Everything AI

As generative AI becomes increasingly accessible and pervasive, many companies are experimenting with these tools to draft full legal agreements. For start-ups and businesses with limited resources, the appeal is obvious: AI can produce contracts quickly and at a fraction of the cost of traditional legal drafting. What once may have required days of attorney time and high legal fees can now theoretically be generated in minutes with a simple prompt.

While this efficiency is attractive, it comes with significant risks. AI-generated agreements frequently contain significant issues that ultimately make them unusable. Unlike an experienced attorney, AI lacks legal judgment, contextual understanding, and knowledge of specific client needs. It can generate provisions that appear polished on the surface but collapse under scrutiny.

One recurring issue is the inclusion of multiple conflicting provisions. For example, a recent AI-generated supply agreement included three different payment mechanisms for the same transaction: upfront payment, milestone-based installments, and payment upon delivery after invoicing. With three competing payment obligations for a single transaction, the contract leaves the payor guessing and sets the stage for potential disputes.

Another problem is language that reads like legal boilerplate but is, in fact, incoherent or legally meaningless. AI may generate complex wording that mimics the tone of a lawyer’s drafting, but it lacks the precision required in contract language. This can make agreements unnecessarily confusing and difficult to interpret. For example, a recent AI-generated commercial contract included the following dispute resolution provision:

“Any dispute, controversy, or claim arising out of or relating to this Agreement, or the breach, termination, or invalidity thereof, shall be finally settled by arbitration in São Paulo, Brazil, in accordance with the rules of the American Chamber of Commerce, Canadian Chamber of Commerce, or other mutually agreed renowned Chamber of Arbitrations.”

This dispute resolution clause has several problems. First, it does not identify the exact set of rules governing arbitration proceedings between the parties. More importantly, there are no such rules of the “American Chamber of Commerce” or the “Canadian Chamber of Commerce” and it is unclear what would be considered a “renowned Chamber of Arbitrations.” Finally, it is highly unlikely that parties resolving a dispute in Brazil would choose to apply American or Canadian-based arbitration rules. An attorney with even minimal experience in drafting arbitration provisions would not make these errors.

The underlying issue is that AI generates text based on patterns, not legal reasoning. It predicts what “sounds” like a contract without understanding enforceability, legal doctrine, or commercial practicality. Given the risks, companies should treat AI-generated contracts with caution. As such, attorney review remains essential to make sure agreements are clear, consistent, and tailored to the business’ needs.

Artificial intelligence (AI) continues to reshape healthcare delivery, particularly in diagnostics, imaging, and care coordination. For established healthcare platforms seeking to integrate innovative capabilities, effective transactions are not always single-step acquisitions. Phased relationships, beginning with commercial alignment and structured with the potential for full-scale integration, can provide a disciplined path to innovation.

For technology and life sciences transactions counsel, this raises a critical question: How can parties structure an initial collaboration to accelerate deployment while preserving long-term strategic optionality?

Phase I: Exclusive Commercial Alignment With Strategic Investment

In the healthcare sector, initial alignment frequently takes the form of an exclusive or field-limited license coupled with a minority equity investment. This structure enables a healthcare platform to secure access to differentiated technology, influence product development, and evaluate performance in a live clinical setting while mitigating integration and regulatory risk.

The effectiveness of this model turns on careful coordination in drafting foundational agreements.

Exclusivity

Exclusivity must be defined with intent and precision. In healthcare technology transactions, scope may be tied to specialty, site of care, modality, patient population, geography, or some combination of the foregoing. Performance milestones often are incorporated to ensure continued development and commercialization progress. Overly broad exclusivity may constrain a developer’s growth and create valuation tension; overly narrow drafting may undermine the strategic value of the arrangement. The objective is to balance long-term alignment with operational flexibility. Of course, in an AI context, parties need to address whether exclusivity extends to future model iterations, retrained versions, or algorithmic enhancements; consideration must be given to the dynamic nature of machine learning systems.

Intellectual Property and Data Rights

Healthcare collaborations frequently involve a convergence of technology developers and data-rich platform organizations. Clear delineation of background intellectual property, derivative works, algorithm improvements, and data usage rights is essential. Platforms typically seek rights to deploy and refine technology across their networks and express concerns around the control and use of their data; developers seek to preserve core intellectual property ownership and the integrity of their models having been trained with partner data. Sophisticated collaborators increasingly distinguish between ownership of model weights, fine-tuned models, embeddings, and generated outputs. Platforms also may be concerned with limiting use of their input data to improve only customer-specific instances rather than also generalized models. These provisions must be drafted with careful attention to the Health Insurance Portability and Accountability Act of 1996, state privacy regimes, cybersecurity obligations, and evolving regulatory scrutiny of AI-enabled tools.

Several intellectual property questions are distinctive in the AI context. First, where a party’s product is a layer built on a third-party foundation model (commonly referred to as a “wrapper”), the scope of any exclusivity grant requires careful scrutiny. The underlying model weights may remain owned and broadly licensed by the third-party foundation model provider. What a wrapper may practically protect is the specific configuration, fine-tuning, and proprietary data contributed by the parties; however, it likely cannot fence off the foundational capabilities beneath. Exclusivity in this context protects differentiation, not the underlying foundation model.

Second, ownership of AI-generated output remains legally unsettled, and third-party platform terms of use continue to shift in ways that increase that uncertainty. Current copyright doctrine in most jurisdictions requires human authorship; outputs produced autonomously by a model may provide no copyright protection at all. In practice, many organizations treat proprietary prompts, fine-tuning approaches, retrieval architectures, and workflow integrations as confidential know-how rather than relying on copyright protections in the outputs themselves. This approach can be effective, but trade secret protection is contingent on maintaining secrecy and on contractual controls that prevent disclosure. Strategic agreements governing AI development should address both the allocation of any rights that do arise as well as the confidentiality obligations, access controls, and use and disclosure restrictions necessary to sustain appropriate trade secret protection.

