In a Thursday (Nov. 19) analysis of four copyright cases, Polsinelli attorneys argued that institutions need to know where their data came from, what rights came with it, what the system does with it and what it produces. For organizations accustomed to treating copyright as a publishing or licensing issue, AI brings those questions into routine decisions about procurement, research and operations.
In “Thomson Reuters v. ROSS Intelligence,” the Third Circuit rejected a fair-use defense for the use of Westlaw headnotes in developing an AI-powered legal research product. The court considered the commercial purpose and the fact that the resulting product competed with Westlaw. The case shows why the purpose of the new tool and its effect on the market for the original material matter.
Federal district courts in “Bartz v. Anthropic” and “Kadrey v. Meta” found that using copyrighted books to train large language models qualified as fair use on the records before them. But the Bartz court treated Anthropic’s acquisition and retention of pirated copies for a general-purpose library as a separate issue. A fair-use finding for training did not excuse building that library with unauthorized copies, the court found.
Output presents its own risk. In “Doe v. GitHub,” the Ninth Circuit upheld the dismissal of a Digital Millennium Copyright Act claim over AI-generated code that omitted copyright management information. The court left open whether substantially similar output could infringe copyright under ordinary infringement principles.
For institutions, the practical implication is to review permissions against the specific workflow. Summarizing one licensed publication can raise different issues from incorporating thousands into a searchable internal system. Polsinelli advised checking ownership, the source of the copies and applicable license restrictions before incorporating material. Confidentiality obligations and internal policies also matter when employees upload proprietary research, personnel records or internal communications. A vendor’s role in supplying the tool leaves the institution responsible for deciding what its staff can put into it. That review can also establish whether a service retains uploaded files, makes excerpts available to other users or uses the material to improve its models.
AI makes governance more important because the same tools that process more information can also shape more consequential decisions. A system introduced to organize grant applications or resumes may start ranking them. If reviewers routinely accept those rankings without examining the underlying material, the institution can effectively delegate judgment without ever formally deciding to do so.
Polsinelli recommended identifying which decisions AI already informs, determining how much influence the system has in practice and designating who can approve each use. For consequential decisions, reviewers need to understand what the system considers, recognize its limitations and have the authority to override its recommendations. The National Institute of Standards and Technology’s AI risk management guidance calls for clearly defined oversight roles. The EU AI Act requires human oversight for certain high-risk systems.
That oversight also helps an institution explain a decision to an employee, applicant, regulator or court. A scholarship allocation or hiring decision may depend on context the system missed or factors the institution weighs differently. As AI moves from summarizing documents to recommending outcomes, leaders need to be able to account for both the information that went into the system and the judgment behind the decision that came out of the process.