A Spoonful Of Sugar: Bryan Sugar's AI Litigation Insights - September 2026

Artificial intelligence continues to move from novelty to necessity in both business operations and IP litigation strategy. Over the past several months, courts, regulators, and litigants have begun to confront foundational questions: Who owns AI‑generated works? When is training an AI model fair use, and when is it infringement? How do we prove or disprove copying when the alleged “copier” is a black‑box model? And what happens when litigants themselves rely on AI and get it badly wrong?

Below is a high‑level survey of several key AI‑related developments with particular relevance to IP litigators, along with some practical takeaways for those navigating these issues in active or anticipated disputes.

1. Authorship, Ownership, and the Limits of Human Creativity Requirements

Courts and agencies have largely converged on a central premise: current U.S. copyright law protects only works with human authorship. In Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025), the D.C. Circuit held that "[t]he Creativity Machine cannot be the recognized author of a copyrighted work because the Copyright Act of 1976 requires all eligible work to be authored in the first instance by a human being," affirming denial of registration for a visual artwork autonomously generated by Stephen Thaler's "Creativity Machine" with no claimed human author. The court was careful to clarify that "the human authorship requirement does not prohibit copyrighting work that was made by or with the assistance of artificial intelligence," explaining that "[t]he rule requires only that the author of that work be a human being—the person who created, operated, or used artificial intelligence—and not the machine itself." This traces back to Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884), which held a photograph is copyrightable only "so far as they are representatives of original intellectual conceptions of the author," and to Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991), which held that "[t]he originality requirement articulated in The Trade-Mark Cases and Burrow-Giles remains the touchstone of copyright protection today." Where AI is involved, the legal inquiry turns on the degree of human control, selection, and arrangement.

The U.S. Copyright Office's March 2023 Registration Guidance (88 Fed. Reg. 16,190) and its January 2025 report, Copyright and Artificial Intelligence, Part 2: Copyrightability, underscore three points:

1. If a work is generated entirely by an autonomous AI system, with no meaningful human creative input, it is not copyrightable. The Copyright Office asks whether AI output reflects the human's "own original mental conception, to which [the author] gave visible form," or instead is the product of "mechanical reproduction" driven by the AI's own determination of expressive detail, and has taken the position that a prompt alone, however detailed, generally does not confer the "ultimate creative control" needed for human authorship because prompts function "more like instructions to a commissioned artist."

2. If a human uses AI as a tool but exercises creative judgment in prompts, editing, and arrangement, those human contributions may be protectable, even if the underlying AI output is not. In the Copyright Office's February 2023 cancellation decision regarding the graphic novel Zarya of the Dawn, the Office concluded that a human author's selection, coordination, and arrangement of AI-generated images into a comic book was copyrightable as a compilation, while the individual AI-generated images themselves were not and had to be disclaimed.

3. Applicants have a duty to disclose the role of AI when registering works and to limit their claimed authorship accordingly. Under the Registration Guidance, applicants must use the Standard Application, describe only the human-contributed authorship in the "Author Created" field, and affirmatively exclude non-de-minimis AI-generated content via the "Material Excluded" field; failure to correct the record risks cancellation of the registration under 37 C.F.R. § 201.7(c)(4) or a court disregarding the registration under 17 U.S.C. § 411(b)(1)(A).

For IP litigators, that framework has concrete consequences. Plaintiffs asserting copyright in AI‑assisted works will face threshold challenges to validity, moving these cases from classic copying disputes into battles over registrability and the definition of authorship. Defendants, in turn, will have new lines of attack in Rule 12 and summary judgment practice, targeting the human authorship requirement and any mismatch between the registration and the actual scope of human contribution.

In parallel, transactional and corporate practices are increasingly revising IP ownership clauses, work‑for‑hire language, and assignment provisions to account explicitly for AI involvement. That drafting trend will inevitably bleed into litigation through contract interpretation disputes when parties disagree over who owns what portion of an AI‑assisted output.

