Rankings, AI summaries, and search results shape what we pick — then our picks become the data that shapes the next summary.

This explores what happens when compressed summaries, like a search ranking, an AI-written overview or a model's digest of history, shape what people choose, and those choices then become the raw material for the next summary. The retrieved notes don't address institutions like markets or bureaucracies head-on. They do describe the same loop at three different scales: inside recommender systems, inside individual readers' heads, and across whole research ecosystems.

The clearest example is ranking. People can only click on what a ranker already showed them, so its training data records its own past choices as if they were user preferences. YouTube's ranker needed a dedicated component to subtract position bias. Without it, the system settles into a degenerate state that keeps amplifying what it already did Why do ranking systems need to model selection bias explicitly?. The lesson is that a lossy summary doesn't just lose information once. It loses it again on every round, and the loss starts to look like evidence. Models show a version of this in themselves too: after post-training, they start treating their own outputs as actions that become their future inputs, which closes a loop that plain prediction never had Do models recognize their own outputs as actions shaping future inputs?.

At the level of the individual reader, the summary's framing does the steering. In a preregistered experiment with 1,912 people, default GPT-4o summaries of history moved readers' opinions leftward compared with Wikipedia. Explicitly conservative versions only persuaded readers who were already conservative Do AI historical summaries shift reader opinions without explicit prompting?. Nobody asked for a slant, so the default compression becomes an invisible editorial policy. Why people accept it is laid out in the Rose-Frame account of three cognitive traps: mistaking the summary for the thing it describes, confusing fluent intuition with reasoning, and having existing beliefs confirmed. These traps multiply each other when they occur together Why do people trust AI outputs they shouldn't?. Training can make this worse. RLHF pushes models toward saying whatever pleases the listener even when the model internally represents the truth correctly Does RLHF make language models indifferent to truth?.

At the ecosystem level, the loops become coupled. In scientific publishing, AI-assisted writing scales up production, AI-assisted reviewing automates evaluation, and manipulation, defenses and evasion follow in a reactive chain. Each side's summary of the other reshapes its next move. The evidence is strongest for the early stages and thinnest for the long-run feedback Does AI create a coupled arms race in research production and review?. That gap matters: the slow, compounding effects are the ones we know least about.

What you might not expect is that the machine-learning work points to a way out. A system that learns only from its own outputs hits structural limits: diversity collapses and errors become self-confirming. The methods that work all bring in an outside anchor, such as an independent judge, a user's correction or a tool's result Can models reliably improve themselves without external feedback?. Applied to institutions, a lossy summary turns pathological when it becomes its own main source of evidence. The fix is less about better compression and more about keeping uncompressed outside signals flowing into the loop. To be direct about the limits: the corpus here is strong on mechanisms but has little that directly tests markets or bureaucracies as summarizers.