AI Capability Claims Under Scrutiny: IP, Disclosure and Governance Risk

Introduction: the AI narrative moves to the centre of corporate risk

The corporate narrative surrounding artificial intelligence has migrated far beyond product marketing brochures. Corporate leadership now routinely incorporates AI automation claims into earnings calls, restructuring announcements, and board communications. In the current economic environment, markets often reward companies that portray AI as a source of efficiency and growth, while penalising companies that report slowing demand. Plaintiffs may therefore claim that some companies used AI narratives to mask underlying financial difficulties.

Navigating this landscape requires corporate counsel to understand how intellectual property, disclosure obligations, and corporate governance increasingly converge. Capability claims rarely create risk in isolation. Rather, exposure often arises when management’s public description of an AI system is not fully aligned with what the company actually owns, how the system actually functions, or the role the technology actually plays within the enterprise.

This article argues that AI capability disputes typically begin with intellectual-property questions but frequently evolve into disclosure and governance disputes. Understanding that progression can help companies design stronger records, more disciplined disclosures, and more defensible oversight structures.

The IP layer: what “proprietary AI” claims actually assert

Every claim a company makes about its AI rests on a basic intellectual property question: what does the company actually own?

Companies like to call their AI “proprietary”, which may suggest ownership, control, exclusivity, or technical differentiation, depending on the underlying rights, but the reality is often more complicated, and getting it wrong can come with potential consequences. Regulators and private plaintiffs have challenged businesses alleged to have marketed products as AI-driven or automated when substantial work was actually performed by humans. In the so-called ghost work cases involving a series of SEC enforcement actions and private securities suits, companies were accused of presenting products as automated when human workers, often overseas, were doing the core tasks. Each company was alleged to have a gap between its public statements about its technology and how the technology actually worked. These cases are hard for plaintiffs to win, however. A recent investor lawsuit, for example, saw its complaint dismissed in April 2026, a point we come back to later.

A company can also get into trouble by overstating how much of its AI it actually owns. If a business licenses a third-party model and layers its own interface, workflow, or training data on top, a counterparty could question whether the company has accurately described which parts it owns, which parts it licenses, and how dependent the product remains on the vendor’s technology. Publicly available models may also be subject to licence terms, attribution requirements, field-of-use limits, redistribution limits, or other conditions that affect whether and how the resulting product can be described as proprietary. Before a company claims to own the technology behind its AI, it is worth confirming what it actually owns and what its contracts and licences permit.

This creates a genuine bind. Backing up its claims can at times require showing that the technology really works to audiences ranging from investors and regulators to opposing parties who, in litigation, can demand its internal records. The challenge is doing so without giving away the confidential know-how that is often its real competitive edge. That tension can often be managed through internal records, vendor documentation, and technical summaries that verify the claim without disclosing source code, model weights, or other competitive details.

Where the data used to train a model came from is its own source of risk. AI systems learn by being fed enormous amounts of existing material, and creators whose work was used without permission have been suing. The picture is still forming, but a pattern is emerging. In 2025, two federal courts found that training a model on lawfully obtained material may support a fair-use defence in circumstances where the use is sufficiently transformative and does not function as a substitute for the original works. Using pirated or improperly obtained material is a different matter, and one AI developer agreed to pay roughly USD1.5 billion to settle claims built on that conduct. These are only trial-court decisions, and higher courts have not yet weighed in, so the law is not settled, but the practical lesson for a company building on someone else’s model is already clear: how the underlying material was sourced matters, and the resulting risk does not automatically stay with the model’s maker. Depending on the contract, indemnity terms, representations, and the user’s own conduct, adversaries may claim that some of that exposure may remain with, or be passed to, the company deploying the model. Companies considering how to most accurately explain their AI usage and ownership may want to consider these topics in connection with their disclosures.

A second question runs in the opposite direction: who owns the copyright and other intellectual property in what the AI itself produces. Current US authority has answered at least the core point: material generated purely by AI, without meaningful human creative input, is not copyrightable, and in 2026 the Supreme Court left that rule in place. A person can still own AI-assisted work, but only to the extent a human genuinely shaped its creative expression. Entering prompts, however detailed, is not enough on its own. For companies, that may make documentation of human selection, arrangement, editing, and creative control especially important when AI-assisted output is described as an owned asset. For a company that publicly describes its AI, or the output that AI generates, as a valuable owned asset, this is a real check: some of what it is calling proprietary may not be something anyone can own.

Keeping careful records of training-data sources, licence conditions, vendor risk allocation, and human contribution to AI-assisted outputs does double duty: it reduces IP risk and gives the company evidence to support public claims about ownership, capability, and value. These IP facts can in turn become disclosure facts when they are material to the AI claim the company has chosen to make.