Third, background intellectual property carries additional complexity when applied to AI development. A developer’s model may incorporate vast quantities of third-party content, potentially including works subject to copyright claims or contested licensing. Training data provenance and any unlicensed works used in the development or subsequent retraining of a model represent material and ongoing diligence considerations. Parties collaborating on the development or deployment of a model should conduct appropriate diligence into training data lineage and should seek representations, warranties, and indemnities addressing third-party intellectual property infringement claims flowing from the model’s underlying composition.

Investment and Governance

A minority equity investment can accompany the commercial arrangement, aligning incentives and providing capital for continued development. Effectively structured governance rights, such as board observation rights, information rights, and other protective provisions, can provide strategic visibility without creating unintended control or regulatory implications. Coordination between the commercial agreement and the investment documents is critical, particularly with respect to transfer restrictions and change-of-control provisions. Moreover, AI governance may extend to oversight of model validation, change management processes, and any regulatory filings associated with algorithm modifications.

At this stage, sophisticated parties also frequently incorporate structural mechanisms that preserve a potential path to acquisition while maintaining flexibility for both parties.

Phase II: Commercialization Proving Grounds

The initial commercial phase often functions as live diligence. The healthcare platform gains insight into clinical adoption, workflow integration, model performance, regulatory positioning, and scalability. This operational visibility informs valuation and integration planning should the relationship mature. This phase can also serve as a period to assess real-world performance drift, explainability, robustness across diverse cohorts, and the sufficiency of monitoring systems required for ongoing compliance.

To preserve optionality for a transition toward a future acquisition, parties often incorporate structural tools with respect to equity, such as rights of first refusal or first offer, exclusive negotiation periods, call options, prealigned valuation frameworks, or some combination of the foregoing. Drag-along and tag-along rights associated with the equity may also be calibrated with a future acquisition scenario in mind. These mechanisms should be drafted with attention to enforceability and regulatory constraints. Early coordination among corporate, regulatory, and competition counsel can materially affect downstream optionality.

Phase III: Acquisition and Integration

If the collaboration demonstrates sustained strategic value, a full acquisition may enable a deeper operational integration and broader commercialization as well as enhance the platform’s market differentiation.

At this stage, issues addressed during the initial licensing phase assume heightened importance. Clear chain of title for intellectual property, documented rights to data and improvements, regulatory compliance history, and alignment across commercial agreements are central to confirmatory diligence. Particular scrutiny is given to training data lineage, model documentation, audit trails demonstrating controlled updates, and explainability materials required for U.S. Food and Drug Administration or equivalent regulatory submissions. Retention and incentive arrangements for founders, developers, and technical personnel often become critical to preserving innovation postclosing.

If lifecycle planning is embedded in the original agreements, the acquisition process typically becomes more efficient. Change-of-control consequences are understood, equity positions are aligned, and commercial rights do not require fundamental restructuring. The transaction can therefore focus on valuation, integration strategy, and long-term growth.

Deliberate Deal Architecture

As technology-enabled innovation continues to transform healthcare delivery, staged collaborations are likely to remain a defining feature of strategic transactions. A phased approach, beginning with commercial alignment and investment while preserving a clear path to integration, allows organizations to validate technology in practice before committing to full-scale acquisition.

For technology and life sciences transactions counsel, the task is not merely to document a license or negotiate a purchase agreement. It is to architect a relationship that accommodates immediate commercialization and future strategic flexibility within a highly regulated environment. Thoughtful lifecycle structuring at the outset can materially shape the trajectory, and ultimate success, of the deal. As AI becomes more deeply embedded in clinical workflows, deal architecture increasingly must anticipate issues such as ongoing model governance, regulatory obligations tied to adaptive algorithms, and safeguards around the responsible use of patient-derived data — all core components of strategic value and risk.

Since August 2, 2026, organizations have been subject to the transparency obligations set out in Article 50 of the EU AI Act. Article 50 introduces transparency requirements for providers and deployers in relation to certain AI system functionalities and use cases that may create transparency risks for individuals. Whilst much of the EU AI Act focusses on obligations on high-risk AI systems, Article 50 obligations may also apply to certain limited-risk systems. As a result, many organizations need to implement governance, disclosure, and content-labelling measures to ensure users are appropriately informed about the use of certain AI systems and AI-generated content.

Importantly, these obligations were largely unaffected by the implementation delays contained in the Digital Omnibus on AI (Regulation (EU) 2026/1744), except for the marking obligations contained in Article 50(2) of the EU AI Act, i.e. relating to AI systems that generate synthetic audio, image, video, or text content and that were placed on the market before August 2, 2026. Following the agreement reached between the Council and the European Parliament on May 7, 2026 (see the European Parliament procedure record), and formally adopted on June 29, 2026, the introduction of these marking obligations has been delayed until December 2, 2026.

In a Bloomberg Law article, Sidley lawyers Greg Lantier, Liv Herriot, and Raneem Eter examine recent federal court decisions clarifying lawyers’ obligations when using generative artificial intelligence (AI) in litigation. The article analyzes how courts are applying Federal Rule of Civil Procedure 11 to AI-assisted filings, emphasizing lawyers’ duties to verify citations, supervise AI-generated work product, and ensure the accuracy of court filings. The authors also outline practical governance measures law firms and legal teams can adopt to reduce AI-related litigation risks and strengthen accountability.

