The Internet is Eating Itself, and AI is Holding the Fork

When AI summarizes the web instead of sending you to it, the economics of online information quietly unravel.

AI search results like Google’s AI Overview are easy to like. Instead of dodging ads, closing pop-ups, scanning a half-dozen pages, and stitching together an answer yourself, you can often get a speedy AI-generated answer at the top of the page. But frictionless is rarely free: this seamless experience masks a profound economic tension that threatens the very source material AI relies upon. In the new working paper “AI and the Collapse of the www,” HBS AI Institute Associate Alex Chan develops a rigorous, theoretical market-design framework to analyze this potential crisis. Chan’s research shows that even with rational users, accurate AI information, and hardworking publishers, the current internet model could be facing structural collapse. 

Key Insight: The Missing Market

“The AI interface can appropriate current answer value without fully paying for the future reproduction value and source-level information value of publisher traffic.” [1]

Chan explains that the open web has long operated through a straightforward exchange: publishers produce content, search engines send users to that content, and those visits accomplish two things simultaneously. First, revenue, from ads, subscriptions, and affiliate links. Second, information, through clicks, return visits, links, and corrections that help future users and ranking systems find the best sources. Chan calls these the revenue event and the measurement event. Generative AI disrupts both: when the reader never arrives at the source, neither event triggers. Chan calls the result a missing market: the AI system captures the value of today’s answer without paying for the future value of reproducing that content. The scarce inputs behind reliable information, like reporting, testing, verification, expert judgment, and maintenance, do not reproduce themselves for free.

Key Insight: The Reproduction Number

“AI diversion deletes source-level observations.” [2]

Today’s internet search-result content creates immediate visits, but also durable attention capital like subscribers, repeat audiences, reputation, and search authority. Those assets help generate tomorrow’s traffic, which helps finance tomorrow’s content. Chan formalizes this dynamic with an open web reproduction number, a measure of whether a site’s traffic, revenue, links, reputation, and subscriber base should be enough to keep costly information production running. If that number is too low, there isn’t enough money or attention being sent to original sources to maintain production levels. What drives the reproduction number downward? Primarily it’s AI diversion, when an AI platform chooses to answer users directly on its own interface instead of passing along traffic to the publishers who created the original content. Chan argues that AI platforms have narrow incentives and engage in over-diversion, keeping more users inside the AI answer interface and sending fewer of them along to publishers. The result, in Chan’s model, is a predictable downward spiral leading to the collapse of the entire system.

Key Insight: When the Clicks Disappear

“Even if an AI answer is useful, answer-level feedback need not identify which source created the useful information, which source was original, which source was stale, or which source corrected an error.” [3]

Chan shows that AI diversion doesn’t uniformly reduce content: topics with high fixed costs relative to direct-click revenue have the greatest risk. That profile maps almost exactly onto the kinds of information society most struggles to fund privately like local news, minority-language content, and long-tail investigative journalism. AI search can also reduce source diversity within topics. If AI systems concentrate attention on a smaller set of sources, they may make discovery less varied. This can weaken web search itself and lead to a negative feedback loop: if conventional search gets worse, users will have more reason to rely on AI answers, which will further reduce producer visits. In Chan’s tipping model, once search quality falls below a threshold, the system may not generate enough non-AI activity to recover. 

Key Insight: The Abundance Paradox

“The observable web becomes bigger and thinner: more pages, fewer costly facts.” [4]

Generative AI also slashes the cost of producing pages because raw page volume and costly human information are simply not the same thing. A synthetic page can be cheap precisely because it contains no new reporting, testing, verification, or local knowledge. Chan identifies an abundance paradox: if publishing costs fall toward zero, synthetic pages can multiply without limit, even as original reporting and maintenance collapse. This can also dilute ranking systems, especially when search lacks trust anchors or protected exposure for verified, genuine sources. Human sources get less exposure; less exposure means less revenue and fewer new human links; fewer new human links means ranking systems continue to degrade.

Key Insight: Rebuilding the Market

“The revenue and measurement events have to move to the new point of attention.” [5]

Chan’s core prescription is actually not to restrict AI answers, but to move the revenue and measurement events to wherever user attention now lands. This includes several mechanisms like visitor-replacement royalties: paying publishers for the visits AI displaces. Another is keystone-topic weights: if compensation budgets are limited, money should flow to the topics that most strengthen the entire reproduction system (such as local news and reporting), rather than to whatever happens to be getting the most clicks at a particular moment (such as low-cost click-baity reposting of news). Chan notes that implementation would require shared data access, common provenance standards (rules for tracking where information came from, how it was used, and how that use can be verified), human-information audits (checks designed to verify that a source actually originated in a real human, institution, etc.), and governance structures.

Why This Matters

Business leaders and executives might recognize some of what Chan is describing through the classic tragedy of the commons. The open web’s stock of human information has functioned as a shared resource, mostly free to anyone with an internet connection, and now to any AI system trained on or retrieving from it. Like a shared fishery or public aquifer, it replenishes only if participants take less than what it needs to regenerate. Meanwhile, an over-harvester reaps the immediate resource rewards but does not bear the full long-term cost of depleting it. The solution to a commons problem lies in designing institutions that use prices, rights, and royalties to align private incentives with the reproduction of the shared resource. That is exactly what Chan’s market design prescriptions attempt to do, and why they deserve serious attention from anyone who depends, as almost every business does today, on the web remaining worth searching. 

Bonus

AI search may make information feel more abundant, but this research shows why abundance can create fragility. Another part of that abundance is AI’s potential to expand coverage and productivity, and another aspect of system fragility is how the expanded content will be less informative without human expertise remaining in the loop. For another look at how generative AI is reshaping the quantity, quality, and economics of content, check out The AI Content Problem. 

References

[1] Chan, Alex, “AI and the Collapse of the www,” Harvard Business School Working Paper, No. 26-088 (June 2026): 2-3.

[2] Chan, “AI and the Collapse of the www,” 3.

[3] Chan, “AI and the Collapse of the www,” 3.

[4] Chan, “AI and the Collapse of the www,” 31.

[5] Chan, “AI and the Collapse of the www,” 44.

Meet the Authors

Watch a video version of the Insight Article here.