for us to buy their subscription.

what we should be worried about is the ads being baked into the model's training data/weights.

[-]