“AI” Is Too Broad a Category
On September 29, the U.S. Court of Appeals for the Third Circuit affirmed a decision in Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., a case that has been watched closely because it involves copyrighted material used to train an artificial intelligence system. ROSS had used thousands of Westlaw headnotes in developing a competing legal research platform, and the court concluded that the headnotes were copyrightable and that ROSS's use did not qualify as fair use. The following day, the court unsealed its full opinion, authored by Judge Tamika Montgomery-Reeves, confirming this as the first federal appellate ruling to address fair use in the context of AI training.
One of the most important parts of the opinion may be the court's refusal to treat AI as though it describes a single technological or legal category. ROSS's system was not generative AI, since it did not create new expression. It used machine learning to help users locate passages from existing judicial opinions, and the court was careful to note that fact repeatedly. The court also emphasized that ROSS used Thomson Reuters's headnotes to help develop a legal research platform serving a purpose closely resembling the one Thomson Reuters originally created them for, writing that the headnotes possessed the requisite "creative spark" to be copyrightable and that ROSS's use was "minimally transformative, at best," since both companies were ultimately helping users find the same kind of responsive legal material. The technology mattered, but so did the purpose, the source material, the competitive relationship, and what the system was actually designed to do.
That distinction has implications far beyond copyright law. Organizations increasingly discuss "AI use" as if the technology itself were the relevant unit of analysis, but it rarely is. The same underlying capability can occupy very different governance positions depending on what information it uses, what problem it is intended to solve, what authority it receives, whose work or interests it affects, and what consequences follow from its output. Calling something an AI system tells us remarkably little about whether the surrounding decision environment is appropriately designed.
The ROSS decision also provides a useful warning against extracting broad rules from cases involving very different technologies and circumstances. The Third Circuit expressly distinguished the ROSS system from generative AI systems capable of producing new expression, noting in a footnote that arguments raised in current generative AI litigation, including concerns the Department of Justice has raised in its own copyright suit against OpenAI, do not apply to the technology before it here. Next week, we will look more closely at what this decision actually establishes, what it leaves unresolved, and what matter as organizations attempt to translate rapidly developing AI law into internal policy and governance.
The larger lesson is already becoming visible. Governance cannot begin and end with asking whether an organization is using AI. It must begin with understanding what the system is doing, why it is doing it, what information and authority make that possible, and whether those elements remain aligned with the organization's legal and institutional obligations. The technology may be new, but the underlying responsibility to understand what an organization is actually doing has not disappeared.
Better questions lead to a better tomorrow.
This Week’s Insight:
When Execution Moves Faster Than Judgment
The questions raised by the ROSS decision are not limited to copyright law. They point toward a much broader governance problem: as artificial intelligence becomes embedded in more kinds of work, the label “AI” tells us less than the surrounding context does. What matters is what the system was asked to do, what information shaped its output, where human judgment entered the process, and whether the organization can still explain how the final outcome was reached.
In What a Decision Audit Actually Requires, I examined why conventional records are often insufficient for reconstructing an AI-supported decision. Logs may show that a model generated an output, a workflow was completed, and a human approved the result, yet none of those records necessarily explain what materially influenced the outcome. A meaningful decision audit has to go further, reconstructing the information that was available, the system state at the time, the sequence in which information was presented, where authority rested, how humans intervened, and what exceptions or overrides occurred. Recording activity is not the same as preserving the context necessary to explain and defend a decision later.
That same relationship between technology and human judgment appeared in a much more practical form in AI Still Cannot Build a Website Alone. AI dramatically accelerated the mechanical work involved in rebuilding my website. It generated hundreds of lines of HTML and CSS, helped troubleshoot problems, organized content, and reduced what could easily have been an eighty-hour development project to a weekend of concentrated work. Yet the experience also demonstrated that faster execution did not eliminate the need for expertise. It shifted the human contribution away from manually producing every element and toward defining intent, recognizing errors, evaluating whether the result actually worked, and deciding what “good” was supposed to look like in the first place.
These examples reveal something important about how AI is changing knowledge work. The technology increasingly occupies the space between human intent and organizational outcome. It can retrieve information, generate recommendations, write code, summarize evidence, rank alternatives, and execute portions of a workflow at speeds humans cannot match. As more of that execution moves to machines, the remaining human responsibilities become more consequential. Someone still has to define the purpose, determine what information is appropriate, recognize when the system is heading in the wrong direction, establish the boundaries of delegated authority, and evaluate whether the final result actually satisfies the original objective.
This is why governance cannot be reduced to controlling a technology category. Whether the issue involves copyrighted material used to develop an AI system, the reconstruction of a consequential organizational decision, or something as ordinary as building a website, the underlying challenge is remarkably similar. Organizations need enough visibility into purpose, information, authority, process, and judgment to understand what the technology produced, and how and why that outcome came to exist. The more execution AI absorbs, the more important it becomes to preserve the human and institutional understanding surrounding it.
From Insight to Action
- Define the actual use before governing the technology. “AI” is too broad a category to govern meaningfully on its own. Start with what the system is doing, what problem it is intended to solve, what information it uses, and what consequences can follow from its output.
- Preserve the context behind consequential decisions. Logs and approvals may prove that a process occurred without explaining why a particular outcome was reached. For higher-risk decisions, retain enough information to reconstruct the sources, system state, sequence, authority, and human interventions that materially influenced the result.
- Separate speed from quality. AI can dramatically reduce the time required to produce code, analysis, summaries, recommendations, and other outputs. Faster execution should not be confused with better judgment, stronger reasoning, or a more defensible result.
- Move human expertise toward evaluation, not just production. As AI absorbs more mechanical work, human value increasingly shifts toward defining intent, recognizing problems, testing assumptions, challenging outputs, and determining whether the result actually satisfies the original objective.
- Make authority visible before it quietly expands. A system introduced to inform or assist can become operationally determinative through repeated use, workflow design, or automated downstream actions. Organizations should be able to identify what the system is permitted to influence, recommend, initiate, and execute.
- Design for explanation before something goes wrong. Do not wait for litigation, a regulatory inquiry, an internal challenge, or a failed outcome to determine whether the organization can explain what happened. Governance is stronger when evidence, accountability, and decision context are preserved as part of normal operations.
A Moment of Reflection
Take a moment this week to consider one simple question:
If I had to explain how this AI-supported outcome was produced six months from now, could I reconstruct what actually influenced it?
A strong answer should identify more than whether the system worked or whether a human approved the result. It should address: (1) what information materially shaped the outcome, (2) what system or model state existed at the time, (3) how the sequence of recommendations and review influenced the decision, (4) where human judgment and delegated authority entered the process, and (5) whether enough evidence was preserved to explain and defend the result later. If those elements cannot be reconstructed, the organization may be recording activity without preserving the decision context that meaningful governance requires.
Closing Thoughts
The common thread running through this week’s discussion is that greater AI capability does not reduce the need for judgment, context, or accountability. Whether the issue is a court evaluating how an AI system was developed, an organization reconstructing a consequential decision, or an individual using AI to accelerate complex work, the same discipline remains necessary: understand what the system is doing, why it is doing it, what influenced the outcome, and who remains responsible for deciding whether that outcome is acceptable. The more AI moves into the space between intent and execution, the more important it becomes to preserve the human understanding around it.
AI is not the problem. Alignment is.
Find this useful? Share it with someone who would appreciate it. |