Stay informed with free updates

The writer is professor emeritus of computer science and engineering at the University of Washington and author of ‘The Master Algorithm’

The internet is in trouble. Thanks to artificial intelligence, traffic to websites is falling. The advertising-supported funding model on which the entire sector rests is at risk.

Content providers like news organisations and social media platforms make money when you go to their website and click on an advert. They receive a cut of the advertiser’s payment to Google or Meta, the tech companies that placed the ad. But with generative AI you can talk to a chatbot — there is no need to visit the websites used to construct its answers. The irony is that if content providers die out, AI chatbots will be starved of new content and users may lose interest, creating a death spiral in which everybody winds up worse off.

This looming crisis creates a rare opportunity to rethink the way the internet is funded and come up with a better solution. That solution could make use of AI’s own construction. While a chatbot’s output may seem like an opaque mash-up of its sources, we in fact know how to quantify exactly how much each source contributed. The key to this is the algorithm that AI systems are built on: backpropagation, known colloquially as backprop.

Backprop was designed to solve what we in AI call the credit assignment problem. When a network that processes data using billions of parameters sees an image of a dog and says that it’s a cat, we need to know how much each parameter must be changed so that it will do better next time. Backprop’s answer is to wiggle each one just a little, see how the output changes and set the parameter to a new improved value.

This system could also be used to solve a different assignment problem: calculating who should be paid when someone uses AI.

If an internet user paid, say, 10 cents for a chatbot’s response to their question, what fraction of that should go to each source? Every parameter in the network is a sum of contributions from all the inputs it was trained on. Backprop tells us how much came from each one. That proportion, minus the chatbot owner’s cut, could be calculated and paid out to the providers. This can be done efficiently using the computing infrastructure that is already in place.

Most likely, however, users won’t be paying per answer, except in cases where the answer is highly valuable and correspondingly expensive to produce. Instead, they could have a chatbot subscription. The revenue from this would be apportioned to content providers, in the same way that Spotify apportions revenue to song creators. And the value of a chatbot subscription would climb as AI becomes more essential to running our lives. A 2018 MIT study found that the average internet user thought Google search was worth $17,000 a year (though there’s no evidence they would pay this much). The value of AI is likely to be even higher.

Of course, in this system the payment to a content provider from a single chatbot answer would be a fraction of a cent. But aggregated over millions of answers it would quickly become a far larger source of income than many providers receive today from advertising. And the same backprop process can be used by AI chatbots to request new, high-quality content from providers when it finds gaps in its ability to produce what a user requests. The chatbot can even rank requests by how much income they’re likely to produce, helping content providers to make the best use of their time.

This dynamic would change the economics of the internet.

Content providers would no longer need to compete for scarce human attention as they do today. Instead of incentives for producing clickbait, they’d be rewarded for producing content that chatbots find useful. We would have higher-quality content for the same reason streaming services tend to produce higher-quality content than broadcast TV: the incentive is to make a subscription valuable.

In the new ecosystem, high-quality information providers will be that most valuable of all things: a trusted source of facts. And they need not remain invisible; their logos and links can appear alongside the chatbot-assembled content they contributed to, encouraging users to check them out.

Why haven’t tech companies told us that this system is possible? Maybe they haven’t realised it yet. Or maybe they’re not telling us because it’s not to their benefit at this stage in the game. Right now, AI companies want to pay as little as possible to the providers whose content trains their models. That is one reason why they like the current ad-funded business model. The rest of us would benefit from change. The sooner we move to a backprop-based internet economy the better.