Operational AI-washing: the emerging theory

Litigation and enforcement to date have largely targeted “product AI-washing”, the overstatement of AI capabilities in what a company sells and the pattern at issue in the matters noted above. The emerging frontier is different. A new potential target for plaintiffs is “operational AI-washing”, a theory under which the plaintiffs’ bar may attempt to frame material corporate decisions, such as AI-driven workforce reductions and restructurings, as securities fraud. No court has yet issued a dispositive ruling on this theory, but the conditions and regulatory signals are already present. It relies entirely on a doctrine typically described as an actionable “half-truth”.

Generally, federal securities laws impose no affirmative duty to disclose broad economic headwinds in the abstract, in the absence of a reasonably likely material impact on the company. However, a company could potentially face disclosure risk if it highlighted AI as a principal driver of a workforce reduction while omitting other factors necessary to make that explanation not misleading. Plaintiffs could attempt to argue that omitting genuine liquidity or margin pressures from disclosure could convert an incomplete explanation into an actionable half-truth under the securities laws, and attempt to press that theory even against explanations management believed to be accurate. The half-truth doctrine has been invoked by plaintiffs to attempt to characterise technically accurate statements about software implementation as misleading claims about business operations. If a company highlights technology as a driver for a layoff, plaintiffs may search internal records for parallel cash-preservation efforts to attempt to build a liability theory.

When a company’s results weaken, leading to board discussions about potential causes and fixes, a plaintiff may argue that the company’s internal board materials no longer match its public messaging, which plaintiffs may attempt to characterise as operational AI-washing. Regulators are also increasingly signalling a co-ordinated focus on these internal narratives. The SEC expressly targeted artificial intelligence representations in its fiscal year 2026 examination priorities. This focus is mirrored by the Investor Advisory Committee’s push for standardised disclosures and growing scrutiny through issuer-level comment letters.

To reverse-engineer an argument of fraudulent or bad-faith intent, plaintiffs may hunt for specific corporate “red flags”. For example, they may look for capital expenditure ledgers devoid of the infrastructure spending required to support genuine transformation. They may target routine executive stock sales executed outside of prearranged trading plans. They may also search for any perceived absence of active board-level oversight.

Risks of these capability claims are likely to spill over into the employment docket. Wrongful termination plaintiffs can attempt to borrow pretext allegations from securities complaints to challenge the legality of layoffs. Separately, the underlying allegations of intentional fraud can generate coverage friction with directors and officers insurance carriers and other stakeholders.

The documentary record decides it

Whether an operational AI-washing claim survives usually comes down to one question: does the documentary record hang together? That record has two halves. The first is internal – the board materials a plaintiff can now reach. The second is external – the public statements the company has made. When the two tell the same story, the defence is strong; when they diverge, the gap may provide evidence for a plaintiff’s case.

On the internal side, the March 2025 amendments to the Delaware General Corporation Law provided additional guardrails on a plaintiff’s ability to obtain books and records from Delaware corporations. By confining pre-suit inspection demands largely to formal board materials, the amended Section 220 made the board record the primary battlefield. When plaintiffs obtain Section 220 documents, the exposure may concentrate in a few parts of the record: a capital-expenditure history showing little technology investment behind a claimed transformation; compensation minutes tying executive bonuses to headcount reduction rather than technological progress; and internal assessments describing a project as a pilot while SEC filings portray it as operational.

The same amendments codified the use of an incorporation-by-reference mechanism that had already been present in case law: a company may condition its 220 production on an agreement that all records produced under Section 220 will be deemed part of any later complaint on the same subject. A clean, consistent record therefore may provide pleading-stage evidence for dismissal, while a contradictory one may jeopardise an early dismissal. In securities actions, board minutes reflecting reliance on realistic technical assessments may help defeat allegations of fraudulent intent, and internal dissent. Recent litigation confirms the force of the scienter requirement in securities cases. In a recent securities suit built on AI-capability allegations, the court dismissed the complaint for failure to plead scienter with particularity, and after the plaintiff declined to replead or appeal; the case was dismissed with prejudice in June 2026.

The external side turns on the same discipline. The Private Securities Litigation Reform Act offers a safe harbour for forward-looking statements in certain circumstances but none for assertions of historical fact, and blending the two creates the mixed-statement trap. Consider a standard announcement: because the company’s automation programme is now fully operational, it is cutting its workforce in a move expected to save USD50 million. That single sentence welds a present-fact claim to a projection, and because federal courts differ on how much protection survives the blend, the safest course may be to draft disclosure for the strictest venue, assuming one historical claim could be argued to compromise the whole statement.

