For two years I have been in rooms where the filter bubble worry about AI was treated as a settled fact. Content strategy got anchored to it. Internal search projects got shelved. Nobody actually tested it.
Last month someone finally did, at the one place where that worry is most intuitive. The prediction broke the other way.
The CCO who paused her project for eighteen months
The head of internal communications at a European industrial group, in a cohort session two weeks ago. She explained why her team had held an AI-powered intranet search in review for eighteen months. "If people get the AI answer first, they stop seeing the long tail. We would narrow what the organisation reads." Everyone in the room nodded. I did too.
Then I asked what evidence she had for the sign of the effect. She paused for a long time. "None, really. It just felt obvious."
Felt-obvious is doing a lot of work in that room. It paused a real project. It held comms strategy in a defensive crouch for eighteen months. Nobody owed it a test.
What the Washington Post experiment actually found
A four-author team ran a 38-day randomised field experiment on washingtonpost.com, October through November 2024. The team came from UT Dallas, Boston University, American University and HKUST. Sample size: 37,561 readers. Readers were randomised at the browser-cookie level on first on-site search, into either conventional results or AI-generated answers with cited articles displayed above them. Same archive, both arms.
"Treatment readers consume 69% more shared information on average than Control readers (P << 0.001)."
More shared information. More topic diversity per reader (effective topics 4.58 → 5.55). Aggregate consumption shifted toward less-popular topics, with top-10 share down 4.4 percentage points. Total information per reader up 82.2%. Four conceptually distinct measures all moving the same direction.
Most of the shared-information gain came from the AI answer itself, before any click on a cited article. The AI delivered a wider set of ideas without requiring the click. The mental model that said AI would route traffic away from the long tail had the sign wrong.
The honest caveat: one field experiment, one high-stakes setting, no preregistration shown. The direction of the effect inside your intranet or your knowledge base is still an open empirical question.
If your content strategy, your internal search project, or your comms platform rollout was shaped by the filter bubble prediction, the anchor just moved. One well-identified experiment ran the other way. The audit is cheap. The decisions downstream of a wrong assumption are not.