Courts are writing their own rules about generative AI, and in patent cases those rules differ sharply from judge to judge. This survey, current as of September 8, 2026, catalogs the generative-AI requirements in the venues where most patent infringement suits land: the Eastern and Western Districts of Texas, the Northern and Central Districts of California, the Northern District of Illinois, the District of New Jersey, and the District of Delaware and the U.S. International Trade Commission, where no AI-specific rules were identified. Only the Eastern District of Texas has a court-wide rule, making every litigant responsible for verifying AI-assisted filings. Elsewhere the obligations come from individual judges’ standing orders and vary widely, from restating the existing duty to verify what you sign to requiring a separate certificate naming the AI tool, confirmation that a human checked every citation, records of the prompts used, or disclosure of AI use in billing. Penalties include US$500 per violation and having the filing struck.

On June 2, 2026, President Trump signed an executive order called “Promoting Advanced Artificial Intelligence Innovation and Security” building on the cyber strategy his administration released in March 2026, which sets deadlines for federal agencies in three areas: using artificial intelligence (AI) to defend government and critical infrastructure systems, evaluating the most capable AI models, and prosecuting people who use AI to break into computers or commit fraud. It directs the Cybersecurity and Infrastructure Security Agency (CISA) to issue binding directives within 30 days, and the United States Department of Treasury to create an industry clearinghouse that coordinates vulnerability scanning and patching.

Within 60 days, agencies must build a classified benchmarking process to decide when a model counts as a covered frontier model, plus a voluntary program letting developers give the government early access before a wider release. For companies, the signal is that AI security and cybersecurity are being treated as one governance topic, and AI-enabled defenses may become a baseline expectation in government and critical infrastructure contracts.

In Fortis Advisors v. Krafton, the Delaware Court of Chancery awarded a striking equitable remedy: it ordered the reinstatement of a CEO who was wrongfully terminated using a strategy developed with ChatGPT, and extended the earnout period by the full duration of his improper exclusion. The buyer had used AI to strategize avoidance of a US$250 million earnout, but because that advice was generated outside the attorney-client relationship, it was not privileged and became damning evidence at trial. The decision puts buyers on notice that depriving seller-managers of bargained-for post-closing operational control may result in not just damages, but equitable remedies that restore control and reset the earnout clock.

The U.S. Food and Drug Administration is deepening its own use of artificial intelligence. Less than a year after launching its internal AI tool, Elsa, the agency released Elsa 4.0 and connected it to a new unified data platform that pulls together roughly 40 previously separate systems, so staff can query agency data directly instead of uploading documents each time. New features include custom AI agents, document generation, data analysis and visualization, and better search. FDA says humans stay involved at every step and that the model is not trained on data submitted by regulated companies.

On the same day, the then-Commissioner announced a pilot of “one-day inspections” for facilities that AI-driven risk analysis flags as lower risk. These are short screening visits meant to supplement, not replace, full inspections. More than 40 have been completed since April 2026, mostly ending with no action indicated, though inspectors can still issue formal observations and can expand the visit on the spot. Companies are not told in advance whether they are in the pilot, so life sciences startups with manufacturing or clinical sites should stay ready for either format. Complex or higher-risk operations, such as sterile manufacturing, remain on the standard inspection track.

Two February 2026 federal court decisions addressed how existing privilege and work product doctrines apply to generative AI, with one court (in United States v. Heppner) denying privilege and work product protection where an executive used an AI platform without adequate confidentiality safeguards and without counsel’s direction, and another (in Warner v. Gilbarco) holding that a pro se litigant’s AI-assisted analysis via a public AI platform was protected work product. These decisions signal that courts may apply existing doctrines to new technology in the legal context. Key considerations include whether the AI platform’s terms and privacy policies preserve confidentiality, the extent of attorney involvement in the AI-assisted work, and whether the materials were prepared in anticipation of litigation. Organizations should inventory AI tools and review their confidentiality terms, establish guardrails against AI use for legal analysis, and consider updating litigation hold protocols to account for AI-generated materials such as prompts and outputs.

Courts are now addressing how protective orders should govern the use of generative AI in discovery. Two federal decisions — Morgan v. V2X, Inc. and Jeffries v. Harcros Chemicals, Inc. — distinguish between “open” public AI tools that may retain and train on user data and more secure “closed” systems, and contemplate contractual safeguards before confidential discovery materials can be submitted to AI platforms. Practical takeaways may include addressing AI use explicitly at the outset of discovery and drafting concrete operational safeguards rather than referencing specific tools. Emerging issues that parties may anticipate include pro se litigants’ access to AI, AI’s influence on user behavior, and the conceptualization of AI within existing privilege and confidentiality frameworks.

The U.S. Copyright Office’s Part III report on AI and copyright signals that companies cannot assume generative AI training is automatically allowed under “fair use.” The report says copyright issues may arise at several stages: when developers collect and copy protected material for training the model, when a model memorizes protected material in its internal “weights,” and when AI outputs closely resemble copyrighted material. It emphasizes that fair use must be judged case by case, looking at why the original work was used, whether the use is commercial or transformative, how much of the original work was copied, whether guardrails prevent copying in outputs, and whether the AI system harms existing or potential markets for the original work. The report also points to licensing and possible policy reforms as ways to address uncertainty, but overall it takes a cautious approach that is more favorable to copyright owners than to unrestricted generative AI training.

As AI vendors become integral to business operations, their potential bankruptcy creates complex challenges under a Bankruptcy Code that has yet to fully adapt to AI technology. The first of a two-part series, this article focuses on intellectual property considerations: specifically, how companies that license AI tools or models can protect their continued access when a vendor enters bankruptcy. Among the key risks is a debtor’s ability to reject executory contracts, which could strip away the ongoing services that make a licensed AI tool usable, even where the bare license itself survives. Proactive contractual measures, including escrow arrangements and source code and deployment rights, can mitigate these risks before a vendor crisis occurs.

