Who owns what we know — and who decides who may use it? That question, which helped drive Aaron Swartz to risk everything a decade ago, has returned with renewed urgency in the age of artificial intelligence. As machine learning systems devour, distill and republish the world’s research, law, commerce and politics collide over a single, uncomfortable fact: the same firms that train the most powerful models also control the most consequential access to knowledge.
Swartz argued that publicly funded knowledge should be openly available to the public. His act of mass download from the JSTOR archive and the federal prosecution that followed exposed a fundamental tension: legal protections and commercial incentives can lock information away from the public, even when that information was produced with public money. That contradiction now animates debates about who may scrape, license or teach models from science, scholarship and cultural works — and what it means when corporations become gatekeepers for knowledge itself.
The moment is not hypothetical. AI’s promise depends on vast troves of text, code and imagery. Firms with the resources to assemble proprietary datasets, build enormous compute clusters and tune closed models have a practical advantage. That advantage translates into market power: control of platforms for discovery, personalized tutoring, automated research assistance and even the tools scholars rely on to find and cite prior work. When access to source materials, high‑quality data or the APIs that serve model outputs is restricted, corporate priorities can shape what research is visible, what questions are answered and whose work is amplified.
Technologists and security analysts warn that the dynamics of scale in AI aren’t just economic; they change how risk plays out. Models trained to find vulnerabilities can automate reconnaissance and accelerate attack campaigns far faster than human analysts could manage, while defenders must marshal context — inventories, dependency graphs and richer telemetry — to respond effectively. That insight, summarized in recent technical commentary, underlines the same structural imbalance that gives a handful of well‑capitalized firms both offensive capability and custodial control over data and tools the public needs to use safely and fairly .
At the same time, the infrastructure that carries corporate and public data remains leaky. Studies of long‑standing communications channels show that vast amounts of corporate and government traffic can be observed with modest equipment, a reminder that control over knowledge also depends on the security and accessibility of the systems that move it .
Why does this matter? There are three overlapping consequences worth considering:
- Research inequality and rent extraction. When publishers or platforms monetize exclusive access to journals, datasets or citation tools, they extract rents from institutions and researchers. Firms that layer AI services on top of paywalled content can further concentrate value, offering superior tools only to customers who pay — and curtailing competition from smaller labs, independent scholars or non‑profits.
- Distortion of the knowledge ecosystem. Models bake in the biases of their training sets and the licensing choices made by their builders. If corpuses privilege English‑language, commercially produced or paywalled sources, the models trained on them will reflect and reproduce those biases — steering attention, funding and policy toward a narrow set of topics and voices.
- Legal and ethical friction. Copyright, database protection, and contract law are being tested by large‑scale scraping and dataset curation. The litigation and licensing responses will shape not only commercial arrangements but the practical reality of who may build or benefit from future models.
Different actors see different threats and opportunities.
- Technologists and open‑knowledge advocates often argue for broader access. They point to the public‑goods role of scholarship and the practical benefits of open datasets for reproducibility, innovation and equitable participation. The moral argument echoes Swartz: when taxpayer money funds research, the public has a stake in the results.
- Publishers and rights holders stress incentives. They contend that revenue from subscriptions, licensing and content partnerships sustains peer review, editorial labor and infrastructure. Without the ability to monetize content, some say, quality could suffer and investment in niche fields could dry up.
- Policymakers face tradeoffs. Some propose rules to require public access to publicly funded research, expand fair use, or mandate interoperable APIs. Others emphasize intellectual property as necessary to reward creators and attract investment. Regulators also worry about security, privacy and the potential for AI systems to amplify misinformation.
- Users and the public want tools that are useful, transparent and affordable. They also want safeguards: clear provenance for AI outputs, mechanisms to contest errors, and equitable access to services that increasingly mediate education, health and employment decisions.
There are practical paths forward that do not require choosing sides between open science and commercial viability. Promising approaches include:
- Policy that ties public funding to open access mandates, while preserving reasonable commercial pathways for added‑value services.
- Standardized, transparent licensing regimes that make it clear when datasets may be used to train models and under what terms.
- Technical and institutional investments to capture operational context — ownership registries, dependency maps and enhanced metadata — so that the provenance and permissions around training data are auditable and enforceable.
- Hybrid business models in which publishers and creators receive compensation for derivative commercial uses while core research remains broadly accessible for discovery and verification.
There are hard problems, too. Enforcement is costly; cross‑border datasets collide with divergent legal regimes; and market concentration rewards scale in ways that drive consolidation. Technical fixes — better metadata, watermarking model outputs, or federated training — help, but they do not by themselves redistribute bargaining power.
To borrow a phrase from cybersecurity analysis: context wins. The same logic applies here. Transparency about datasets, clear legal frameworks for training and reuse, and institutional investments in public infrastructure shift the balance away from opaque corporate gatekeeping toward accountable stewardship of knowledge .
If we value a knowledge commons that serves scholarship, civic life and the innovation economy, the policy questions are urgent and practical: which parts of the research ecosystem are treated as public infrastructure, what compensation models are acceptable for commercial layering, and how do we ensure that AI amplifies, rather than narrows, human inquiry?
Aaron Swartz’s tragedy was not merely about one prosecution; it was about the friction between law, market power and moral conviction over access to knowledge. Today, as models ingest and re‑express vast swaths of human work, that friction has become institutionalized. The choices we make now — about licenses, standards, competition policy and public investment — will determine whether AI becomes a democratizing force for shared knowledge or a mechanism for corporate capture.
Are we prepared to treat knowledge as an asset to be guarded by a few, or as a public resource that fuels many? The answer will shape not only who profits from AI, but who gets to understand the world it creates. Source: https://www.schneier.com/blog/archives/2026/01/ai-and-the-corporate-capture-of-knowledge.html