Practice implications for IP litigators:

• Expect more motions challenging copyright registrations of AI‑assisted works, particularly when the registration does not disclose AI use.

• In discovery, focus on prompt logs, version histories, and human editing workflows to assess the extent of human authorship.

• When counseling clients pre‑dispute, encourage them to document human contribution to AI‑assisted content, anticipating later validity challenges.

2. Training Data, Fair Use, and Emerging Theories of Infringement

The most hotly watched AI IP matters involve claims that large language models and generative systems infringe copyright through their ingestion of vast corpora of text, images, code, and other creative works. A cluster of cases now illustrates how these theories are playing out: Bartz v. Anthropic PBC, 787 F. Supp. 3d 1007 (N.D. Cal. 2025); Kadrey v. Meta Platforms, Inc., No. 23-cv-03417-VC (N.D. Cal.); Andersen v. Stability AI Ltd., No. 23-cv-00201-WHO (N.D. Cal.); The New York Times Co. v. Microsoft Corp., 777 F. Supp. 3d 283 (S.D.N.Y. 2025); In re OpenAI, Inc. Copyright Infringement Litigation, MDL No. 25-md-3143 (S.D.N.Y.); Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc., 765 F. Supp. 3d 382 (D. Del. 2025); Doe v. GitHub, Inc., No. 22-cv-06823-JST (N.D. Cal.); and In re Google Generative AI Copyright Litigation, 809 F. Supp. 3d 903 (N.D. Cal. 2025). Plaintiffs across these cases have advanced direct infringement in the copying of works into training sets, unauthorized derivative works through model outputs, vicarious or contributory liability, and DMCA Section 1202(b) claims for removal of copyright management information. "Model-as-derivative-work" theories have uniformly failed at the pleading stage: in Kadrey, the court held there is "no way to understand the LLaMA models themselves as a recasting or adaptation of any of the plaintiffs' books," and Andersen likewise held it "simply not plausible" that every output relies on copyrighted material absent allegations of similarity to protected elements of specific training images.

Defendants have increasingly invoked fair use, analogizing model training to familiar uses such as search indexing and text/data mining. In Bartz, Judge Alsup held that copying books to train Anthropic's Claude models was fair use because "[e]very factor but the nature of the copyrighted work favors this result," describing the technology as "among the most transformative many of us will see in our lifetimes," while separately holding that copies pirated from shadow libraries to build a permanent "central library" were not fair use because "[e]very factor points against fair use" for that separate, non-transformative use. In Kadrey, Judge Chhabria granted summary judgment to Meta on fair use as to the named plaintiffs, but expressly cautioned the ruling "stands only for the proposition that these plaintiffs made the wrong arguments and failed to develop a record," and that in future cases "it seems like the plaintiffs will often win" if they present evidence of market dilution, the theory that AI-generated works compete with and displace the human-authored works used to train the model. Plaintiffs counter that large‑scale copying of expressive works for commercial AI products exceeds any reasonable fair use boundary and threatens traditional licensing markets, particularly for code, news, and visual art; in Thomson Reuters v. Ross Intelligence, the court granted summary judgment for Thomson Reuters on fair use because Ross's non-generative AI legal-research tool was built to directly compete with Westlaw, though the court expressly flagged that "only non-generative AI is before me today," a distinction Bartz later relied on to distinguish Ross Intelligence when finding LLM training itself transformative.

Courts confronted with motions to dismiss in these cases have often drawn a line between claims tied to concrete, allegedly infringing outputs and broader, more abstract attacks on the existence of training itself. In Andersen and Kadrey, courts dismissed output-based derivative-work and vicarious-infringement theories at the pleadings stage because plaintiffs failed to allege that specific outputs were substantially similar to or incorporated protected elements of specific copyrighted works, while direct-infringement claims based on the copying of works into training sets have generally survived motions to dismiss. In The New York Times Co. v. Microsoft Corp., the court allowed a contributory-infringement theory to proceed based on allegations of actual infringing outputs and defendants' at least constructive knowledge, declining to apply the Sony/Grokster substantial-noninfringing-use defense at the pleadings stage. Many judges appear reluctant, at the pleadings stage, to declare AI training categorically fair or unfair without detailed factual records about datasets, model architectures, and practical market effects, reserving that determination for summary judgment, as in Bartz, Kadrey, and Thomson Reuters.