The second of a two-part series, this article examines consumer privacy risks and strategic implications that arise when an AI vendor files for bankruptcy. Key issues include how trained models, fine-tuned weights, and training data derived from sensitive customer-provided data may be treated as bankruptcy estate assets that can be sold to third parties, how those model artifacts can enable reconstruction or extraction of personally identifiable data even without a sale of the raw data, and whether courts will classify AI vendors as “critical vendors” or “utilities” — designations that carry significant consequences for service continuity and restructuring negotiations. Companies should address these risks proactively through contract terms and vendor risk management protocols before a crisis occurs.

With AI becoming central to business strategy, the legal questions surrounding it are keeping pace. As companies race to build and deploy AI systems, a critical question is emerging: Do their teams understand the data that powers these tools well enough to manage the risks? For some, the learning curve has come with real legal consequences.

Two categories of AI data sit at the center of these questions: training data and grounding data. Each raises meaningfully different legal and contractual concerns. For organizations working with AI developers and building their own models, understanding the distinction is an increasingly important part of getting AI right.

Training Data: The Textbook

Training data refers to the large datasets used during the development of an AI model. During training, the model analyzes large quantities of information, such as text, images, or software code, to identify patterns that allow the model to generate outputs when responding to prompts.

Think of training data as a student’s textbook.

Once students have read a textbook, they don’t carry it around to solve problems.

Training datasets can originate from a variety of sources, including:

• publicly available internet content, such as websites or online forums;
• digitized books, academic journals, and research publications;
• open-source software repositories;
• commercially licensed databases; and
• proprietary enterprise datasets (typically owned by the developer or provided to the developer by an external partner, customer, or collaborator).

Because these datasets often contain copyrighted or proprietary materials, and potentially personal information or other regulated data, the legal rights associated with training data can vary significantly. In some situations, the AI developer may own or curate the dataset itself (including, with respect to curation, to mitigate regulatory risks by using deidentified or synthetic data). In others, the underlying materials may remain owned by third-party publishers, content creators or owners, or database providers (and may still include data subject to various regulatory regimes).

Companies developing AI models internally may also incorporate proprietary operational data, customer information, or research datasets. When external collaborators are involved, training data may be governed by contractual agreements that define ownership, permitted uses, and rights to derivative models and datasets.

Importantly, training data is typically used during the model development phase. Once the model has been trained, the system generally no longer needs direct access to the underlying training dataset in order to function. The distinction can become relevant when licensing arrangements change or when organizations reevaluate their data governance strategies after a model has already been deployed.

Grounding Data: The Exam Fact Pattern

Grounding data serves a different purpose. Rather than teaching the model general patterns, grounding data provides contextual information at the time a model generates a response to a prompt. Organizations frequently use grounding to ensure that AI systems reference authoritative or up-to-date sources when producing outputs.

Think of grounding data as an exam fact pattern with specific facts and context.

Grounding data can include:

• internal company policies or compliance manuals (e.g., an employee handbook retrieved to answer an HR query);
• corporate knowledge bases and document repositories (e.g., past deal memos surfaced during contract negotiation);
• regulatory guidance or industry standards (e.g., Food and Drug Administration guidance retrieved to support a compliance question);
• proprietary databases (e.g., a licensed case law database queried at inference time); and
• user-provided documents or uploaded files (e.g., a contract the user uploads and asks the AI tool to summarize).

To make this concrete: When an attorney asks an AI tool, “What are the notice requirements under this contract?” the user input is that question. The grounding data is what the system automatically retrieves behind the scenes to answer it (e.g., the specific contract, the firm’s internal playbook, a relevant regulatory database). The user does not see that retrieval, but it drives the response.

A common implementation approach for grounding data is retrieval-augmented generation (RAG), where an AI system retrieves relevant documents and information from a database or search index and incorporates those materials into its response.

For businesses deploying AI tools internally, grounding can play an important role in maintaining guardrails around model outputs. By referencing trusted materials, grounding mechanisms can help improve model accuracy and reduce the likelihood that unsupported or misleading responses are generated.

From a legal perspective, grounding data presents a distinct set of considerations because it is actively retrieved and potentially reproduced at the time of each query rather than absorbed into the model’s parameters during the development phase. This distinction also has practical consequences postdeployment: A trained model generally continues to function even if access to its training data is later lost or a license expires, because the learned patterns are encoded in the model’s weights. Grounding data, by contrast, must remain continuously accessible, and loss of access or limitations on access to support risk management needs (e.g., cybersecurity, privacy, or other data governance purposes) may directly impair the model’s outputs.

The following table summarizes the key distinctions between training data and grounding data.

Litigation Trends: Focus on Training Data

To date, the most significant AI copyright decisions have focused on training data practices — specifically, whether copying protected works to build or improve a model is authorized and lawful. No court has yet issued a comparable ruling centered on grounding data or RAG systems, though legal risks in that area are building, as discussed below.

In Thomson Reuters Enterprise Centre GmbH v. ROSS Intelligence Inc., Thomson Reuters, the owner of Westlaw, sued ROSS Intelligence, a developer of an AI-powered legal research tool, alleging that ROSS used Westlaw headnotes as training data. ROSS argued that its use of the headnotes to train its model constituted fair use because the system learned from the material without reproducing it (i.e., without including the material in outputs). The U.S. District Court for the District of Delaware rejected that argument (in part) and found infringement, emphasizing that the use of proprietary, curated content to train a competing product weighed heavily against fair use. The decision suggests that training on copyright-protected materials, particularly where the resulting system functions as a market substitute, can give rise to copyright infringement risk even if the model does not use the original content in its output. Notably, the Court explicitly mentioned in its holding that only nongenerative AI was at issue in this case.¹

Two decisions issued in June 2025 by the U.S. District Court for the Northern District of California reached different conclusions in cases involving generative AI and together illustrate how fact-specific the fair-use analysis has become.