A short drafting protocol may help to minimise risk. Companies may want to keep statements of present operational reality separate from future expectations; pair projections with cautionary language specific to the actual risks debated in the boardroom, such as data-governance hurdles, vendor constraints, integration delays, or the persistent need for human oversight; and apply the same discipline across every channel. The most damaging slips may surface in lightly reviewed arenas, including earnings call remarks and investor day decks, which can carry the same exposure as disclosures in a formal filing. Aligning the total mix of information is what keeps the external record consistent with the internal one.

The board’s duty: Caremark in the AI era

As artificial intelligence permeates revenue generation and critical operations, the Caremark duty of oversight may also come into play. Although no Delaware court has assessed a Caremark claim in the AI context, it seems likely that such a claim would be evaluated through the frameworks established in cases like Boeing (aviation safety) and Marchand (food safety). Importantly, Caremark is a process-based standard: the question is whether the board made a good-faith effort to oversee the risk, not whether it adopted any particular mechanism. It is important for boards to consider this framework as they implement their AI protocols.

A practical starting point for any board in assessing the company’s use of AI is mapping the enterprise’s software footprint, since reconstructing that picture after a crisis emerges is considerably harder. From there, directors can focus oversight on four recurring dimensions: algorithmic bias; data integrity and privacy; model reliability; and AI-washing, the overstatement of AI capabilities in public disclosures, whether about products sold or operational decisions. Treating the risk of inaccurate disclosure as a governance question is part of the same exercise, and intellectual property questions around data-sourcing rights fall naturally within the privacy and integrity dimension.

This kind of oversight tends to rest on a measure of director fluency. Board members do not need engineering degrees, but directors who can critically evaluate management’s technical claims are better equipped to catch problems early. In practice, that means, among other things, noticing contradictions between internal metrics and public messaging and memorialising that scrutiny, which is also what builds a protective record.

Boards can strengthen their position by formalising this governance architecture. That can mean, among other things, designating committee ownership in a specific charter; setting written escalation thresholds for when an algorithmic failure or integration delay reaches the directors; expanding the internal audit scope to cover the gap between public capability claims and actual deployment status; and designating a single owner to check every external capability claim against the latest internal audit before publication. Together, this kind of reporting structure strengthens the company’s prelitigation defence.

Practical takeaways and the year ahead

Navigating change that comes with transformative new technologies comes down to rigorous alignment between internal reality and external narratives. Counsel can put several practical protocols in place to manage these capability claims:

• consider the private record, in the form of board decks and minutes, thoughtfully before authorising any press release;
• conduct vendor and licensing diligence to verify the actual intellectual property ownership of touted capabilities;
• document training-data provenance and confirm how infringement and indemnity risk is allocated with model vendors before publicising AI capabilities;
• ensure public efficiency claims are supported by corresponding technical budget line items;
• bifurcate historical facts from future projections in all public disclosures;
• coach executives on the legal risks posed by casual or imprecise statements during analyst calls; and
• give a named board committee formal ownership of AI oversight in its charter.

While effectuating all of these changes may not be required or necessary, depending on the facts, scale of the company, and its relative risk, consideration of whether to effectuate them may be appropriate.

Looking forward, we expect compliance burdens will compound. State-level algorithmic regulations are advancing. Illinois’ amendment to its Human Rights Act is already effective as of 1 January 2026. California’s automated decision-making regulations took effect on 1 January 2026, with specific obligations for significant employment decisions phasing in on 1 January 2027. Colorado, by contrast, repealed and replaced its landmark AI Act in May 2026, weeks before its effective date, substituting a narrower disclosure-based regime that takes effect on 1 January 2027. Combined with the governance and oversight obligations imposed by the European Union AI Act, whose high-risk obligations are being phased in through 2027 and beyond following recent amendments, the trajectory is moving toward documentation and highly standardised corporate transparency.

For the substantial concentration of public companies headquartered in Florida, geographic location offers no shield from Delaware’s corporate law and legal requirements. Because the vast majority of these companies incorporate in Delaware, they remain fully subject to Section 220 books-and-records demands and Caremark oversight requirements, regardless of where their executive suites reside. Florida’s own posture illustrates the national tension over AI regulation. The Governor-backed Artificial Intelligence Bill of Rights passed the Florida Senate overwhelmingly in the 2026 regular session and again in an April special session, only to be blocked both times by House leadership that views AI regulation as a federal matter, a legislative standoff that is likely to continue through future sessions. For Florida-based boards, the practical lesson is that statutory requirements remain in flux, while the disclosure and fiduciary exposure described in this article exists today, regardless of how the legislative standoff resolves.

Ultimately, the companies best positioned to withstand this scrutiny are those whose intellectual property rights, internal governance records, and public disclosures consistently tell the same story.