AI Won't Kill Us
(but Ad-Supported AI Might)
If the future of AI depends on an ad-based business model like the current internet, it will likely go very badly for society.
When future historians write of our era, they will likely draw a straight line from Edward Bernays through to the collapse of American experiment. It is easy to write the first part of that history already. Once we had too much stuff to sell, advertising surged in importance, leading to the post-war decades of generating endless slop, mostly on television, purely for the sole purpose of injecting ads into peoples brains and manipulating them into buying things they mostly didn’t need.
While this was mostly benign early on (who cares which brand of dish soap or soda people buy), it set up the playbook for mass manipulation within a democracy, gradually replacing media as a source of information with entertainment that may or may not contain some information. We learned quickly that this entertainment usually galvanized attention better than reality and that attention = more ad revenue, leading to a few decades of mostly mindless sitcoms, soap operas, and game shows. Again, this was relatively benign until we discovered that we could get even more engagement if we combined the salience of real world threats with entertainment elements, giving rise to the era of news as entertainment (CNN, Fox, etc) in the 80s and 90s (see Postman’s Amusing Ourselves to Death for more).
The rise of the internet was in many ways a revolt against this new infotainment world, built by nerds with mostly noble goals of sharing information freely w/o the corruption of the ad-financed gatekeepers keeping the news shallow and stupid. And this worked for a while. We got open protocols and the other key elements that allowed us to bypass those corporate and government bottlenecks but, in our techno-optimistic naiveté, we forgot about finding a business model. So, first we got popup ads but those were so badly done that they quickly failed. That should have been the warning shot but it was ignored until Google finally found the perfect internet-native ad format with AdSense and eventual acquisition of DoubleClick. This basic algorithmic concept was copied by all and we were promised contextually relevant ads that we would “actually want to see” over and over again for the next several years.
In reality, we instead got several large companies that figured out, surprise, that attention = more ad revenue. More importantly, they figured out that manipulating their users to spend more time on site, viewing and clicking ads, was more profitable than anything else they could do - far, far more profitable (Meta made $162.4 billion on ads in 2024). Several books can be written on this transformation but, in short, it led to the current landscape of a few massive walled gardens where companies explicitly manipulate billions of people every day into performing the preferred actions that have led to their trillion dollar valuations.
At this point, one can maybe still make the argument that the ad-supported internet has made information widely and freely available and so was worth the trade-off for dysfunctional governance, failed and failing institutions, widespread distrust, an explosion of disinformation, misinformation, and explicit propaganda, and the replacement of our civil discourse and politics with whichever random characters figure out how to best manipulate the algorithms designed for monetization. Oh, and a populace so dopamine addicted that we’re now turning everything in existence into some form of gambling.
One could make that argument but only by blindly and willfully ignoring that fact that we did not actually need those companies making those kinds of profits to get the kind of information distribution we now have (or probably much better), but that’s a different essay.
Whether or not it was worth it, we’re here now and all that profligate ad-supported information dissemination created the internet-scale data that has led to the current crop of LLMs. It is still too early to tell how far this form of AI may enable us to advance but the promise is still there. Even just by wisely implementing the current models, we may be able to revolutionize many industries simply by doing much more, much faster. And it is clear we have not reached their limits, especially in regards to the more difficult areas of real science and research (and by “difficult” I mean “less profitable over the 12 months”).
Unfortunately, the internet had the same, if not more, promise in the early days. We could have built a massive, truth-seeking collective intelligence and already solved so many problems we still face. We never truly even scratched the surface of what could be done. There are multiple reasons (copyright, IP, international law, etc) but the big one was that it was just much easier and so, so much more profitable to sell ads. And because everything in capitalism is ruled by power laws, most the capital got dumped into dopamine slop machines that churned out extraordinary ad revenue year after year.
So, yes, LLM-based AI does have tremendous potential but the last 100 years of history and every reading of capitalism since the dawn of the internet suggests that our default path will be to mostly ignore that potential and instead pursue the proven and extremely profitable path of manipulating billions of people in order to sell them ads and propaganda. More importantly, the internet pales in comparison to the level of manipulation that can be achieved with an LLM, making that path all but inevitable.
Given the damage already done by ads and the potential damage that could be done with even more effective and profitable methods, the most important technical challenge of today might be simply finding a way to keep ads away from AI, to find a better business model to replace advertising. If we don’t, then the inevitable will happen and we’ll not only again lose out on all the progress we could have achieved but we’ll create an information ecosystem so hopelessly corrupted that it’s unlikely our institutions will survive long.