In Kadrey v. Meta Platforms, a group of authors sued Meta alleging that training its LLaMA large language model on copyrighted books constituted infringement. The Court found for Meta, concluding that the training use was a highly transformative fair use. The Court also found that the plaintiffs had failed to develop sufficient evidence of market harm (one of the central fair-use factors) and analyzed the copying and the training as a single, unified act rather than as separate instances of reproduction and use. The decision is notable, but its reach is limited: The Court declined to issue a broad ruling that AI training on copyrighted materials is categorically fair use, and the outcome turned heavily on the specific evidence presented (or lack thereof).²

In Bartz v. Anthropic, also decided in June 2025, the Court drew a sharper line based on how training copies were obtained. The Court found that training the Claude model on lawfully acquired or purchased books constituted fair use but declined to extend that protection to a central library Anthropic had compiled from pirated sources. The distinction (fair use for legally obtained copies, no fair use for pirated ones) has practical implications for how AI developers source and document their training data. The case settled for approximately US$1.5 billion following the ruling.³

Taken together, ROSS, Kadrey, and Bartz reflect a landscape in which outcomes turn on particular facts: what kind of content was used, how it was obtained, whether the resulting system competes with the original market, and what evidence of harm was presented. No appellate court has yet weighed in on the merits of AI training and fair use, meaning none of these decisions are binding precedent beyond its own jurisdiction. Significant cases remain in active litigation, including The New York Times Co. v. OpenAI and Microsoft in the Southern District of New York (the first major generative AI case brought by a news organization) and Disney v. Midjourney in the Central District of California, both of which are pending.^{4, 5} The appellate trajectory of ROSS, which is on interlocutory appeal to the Third Circuit, may also provide the first appellate guidance on these questions. For now, the law remains unsettled, and companies with significant exposure should monitor developments closely.

RAG Systems, Scraping, and Emerging Risks

As AI systems increasingly rely on grounding techniques such as RAG, new legal issues are beginning to emerge. Many RAG implementations retrieve information from external sources in real time, which may involve web scraping, application programming interface (API) access, or database queries. These practices raise potential risks that differ from traditional training-data disputes. That distinction drives the broader IP risk profile:

• Terms of Service — RAG systems that query external sources in real time may exceed what a website or API permits.
  
• Circumvention of access restrictions — retrieving content from paywalled or restricted sources can trigger additional liability beyond copyright (e.g., risks related to computer crime, privacy, or cybersecurity laws).
  
• Output reproduction — because retrieved text may appear verbatim in model responses, copyright infringement risk arises at the moment of each query, not just during development.
  
• Contractual limitations on database usage — licensed databases often do not contemplate AI systems incorporating their content into generated outputs, creating potential breach risk.

Organizations deploying RAG-based systems should therefore evaluate how their AI models retrieve and present external information. Practical risk-mitigation steps may include prioritizing vetted and licensed data sources, reviewing website terms before scraping content, and implementing technical safeguards that limit the amount of retrieved material incorporated into model outputs.

Contractual Considerations

Understanding the legal distinctions between training data and grounding data has direct implications for how organizations should structure their AI-related agreements. The two data types raise different contractual concerns, and companies entering into AI development or deployment arrangements should evaluate each category separately.

Training Data Agreements

When training data is sourced from third parties, whether licensed databases, research partners, or content providers, agreements should address several core issues. First, ownership and chain of title should be clearly defined: who owns or controls each component of the dataset, and who owns any derived datasets, curated subsets, or annotation outputs produced during the development process. This is particularly important where multiple parties contribute data or where the dataset is assembled by a third-party curator.

Second, the scope of the license should be carefully defined. Agreements should specify which model versions may be trained on the data, whether use is limited to internal deployment or extends to commercial products, and whether fine-tuned derivatives of the trained model fall within the license. Restricting sublicensing of training data to third parties is also worth addressing explicitly, particularly in multiparty development arrangements.

Third, agreements should address model ownership. Whether the trained model constitutes a derivative work of the training data (and if so, who owns it) is a contested question. Parties should allocate model ownership explicitly rather than leaving it to inference. This can be a significant negotiation point, particularly where a data provider is contributing proprietary content that materially shapes the model’s capabilities.

Finally, training data agreements should include representations and warranties from the data provider confirming that it has sufficient rights to license the dataset and that the licensee’s use of the dataset as permitted will not infringe third-party intellectual property rights. Indemnification provisions covering third-party infringement claims arising from the training data are increasingly standard in well-negotiated AI development agreements. Companies should also consider what happens when a training data license expires or is terminated postdeployment. As discussed, a trained model typically continues to function even without ongoing access to its training data, but audit rights and data deletion obligations may still apply and should be addressed in advance.

Grounding Data Agreements

Grounding data raises a distinct set of contractual issues, driven by the fact that grounding data likely must remain accessible throughout the life of the deployment. Unlike training data, which is consumed during the build phase, grounding data is actively retrieved on each query, making ongoing access a functional dependency rather than a historical one.

License agreements covering grounding data should be drafted to cover runtime retrieval specifically, not just one-time processing or storage. Traditional database licenses often do not contemplate an AI system incorporating retrieved content into generated outputs, a use that is meaningfully different from a human reading a record. Companies should audit their existing data licenses to confirm they extend to AI inference use before deploying RAG-based systems that draw on licensed sources.

Because grounding data may be directly reproduced in AI outputs, agreements should also address output ownership and attribution: whether the data provider retains any rights in AI-generated responses that incorporate its content and whether attribution or citation requirements apply. This is an emerging area, and parties should not assume that silence in a license means no rights are retained.

