Open or Invisible: The Question Facing Every Scholarly Publisher

For half a century, the way a researcher found a paper was roughly the same: a search engine, a citation chain, a journal. That interface is being rebuilt around large language models, and the Association for Computing Machinery (ACM), the largest scientific society in computing, founded in 1947, is now publicly arguing it cannot afford to be left out of the rewrite.

In a Communications of the ACM (CACM) opinion column this month, an ACM author makes the case that the society's subscription Digital Library should be opened to large language models for both training and retrieval-augmented generation (RAG), a technique where an AI system fetches relevant documents at query time to ground its answers. The argument: if ACM stays walled off, peer-reviewed computing research will be quoted less, surfaced less, and cited less inside the AI tools students, engineers, and policymakers are already using.

The column is not ACM policy. The ACM Publication Board told the community in a note relayed through SIGSIM that the society "does not currently license content from the Digital Library to train LLMs" and is "committed to dialogue with SIGs, authors, and the community before any change to the status quo." That gap, between an advocacy column on CACM and the society's actual posture, is the story. It shows that the question of whether to license peer-reviewed science to AI labs is being treated as an open negotiation rather than a settled decision.

The column's argument is one-sided. It warns that publishers absent from AI ecosystems "may become repositories of record rather than sources of influence." That is a strategic claim about citation flows, not a fact about them. The author is also a participant in the governance structure of the society whose choices they are recommending. The piece should be read as a position paper, not as a neutral assessment of tradeoffs.

The tradeoffs the column does not engage with are the ones the negotiation will turn on. Re-purposing that work for model training or RAG is a use the original agreements may not cover. Compensation: the work was produced by academic authors and volunteer peer reviewers. If a commercial AI lab pays for a license, who receives the money: the society, the authors, the institutions? The Hacker News thread reacting to the column surfaced a third concern: that the licensing debate is being conducted by publishers with very little member democracy about terms, prices, or revenue sharing. Several commenters raised the question of whether "responsible inclusion" is a meaningful commitment or a brand line.

There is also a structural question the column does not address. Scholarly publishers are non-profits whose mission is to serve their disciplines. The AI labs they would license to are some of the most heavily capitalized private companies in technology. The negotiating leverage is asymmetric, and the deals that result will shape which peer-reviewed research AI systems are trained on, retrieve from, and cite. They will shape what those systems know. That is a chokepoint, not a procurement decision.

Other societies face the same choice. The Institute of Electrical and Electronics Engineers (IEEE), the American Physical Society, and a long list of university presses are running parallel conversations. None of them have published a public framework for evaluating LLM licensing terms, and almost none have committed to author-level revenue sharing or opt-in consent regimes. ACM's column is the most visible argument for opening the door; the equally visible case for keeping it closed, or for opening it only under strict terms, has yet to be made with the same clarity.

The next move is on ACM's side. A working group, a member consultation, or a published licensing framework would convert the column from advocacy into a posture. Whether the authors whose work is at stake get a vote in the answer is the part the column does not address.