What happened
In an October 28, 2026 essay in TechPolicy.Press, Lisa Macpherson of Public Knowledge and Anna Tumadóttir of Creative Commons describe what they call a new internet bargain. They say the former arrangement—creators and publishers make content available, while platforms return acknowledgement and traffic—has weakened. Generative AI has intensified the conflict: model developers seek large, high-quality datasets, while rights holders respond with licensing, lawsuits, legislation, paywalls, and technical blocking.
The authors distinguish two kinds of enclosure. “Extractive enclosure” is the conversion of shared human knowledge into raw material controlled by powerful technology firms, including through exclusive licensing, bypassing access signals, or obscuring crawler identity. “Defensive enclosure” is the restriction of access by creators and publishers who believe their work is being exploited. The essay also describes a quality problem: AI-generated low-quality material and deliberate “LLM poisoning” may contaminate sources used for training or retrieval, while restrictions may disproportionately remove valuable human-created material from the accessible web.
To support the claim that the commons is shrinking, the essay cites the 2024 report *Consent in Crisis: The Rapid Decline of the AI Data Commons*. It reports that, in 2023–2024, more than 5% of tokens in the C4 dataset and more than 28% of its most actively maintained sources became fully restricted from use; restrictions in terms of service covered 45% of C4. It also cites a study measuring an exclusion rate of 60% for AI training bots on business and travel domains, 57.1% on community and knowledge-sharing sites, 40% on news and media, and 10.9% on public research. These figures are presented by the essay as evidence that access is shifting from default readability toward tiered permission.
The authors argue that both unrestricted extraction and defensive blocking can make the commons less open and less useful. They raise concerns that exclusive data deals could advantage large firms, while legal uncertainty and licensing costs could impede smaller and open-model developers. Their proposed direction is a set of principles for a fairer exchange: preserve fair use, improve transparency and preference signaling, and develop workable ways for rights holders and AI developers to share value without making public knowledge inaccessible.
What the archive adds
New Runtime’s earlier coverage, “Fair use benefits everyone” (October 31, 2026), noted OpenAI’s argument for fair use in AI training in the New York Times case. This essay is a new contribution to the same debate: it broadens the focus from a legal defense to the economics and governance of the open-web knowledge commons.
Trend signal
This supports a shift in the web’s access model, but it does not directly fit a pattern New Runtime tracks. The essay’s central claim is about competing incentives and public policy, rather than a shipped agent capability or a concrete interoperability layer. This does not yet fit a pattern New Runtime tracks.
Why now
The authors describe an accelerating conflict: AI developers are seeking valuable data, while publishers and creators are adopting restrictions and pursuing legal or legislative remedies. They also point to current disputes over crawler access and emerging data-access rules, making the terms of a durable exchange an immediate policy question rather than a distant design concern.
Connections to other events
New Runtime’s earlier coverage on October 31, 2026, “Fair use benefits everyone”, described OpenAI’s fair-use argument in the New York Times case. Macpherson and Tumadóttir frame that legal issue within a wider contest over whether public web material remains accessible and how creators might be compensated.
Why it matters
For teams building AI products, web access is not a stable assumption. A dataset that is technically crawlable today may later be restricted, licensed exclusively, or treated differently under law or platform policy. That can affect model training, retrieval quality, product provenance, and the cost of maintaining a dependable data pipeline.
The essay also highlights a difficult trade-off: respecting restrictions may reduce the amount of high-quality material available to developers, while ignoring them can deepen distrust and invite defensive responses. Its proposed bargain is an argument, not a settled operating standard, but it gives product and policy teams a useful frame: durable access depends on credible permission signals, transparency, and incentives that do not reserve the best sources for only the largest buyers. Builders should treat data rights and source quality as product risks, not merely acquisition details.
Limits and uncertainty
The essay is a perspective by two advocacy organizations, not an independent evaluation of its proposed policy approach. Its cited studies use different datasets, definitions, and measures of restriction, so their percentages are not directly interchangeable. The claimed effects on future model quality, competition, and access are risks argued by the authors, not measured outcomes established here. The essay also discusses contested legal questions, including fair use, whose application varies by jurisdiction and remains subject to litigation and policy change.
Reader takeaway
The dispute is not simply whether AI companies may use web content. It is whether the web can sustain a reciprocal system in which creators retain meaningful choices, developers can build on shared knowledge, and public access does not collapse into either extraction or exclusion.