Where enterprise grounding data contains sensitive business information or trade secrets, confidentiality and data security provisions are essential. Agreements with AI vendors should specify how grounding data is stored and accessed, who within the vendor’s organization can retrieve it, and whether the vendor may use the data to train or improve its own models. This last point is often buried in vendor terms of service and deserves careful review.

Finally, because loss of access to grounding data can immediately impair a live AI system, agreements should include provisions addressing continuity and license termination. This may include uptime and availability service-level agreements for critical data sources, step-in rights if access is revoked, and data portability provisions allowing the organization to migrate to alternative sources without disruption. The operational stakes here are often higher than with training data, where the model continues to function even after the underlying data is no longer accessible.

Key Takeaways for Businesses

• For companies adopting AI technologies, data governance is a critical component of responsible AI deployment.
  
• Training data raises questions about IP rights, licensing, and model development practices, as well as privacy, regulated data, and cybersecurity-related risks.
  
• Grounding data affects how AI systems operate in real time and how organizations maintain reliability and control over model outputs and may present similar privacy, regulated data, and cybersecurity risks.
  
• As AI adoption continues to expand across industries, companies that carefully evaluate both aspects of the AI data pipeline will be better positioned to capture the technology’s benefits while minimizing legal and operational risk.

AI is reshaping drug discovery and development by accelerating research timelines, optimizing clinical strategies, and improving manufacturing efficiency, but adoption brings a complex web of legal considerations that life sciences companies must navigate. This article, published in Drug Discovery World, examines regulatory pathways for AI-assisted drug development, intellectual property ownership challenges, and data protection and privacy obligations arising from AI model training, and outlines practical steps organizations can take to capture the benefits of AI while managing regulatory, IP, and contractual risks.

On July 24, 2025, the California Privacy Protection Agency approved sweeping new regulations covering cybersecurity audits, risk assessments, and automated decisionmaking technologies (ADMT), pending review by California’s Office of Administrative Law. Certain businesses will be required to undergo annual independent cybersecurity audits beginning in 2027, phased in by company revenue, covering 18 components of a cybersecurity program and culminating in executive certification. Businesses must also conduct risk assessments before engaging in activities such as selling personal information, processing sensitive data, or using ADMT for significant decisions, with California uniquely requiring evaluation of “negative impacts” to consumers. Beginning April 1, 2027, businesses using ADMT for significant decisions must provide pre-use notices, access to information about how the ADMT works, and opt-out rights, though the regulations clarify that ADMT excludes technologies where a human retains authority to review and override the output. Additional provisions strengthen opt-out notification requirements and prohibit dark patterns in consent mechanisms, with several of these provisions already in effect as of January 1, 2026.

On May 21, 2026, the New York State Department of Financial Services (“DFS”) issued two coordinated Industry Letters: a letter on Heightened Cybersecurity Risks Associated with Frontier AI models (the “AI Advisory”) and accompanying Guidance on Measures Regulated Entities Should Consider in a Heightened Cybersecurity Threat Environment (the “Guidance,” and together, the “May 2026 Publications”). The AI Advisory builds on DFS’s October 2024 guidance on cybersecurity risks arising from artificial intelligence (AI), but is narrower in focus. Specifically, it addresses frontier models that may materially increase the speed and effectiveness of vulnerability discovery and exploitation.

The May 2026 Publications are not new rulemaking — both Industry Letters explicitly state so — but they are meaningful supervisory guidance: DFS identifies frontier AI Models as a technological development that may materially change the threat environment and instructs covered entities to evaluate whether their existing Part 500 programs remain adequate in light of that changed risk. The publications merit attention from DFS-regulated entities because they identify a specific class of emerging technology that DFS views as material to cybersecurity risk management under Part 500. That attention is warranted not only because the May 2026 Publications identify risks DFS views as material under Part 500, but also because DFS has cited prior Industry Letters in Part 500 consent orders, underscoring that such guidance can have practical supervisory and enforcement significance.

Generative artificial intelligence (GenAI) tools are becoming increasingly utilized by employees in the workplace to enhance productivity. However, employees are frequently leveraging these tools without formal approval or oversight. While the benefits of GenAI tools may be substantial, use of GenAI may pose significant risks to organizations, including legal liabilities, data breaches, and reputational damage. This article explores certain key concerns associated with employee use of GenAI tools and essential components of an AI use policy.

Understanding the Risks

Data Privacy and Security

One of the most significant risks with the use of GenAI is that information entered into prompts or uploaded to common GenAI platforms may reveal to the AI tool sensitive or confidential company information or information about clients or customers, including information the company may have contractually agreed to keep confidential. Similar risks exist with inputs of personally identifiable information, such as email addresses or other identifiers. Typically, on publicly-available versions of such tools, data that is entered is used to train the AI or may otherwise be stored or used by the AI tool developer.

To mitigate risks of exposing potentially sensitive company data or potentially breaching customer confidentiality agreements, companies may want to consider purchasing enterprise versions of GenAI tools that operate in a closed-loop system, meaning the developer is not permitted to use or store inputs for purposes other than to provide the GenAI tool. Companies may also want to have an approved list of AI tools, which might include enterprise versions of common GenAI tools. Also helpful in mitigating risks is employee training and raising awareness of these risks.

Another significant and emerging risk is that GenAI tools are increasingly being used as a tool for cybercriminals, including through the orchestration of complex injection attacks that exploit the fact that these tools do not think, but rather, simply follow instructions. For example, attackers can embed instructions in publicly available links that, when uploaded to the GenAI tool, tell the tool to maliciously attack the host system and override instructions. Alternatively, the embedded instructions in public links could produce PDFs or other outputs that themselves contain malicious code. Companies should consult with their security team about the risks unique to their business and implement recommended controls if applicable. In addition, companies may want to consider working with external cybersecurity training providers to ensure their materials include coverage of AI tool injection attacks in employee training modules.