As matters move into discovery, litigators are grappling with several novel problems:

• The extent to which source datasets, model weights, and training pipelines are discoverable, given trade secret and security concerns.

• How to frame “substantial similarity” when a model is alleged to have ingested millions of works, and the outputs are statistical composites rather than conventional copies.

• Whether and when “memorization” of training data in a model can convert a statistical system into a reproducing device for specific works.

Practice implications for IP litigators:

• Anticipate heavily contested protective orders and confidentiality regimes around datasets and model internals.

• Consider early expert involvement not just for damages, but to shape the technical narrative on how models learn and what they retain.

• For defendants, develop a record distinguishing non‑expressive intermediate copying from market‑substituting outputs.

• For plaintiffs, focus on concrete instances of near‑verbatim or highly recognizable outputs, and on evidence of licensing markets harmed by unlicensed training.

3. AI in the Courtroom: Litigation Conduct, Sanctions, and Evidentiary Concerns

While much attention has focused on AI as a subject of litigation, courts are also increasingly dealing with AI as a tool used by litigants themselves. In the foundational case, Mata v. Avianca, Inc., 678 F. Supp. 3d 443 (S.D.N.Y. 2023), plaintiff's counsel filed a brief citing fabricated case law generated by ChatGPT; Judge Castel held that "[a] fake opinion is not 'existing law' and citation to a fake opinion does not provide a non-frivolous ground for extending, modifying, or reversing existing law," describing the conduct as "an abuse of the adversary system," and imposed a $5,000 sanction under Rule 11 jointly and severally on the attorneys and their firm, along with orders requiring notice to the client and to each real judge falsely identified as an opinion's author. Since then, fabricated AI-generated citations have proliferated across federal courts nationwide, and a tracking database cited by the Fifth Circuit identified 239 U.S. lawyer AI-hallucination cases as of mid-2025, later cited by an Eastern District of New York court as 919 cases by April 2026. As one federal court put it, "[i]t is . . . well-known in the legal community that AI resources generate fake cases." Sanctions have followed in numerous cases, including the Second Circuit's referral of counsel to its Grievance Panel in Park v. Kim, 91 F.4th 610 (2d Cir. 2024), a $2,000 fine and mandatory CLE training in Gauthier v. Goodyear Tire & Rubber Co., No. 1:23-CV-281 (E.D. Tex. Nov. 25, 2024), and a $2,500 sanction in Fletcher v. Experian Information Solutions, Inc., 168 F.4th 231 (5th Cir. 2026).

Judges have responded in a few ways:

• Requiring certifications that any AI‑assisted filings have been independently verified by counsel. For example, the Northern District of Texas's Local Civil Rule 7.2(f) requires that "[a] brief prepared using generative artificial intelligence must disclose this fact on the first page," and treats the absence of disclosure as a certification that no AI was used.

• Issuing standing orders or general orders governing the use of generative AI in drafting submissions, with an emphasis on human responsibility. The Southern District of Texas's General Order 2025-04, for example, cautions litigants against submitting AI-drafted filings "without checking the submission for accuracy."

• Emphasizing that reliance on AI does not excuse violations of Rule 11 or professional responsibility obligations. As the Second Circuit put it in Park v. Kim, "the duties imposed by Rule 11 require that attorneys read, and thereby confirm the existence and validity of, the legal authorities on which they rely," and the Fifth Circuit's AI Subcommittee ultimately declined to adopt an AI-specific local rule, concluding that existing Rule 11 and Federal Rule of Appellate Procedure 46 obligations already required no less.

