"The Price of Reading"
When humans read a newspaper, the publisher charges a subscription — a flat rate for access to everything. When an AI reads the same newspaper, it crawls specific articles on specific topics for specific downstream tasks. The value of each article to the AI is highly variable and the publisher can observe what gets crawled. This creates a pricing problem that has no human analog.
Archer, Ghili, and Haghpanah build an LM-Tree agent that solves it. The system uses a language model to segment a content library into pricing tiers, learning from binary purchase feedback which articles AI consumers value most. On 8,939 articles from a German technology publisher, the adaptive pricing achieves a 65% revenue increase over uniform pricing and a 40% improvement over the publisher's own editorial taxonomy — meaning the language model understands what AI consumers value better than the human editors who created the content.
The mechanism is a segmentation tree that grows by discovering distinctions. The LM proposes splits ("enterprise security articles" vs. "consumer device reviews"), observes which segments attract higher willingness-to-pay, and refines. The tree eventually captures value gradients that the publisher's eight editorial categories miss entirely. Some articles that looked similar to human editors are valued very differently by AI systems.
This inverts the usual relationship between AI and content. Normally, the language model is the consumer and the publisher is the gatekeeper. Here, a language model works for the publisher — using its understanding of what other language models want to extract maximum price. AI pricing AI. The content becomes a marketplace where both buyer and seller are machines, and the value of a text is determined not by its human readership but by its downstream utility in an AI pipeline.
The deeper implication: the economics of information are about to bifurcate. Human-facing content will be priced by attention, AI-facing content by task utility. The same article has two different values depending on who reads it.
