You're assuming ads would be subtly worked into the answers. There's no reason it has to be done that way. You can also have a classic text ads system that's matching on the contents of the discussions, or which triggers only for clearly commercial queries "chatgpt I want to eat out tonight, recommend me somewhere", and which emits visually distinct ads. Most advertisers wouldn't want LLMs to make fake recommendations anyway, they want to control the way their ad appears and what ad copy is used.
There's lots of ways to do that which don't hurt trust. Over time Google lost it as they got addicted to reporting massively quarterly growth, but for many years they were able to mix in ads with search results without people being unhappy or distrusting organic results, and also having a very successful business model. Even today Google's biggest trust problem by far is with conservatives, and that's due to explicit censorship of the right: corruption for ideological not commercial reasons.
So there seems to be a lot of ways in which LLM companies can do this.
Main issue is that building an ad network is really hard. You need lots of inventory to make it worthwhile.
There are lots of ways that advertising could be tied to personal interests gleaned by having access to someone's ChatBot history. You wouldn't necessarily need to integrate advertisements into the ChatBot itself - just use it as a data gathering mechanism to learn more about the user so that you can sell that data and/or use it to serve targetted advertisements elsewhere.
I think a big commercial opportunity for ChatBots (as was originally intended for Siri, when Apple acquired it from SRI) is business referral fees - people ask for restaurant, hotel etc recommendations and/or bookings and providers pay for business generated this way.
Right, referral fees is pay-per-click advertising.
The obvious way to integrate advertising is for the LLM to have a tool to search an ad database and display the results. So if you do a commercial query the LLM goes off and searches for some relevant ads using everything it knows about you and the conversation, the ad search engine ranks and returns them, the LLM reads the ad copy and then picks a few before embedding them into the HTML with some special React tags. It can give its own opinion to push along people who are overwhelmed by choice. And then when the user clicks an ad the business pays for that click (referral fee).
> You're assuming ads would be subtly worked into the answers. There's no reason it has to be done that way.
I highly doubt advertisers will settle for a solution that's less profitable. That would be like settling for plain-text ads without profiling data and microtargeting. Google tried that in the "don't be evil" days, and look how that turned out.
Besides, astroturfing and influencer-driven campaigns are very popular. The modern playbook is to make advertising blend in with the content as much as possible, so that the victim is not aware that they're being advertised to. This is what the majority of ads on social media look like. The natural extension of this is for ads to be subtly embedded in chatbot output.
"You don't sound well, Dave. How about a nice slice of Astroturf pizza to cheer you up?"
And political propaganda can be even more subtle than that...
There's no reason why having an LLM be sly or misleading would be more profitable. Too many people try to make advertising a moral issue when it's not, and it sounds like you're falling into that trap.
An ideal answer for a query like "Where can I take my wife for a date this weekend?" would be something like,
> Here are some events I found ... <ad unit one> <ad unit two> <ad unit three>. Based on our prior conversations, sounds like the third might be the best fit, want me to book it for you?
To get that you need ads. If you ask ChatGPT such a question currently it'll either search the web (and thus see ads anyway) or it'll give boring generic text that's found in its training set. You really want to see images, prices, locations and so on for such a query not, "maybe she'd like the movies". And there are no good ranking signals for many kinds of commercial query: LLM training will give a long-since stale or hallucinated answer at worst, some semi-random answer at best, and algorithms like PageRank hardly work for most commercial queries.
HN has always been very naive about this topic but briefly: people like advertising done well and targeted ads are even better. One of Google's longest running experiments was a holdback where some small percentage of users never saw ads, and they used Google less than users who did. The ad-free search gave worse answers overall.
Wouldn't fewer searches indicate better answers? A search engine is productivity software. Productivity software is worse when it requires more user interaction.
Also you don't need ads to answer what to do, just knowledge of the events. Even a poor ranking algorithm is better than "how much someone paid for me to say this" as the ranking. That is possibly the very worst possible ranking.
Google knows how to avoid mistakes like not bucketing by session. Holdback users just did fewer unique search sessions overall, because whilst for most people Google was a great way to book vacations, hotel stays, to find games to buy and so on, for holdback users it was limited to informational research only. That's an important use case but probably over-represented amongst HN users, some kinds of people use search engines primarily to buy things.
How much a click is worth to a business is a very good ranking signal, albeit not the only one. Google ranks by bid but also quality score and many other factors. If users click your ad, then return to the results page and click something else, that hurts the advertiser's quality score and the amount of money needed to continue ranking goes up so such ads are pushed out of the results or only show up when there's less competition.
The reason auction bids work well as a ranking signal is that it rewards accurate targeting. The ad click is worth more to companies that are only showing ads to people who are likely to buy something. Spamming irrelevant ads is very bad for users. You can try to attack that problem indirectly by having some convoluted process to decide if an ad is relevant to a query, but the ground truth is "did the click lead to a purchase?" and the best way to assess that is to just let advertisers bid against each other in an auction. It also interacts well with general supply management - if users are being annoyed by too many irrelevant ads, you can just restrict slot supply and due to the auction the least relevant ads are automatically pushed out by market economics.
The issue is precisely that "did the click lead to a purchase" is not a good target. That's a target for the advertiser, and is adversarial for the user. "Did the click find the best deal for the user (considering the tradeoffs they care about)" is a good target for the user. The winner in an auction in a competitive market is pretty much guaranteed to be the worst match under that ranking.
This is obvious when looking at something extremely competitive like securities. Having your broker set you up with the counterparty that bid the most to be put in front of you is obviously not going to get you the best trade. Responding to ads for financial instruments is how you get scammed (e.g. shitcoins and pump-and-dumps).
You can't optimize for knowing better than the buyer themselves. If they bought, you have to assume they found the best deal for them considering all the tradeoffs they care about. And that if a business is willing to pay more for that click than another, it's more likely to lead to a sale and therefore was the best deal, not the worst.
Sure, there are many situations where users make mistakes and do some bad deal. But there always will be, that's not a solvable problem. Is it not the nirvana fallacy to describe the potential for suboptimal outcomes as an issue? Search engines and AI are great tools to help users avoid exactly that outcome.