More subtly, evidentiary questions are emerging around the admissibility and weight of AI‑generated analyses, summary charts, or technical reports. In IP cases, parties may be tempted to use AI for prior art searches, claim chart generation, or similarity analysis between works. The line between permissible internal use and overreliance in sworn submissions is not always clear.

For IP litigators, particularly in complex patent and copyright cases, AI will increasingly become part of the litigation toolkit, from document review and e‑discovery to analytics for judge and venue tendencies. The core message from the bench remains consistent: AI can assist, but cannot replace, the attorney’s independent judgment and verification.

Practice implications for IP litigators:

• Treat any AI output as a draft or lead, not as final authority; independently confirm all cases, statutes, and factual assertions.

• Monitor standing orders in key venues for specific rules governing AI use in filings.

• When using AI‑assisted tools for discovery or analytics, be prepared to explain methodologies if later challenged on proportionality, completeness, or bias grounds.

4. Trade Secrets, Data Access, and Reverse Engineering of AI Systems

Beyond copyright, AI systems have triggered an uptick in trade secret claims. Companies allege misappropriation where employees use confidential datasets, code, or architectures to build competing models, or where access to APIs and interfaces is allegedly abused to scrape data or reverse engineer protected functionality. In X.AI Corp. v. OpenAI, Inc., No. 25-cv-08133-RFL (N.D. Cal. Feb. 24, 2026), xAI alleged that former employees took source code and confidential AI model training and tuning materials to OpenAI; the court dismissed the claims against OpenAI with leave to amend because xAI failed to plausibly allege that OpenAI induced the taking or that the employees actually used the trade secrets once at OpenAI, holding that "mere possession of trade secrets is not sufficient to" establish misappropriation by the new employer. In Juries.ai, Inc. v. Sheu, No. 5:25-cv-10188-BLF (N.D. Cal. Nov. 25, 2025), by contrast, the court granted a temporary restraining order against a departing co-founder who retained an AI litigation-simulation platform's source code and methodologies.

At the same time, plaintiffs challenging AI systems, for bias, IP infringement, or other harms, often seek deep technical discovery that defendants resist as trade secrets. This tension is particularly acute in IP litigation, where plaintiffs may argue that they cannot prove copying without insight into training corpora and internal processes. In the consolidated In re OpenAI, Inc., Copyright Infringement Litigation, MDL No. 25-md-3143 (S.D.N.Y.), the July 2025 stipulated protective order created a four-tier confidentiality regime, including a "HIGHLY CONFIDENTIAL — SOURCE CODE" tier restricted to outside counsel and experts, with source code produced only on standalone, non-networked computers subject to strict inspection protocols. Courts have also policed over-designation: the magistrate judge in that MDL flagged her "concern with OpenAI's overuse of the AEO designation" where documents "do not mention particular Plaintiffs at all," and in Concord Music Group, Inc. v. Anthropic PBC, No. 24-cv-03811 (N.D. Cal. May 23, 2025), the court similarly found Anthropic's use of the AEO designation for usage statistics "overbroad." In Kadrey v. Meta Platforms, Inc., No. 23-cv-03417-VC (N.D. Cal. Jan. 8, 2025), the court compelled production of fine-tuning datasets tied to Meta's copyright-related safety mitigations while denying discovery into unrelated safety categories and raw pre-cleaning training data as disproportionate.

Courts are beginning to strike balances through:

• Carefully tiered protective orders with “attorneys’ eyes only” designations and expert access limitations, as in the In re OpenAI MDL's four-tier confidentiality regime.

• On‑site or virtual inspections of source code and model components under strict security protocols, following the source-code inspection framework used in Waymo LLC v. Uber Technologies, Inc., No. C 17-00939 WHA (N.D. Cal.), and later adopted with AI-specific variations in the In re OpenAI MDL protective order.