Intellectual Property Concerns

Companies should also seriously consider the significant intellectual property (IP) risks posed by employees’ use of GenAI tools. These IP risks arise from two factors: (1) the material used as “input” to train the AI models; and (2) the material generated by the AI model as “output”. AI models are frequently trained on vast datasets which include copyrighted material used without the author’s permission. Furthermore, use of such copyrighted material may cause the GenAI tool to generate outputs that replicate or closely resemble the copyrighted material, further exposing the user to potential copyright infringement claims.

Another IP risk associated with employees’ use of GenAI tools is the lack of IP protection for works and innovations created using GenAI. In the United States, works and innovations created solely by AI without human contribution are not eligible for copyright or patent protection. Without such protections, a company’s competitive advantage may fall into the public domain, where anyone can use it for free.

Use of GenAI may also compromise trade secret protection of a company’s proprietary information. If an employee inputs confidential proprietary information into GenAI tools without proper contractual protections, such information may be used to train the underlying AI models. As noted above, GenAI tools may generate outputs that replicate or closely resemble the inputs, in this case, the proprietary information, potentially exposing the trade secret to other users and compromising the company’s trade secret protection of the information.

Hallucinations and Bias

AI systems, particularly generative models, can produce outputs that appear plausible but are factually incorrect or nonsensical, a phenomenon known as “hallucinations.” These inaccuracies can lead to misguided business decisions, flawed reports, and erosion of trust in AI tools. For instance, an AI-generated report might cite non-existent studies or fabricate data, potentially leading to compliance issues or reputational damage.

Shadow AI Usage

Employees may use unauthorized AI tools (sometimes referred to as “shadow AI”) without the knowledge or approval of IT departments. This practice can result in data security breaches, compliance violations, and loss of control over sensitive information. For example, inputting confidential company data into unvetted AI platforms could inadvertently expose proprietary information. Companies should implement strict policies around shadow AI use and consider limiting access to popular AI tools on workplace computers.

Regulatory Compliance Challenges

The rapidly evolving landscape of AI regulations, such as the European Union’s AI Act, presents compliance challenges for organizations. Failure to adhere to these regulations can result in legal penalties and hindered innovation. Organizations must stay informed about regulatory changes and implement robust governance frameworks to ensure compliance.

DO’S & DON’TS:

Legal Compliance & Policy Design

DO:

• Establish a clear, tailored AI usage policy in collaboration with legal, IT, and compliance teams.
• Define “AI” in plain language, and specify which tools are approved.
• Include a mission or purpose statement to ground the policy in transparency, fairness, and innovation.
• Update the policy regularly to reflect evolving laws (e.g., EU AI Act, state legislation like Colorado’s).

DON’T:

• Don’t use one-size-fits-all policy templates. Each policy should reflect specific business risks and use cases.
• Don’t treat the AI policy as a one-time exercise. Legal exposure grows without active maintenance.

Data Security & Confidentiality

DO:

• Prohibit inputting sensitive, proprietary, or customer data into public AI tools unless specifically authorized.
• Include clear confidentiality provisions related to AI use in employee and vendor contracts.
• Conduct vendor due diligence to confirm AI platforms’ compliance with privacy laws and data handling standards.
• Consider using enterprise versions of AI tools to keep data inputs confidential.

DON’T:

• Don’t assume that AI tools won’t retain user input; review each tool’s terms of service and retention practices.
• Don’t allow shadow IT (e.g., unauthorized browser plugins, personal-use tools) without approval or audit trails.

Intellectual Property & Ownership

DO:

• Review AI vendor agreements to ensure the vendor has proper rights to use training data and provides adequate protections if infringement claims arise.
• Implement human review processes for AI-generated materials to identify potentially infringing content.
• Document human contributions in connection with the development of AI-assisted works and innovations.

DON’T:

• Don’t allow employees to generate customer-facing or public materials without review for potentially infringing content.
• Don’t use AI models trained on unlicensed data without understanding the risk of potential IP infringement claims.
• Don’t allow employees to input confidential proprietary information into AI tools without proper confidentiality and use restrictions.

Training, Governance, & Oversight

DO:

• Require regular employee training on the capabilities, limits, and risks of AI tools.
• Designate responsible roles (e.g., AI Compliance Lead, Tool Vetting Committee) to oversee implementation and approvals.
• Monitor AI tool usage, audit outputs periodically, and solicit user feedback to improve policies.

DON’T:

• Don’t deploy AI tools enterprise-wide without pilot testing, user education, and risk assessment.
• Don’t assume all departments or roles should have equal access to AI tools — access should be role-specific.

Ethical Use & Bias Prevention

DO:

• Embed fairness, transparency, and accountability principles into the company’s AI governance framework.
• Monitor AI systems for evidence of bias, particularly in HR, marketing, or customer interaction tools.
• Require disclosures when AI is used to create or assist in work product — internally or externally.

DON’T:

• Don’t use generative AI content without fact-checking — misleading outputs could carry reputational or legal costs.
• Don’t let AI output go unlabeled in sensitive contexts (e.g., client deliverables, compliance filings).

Although AI-related patent filings have surged dramatically over the past two decades, patent litigation involving AI technologies has not yet kept pace, a trend the article suggests may mirror the delayed wave of smartphone patent litigation that followed the technology’s commercial growth. In this piece for Bloomberg Law, Sidley partner Alexis Cohen argues that AI companies should begin preparing now for increased future litigation by evaluating and strengthening their intellectual property protections, including patents and trade secrets, while also improving documentation, ownership records, employee agreements, and internal processes.

