Fascinating paper by Alex Chan (NBER, 2026) on what generative AI does to the open web. The web runs on a simple trade: publishers make things worth reading, search sends readers to them, and the visits pay for the work. The visits do something else as well: every click, link, and citation tells search engines which sources to trust. LLMs put an end to the trade: the assistant reads the page and gives you the answer, and the source never gets the visit, the money, or the signal. Chan proves something troubling: even if the AI is accurate and the users are rational, the expensive kinds of information start to disappear. Reporting, expert testing, work that requires someone to go and look. And the platforms will not stop it, because each one profits from keeping you on its own screen while the verified knowledge it depends on quietly runs out. This looks like a publishers’ problem. Governments do not lose a benefits portal when ad revenue falls. Their content is paid for by budgets and required by law. The collapse Chan describes seems, at first, not to affect governments.
But we are left with a public-sector alignment problem. When the internet stops being a place people browse, nobody will bother with a government portal anyway, given how bad most of them are. They ask an assistant, and the assistant answers for the state. The website is still sitting there with its poor UX and its failed life-events design. It is just not where the citizen goes anymore. It is a back end now, something an AI reads on the citizen’s behalf, or gets wrong on the citizen’s behalf. Watch what happens when it gets a tax deadline wrong, or a benefit eligibility, or a public health question. The state did not write that answer and cannot fix it once the citizen has it. The firm running the assistant produced it and calls itself a pipe, not the author. The citizen followed the guidance and got hurt. The mistake has no owner, so it lands on all of them. Lawsuits, anyone?
Governments and the firms building these assistants both have a stake in sorting out public-sector information alignment before it turns systematic. Governments can publish their services as authoritative, machine-readable, signed statements of what each one actually is. The firms can treat those signed sources as ground truth for anything said in a government’s name. Neither move is expensive, and both are cheaper than a single wrongful-harm lawsuit.