• Use of special masters or technical neutrals to evaluate sensitive material and report in more generalized terms, as in Waymo LLC v. Uber Technologies, Inc., where the court appointed a special master to monitor and verify compliance with data-return and information-barrier obligations.

For IP litigators, these disputes overlay familiar source code discovery battles with the added complexity of large‑scale datasets, model weights, and proprietary evaluation pipelines. The complexity often exceeds any simple formulaic solution, requiring case‑specific negotiation and creativity.

Practice implications for IP litigators:

• Be prepared for early motion practice and negotiation around the scope and mechanics of technical discovery into AI systems.

• Consider whether targeted sampling, hashing, or statistical methods can provide evidence of copying without full dataset disclosure.

• For defendants, build a clear record of trade secret measures and the competitive value of model and dataset information to support protective order positions.

5. Trademark, Right of Publicity, and AI‑Generated Personas

AI‑generated text, images, and audio have also raised issues at the intersection of trademark, false endorsement, and publicity rights. No reported case yet directly adjudicates a dispute against a commercial AI voice-cloning company, but the doctrinal groundwork is well established. 

• Voice cloning of well‑known artists used in songs or ads, an area governed by the voice-misappropriation tort recognized in Midler v. Ford Motor Co., 849 F.2d 460 (9th Cir. 1988) ("[a] voice is as distinctive and personal as a face"), and Waits v. Frito-Lay, Inc., 978 F.2d 1093 (9th Cir. 1992), which held that imitating a singer's distinctive voice for commercial use can support a Lanham Act false-endorsement claim as well as a state-law tort.

• AI‑generated avatars that resemble public figures without explicit naming, analogous to White v. Samsung Electronics America, Inc., 971 F.2d 1395 (9th Cir. 1992), which held that a robot dressed and posed to evoke Vanna White's identity could support a Lanham Act false-endorsement claim even without a literal photograph or recording of her.

• Chatbots or AI “assistants” branded or presented in ways that suggest association with a known company or personality, implicating the trademark “use as a mark” threshold the Supreme Court addressed in Jack Daniel's Properties, Inc. v. VIP Products LLC, 599 U.S. 140 (2023), which held that the Rogers First Amendment test for expressive works “does not apply when the challenged use of a mark is as a mark,” that is, as a source identifier for the defendant's own goods or services.

IP litigators should expect more hybrid cases where copyright, trademark, and publicity theories are pled together around the same AI‑generated content. The analysis may hinge on whether the AI output is expressive or commercial, transformative or exploitative, and on how the content is marketed. 

Practice implications for IP litigators:

• When advising clients deploying AI‑generated marketing or branded content, stress clearance and guardrails around the use of real‑world personas and marks.

• In litigation, be prepared to address overlapping theories, including the Lanham Act, state publicity statutes, and unfair competition, and to brief constitutional defenses where expressive uses are alleged.

• Evidence about technical controls (e.g., filters to prevent generating specific celebrities) may become relevant to intent and willfulness.

6. Looking Ahead: Strategic Considerations for IP Litigators in an AI‑Heavy Docket

As AI becomes embedded in both the subject matter and conduct of IP litigation, several strategic themes are emerging:

• Venue and judge selection matter more than ever. Different courts and individual judges have shown varying levels of comfort with technical and AI‑specific issues. 

• Early technical case framing can shape outcomes. The party that first offers a clear, accessible explanation of how a model actually works, and what it does not do, often gains credibility.

• Expert selection is critical and challenging. Effective experts must bridge law and technology, explaining concepts like training, inference, and memorization without jargon.

• Documentation is king. For clients building or using AI systems, contemporaneous records of data provenance, licensing decisions, human oversight, and risk assessments will heavily influence litigation risk and defensibility. 

While the legal landscape remains fluid and many of these issues are too complex for any definitive resolution at this stage, IP litigators can position clients effectively by tracking these trends, building technical fluency, and treating AI as both a subject and a tool, powerful, but always subordinate to human legal judgment.