Reproduced with permission. Published May 21, 2026. Copyright 2026 Bloomberg Industry Group 800-372-1033. For further use, please visit: Bloomberg Industry Group copyright and usage guidelines.

On July 23, 2025, the Trump administration released its AI Action Plan, aimed at ensuring U.S. global dominance in artificial intelligence, outlining 90 federal policy positions across three pillars: Accelerating Innovation, Building American AI Infrastructure, and Leading in International Diplomacy and Security. The plan calls for streamlining regulatory pathways, expanding public-private partnerships, increasing federal AI research investment, and promoting open-source AI models as potential global standards, while simultaneously encouraging export of the full “American AI Technology Stack” to allied nations as a geopolitical strategy. Three accompanying executive orders promote export of American AI technology, streamline federal data center permitting, and mandate adoption of ideologically neutral “Unbiased AI Principles” in federal procurement. The Action Plan emphasizes deregulation at the federal level and conditions federal discretionary funding on states’ AI regulatory environments, reflecting growing tension between the administration’s national AI framework and state-level regulatory efforts.

England’s High Court has issued the UK’s first ruling in a copyright fight against an AI developer with an outcome that favors the AI industry. In Getty Images v. Stability AI, Getty (which licenses photos and video) alleged that Stability scraped millions of its images to train the Stable Diffusion image generator. During trial Getty dropped its main copyright claim because it could not show the training actually happened in the UK. Copyright is territorial, so no act in the UK means no UK infringement.

That left a narrower question: whether importing the already-trained model into the UK made it an “infringing copy” under the Copyright, Designs and Patents Act 1988. The court said no. A model’s weights are just numbers derived from statistics; they do not store or contain the underlying images, so supplying the model is not secondary infringement. Getty won only on a historic and limited trademark point about its marks appearing in outputs. The decision settles little about whether UK-based training would infringe, and Getty may appeal.

The UK’s Medicines and Healthcare products Regulatory Agency moved on several fronts in the first half of 2026, with the intention of developing a more innovation-friendly and AI-aware regime. The developments include a landmark program deepening regulatory cooperation between the UK and the United States, a regulatory “sandbox” for AI used as a medical device, meaning a supervised space where companies can test products with the regulator before full authorization, an evidence-gathering exercise on how AI in healthcare should be regulated, a consultation on a draft Rare Disease Therapies Regulatory Framework, and the largest package of UK clinical trial reforms in more than twenty years. For a founder building AI-enabled diagnostics, software, or therapeutics, this changes both the route to market and its timing: earlier regulator engagement is becoming possible, trial rules are being streamlined, and closer UK-US alignment may reduce duplicated work across the two markets.

President Trump’s December 11, 2025 Executive Order aims to protect American AI innovation from what it characterizes as overly burdensome state AI laws by creating a Department of Justice AI Litigation Task Force to identify and challenge problematic state regulations, conditioning federal Broadband Equity, Access, and Deployment (BEAD) Program funding on state AI policy alignment, directing the Federal Communications Commission (FCC) to consider preemptive federal AI reporting standards, and instructing the Federal Trade Commission (FTC) to clarify how Section 5 of the FTC Act applies to AI in ways that could preempt conflicting state laws. The Executive Order stops short of imposing the broad 10-year moratorium on state AI laws that was previously proposed in legislation, instead taking a more targeted approach that carves out state authority over child safety, AI infrastructure, and government AI procurement. The Executive Order also calls on Congress to enact a uniform federal AI framework.

On July 7, 2026, the European Data Protection Board published draft guidelines on web scraping for generative AI, aimed at giving practical guidance on how EU data protection law applies when personal information is pulled from the open internet to train models. The guidelines cover both organizations that do the scraping themselves, directly or through a contractor, and organizations that license or reuse datasets someone else already scraped. For any company buying a generative AI system trained on internet-sourced data, the guidelines, as drafted, do not shift the data protection responsibility away from the buyer. The EDPB’s public consultation on the draft remains open until October 30, 2026, so the text can still change before it is finalized.

In an article for Life Science Leader, partners Sally Wagner Partin, Sharon Flanagan, and Torrey Cope examine how artificial intelligence is reshaping life sciences M&A by changing the pace and economics of drug development. They explore how AI’s growing role in target identification, clinical development, and regulatory decision-making is influencing valuation, due diligence, milestone structures, transaction timing, and risk allocation, and outline key considerations for dealmakers as AI continues to transform the life sciences industry.

A federal judge in Manhattan ruled in May 2026 that the government could not blame ChatGPT for its own decision to terminate more than 1,400 federal grants that the tool had flagged as related to diversity, equity, and inclusion. Staff fed short grant descriptions into the model and asked whether each one related to DEI, then acted on the answers without reading the underlying applications, without defining what DEI meant, and without a single instance of a reviewer disagreeing with the tool. The Court treated the prompts and outputs as discoverable evidence, noting that a decision-maker who chooses a tool cannot escape responsibility by blaming it.

For any company building generative or agentic AI into core business functions, the practical points are about governance: who is allowed to write prompts, whether reviewers have the expertise and authority to override an output, and whether the process is documented. AI-generated records can later surface in litigation or regulatory review, accurate or not.

On June 22, 2026, the White House issued two executive orders on quantum technology: “Ushering in the Next Frontier of Quantum Innovation” and “Securing the Nation Against Advanced Cryptographic Attacks.” Quantum computers can solve certain problems far faster than ordinary computers, which opens up new commercial possibilities but also threatens the encryption that protects most online communication and stored data today. The orders push federal agencies past pure research toward deployment, commercialization, workforce building, and national security uses.