AI and Emerging Technology Law: A Roadmap and Resource Toolkit

Technology > Artificial Intelligence | Published 27 February 2025 - Updated 12 March 2025 | UpLaw - uplaw.us

A stage-by-stage map of AI and emerging-technology law: output ownership, machine inventorship and authorship, the training-data fair-use war, biometric and privacy constraints, products liability, AI contracting, governance programs, and the drone and crypto precedents that show where this is all heading—each stage linked to primary authority and deeper UpLaw resources.

Keywords: artificial intelligence law, emerging technology law, generative AI, AI inventorship, AI authorship, training data fair use, Thaler, biometric privacy, BIPA, web scraping, AI products liability, AI contracting, AI governance, NIST AI RMF, EU AI Act, drone regulation, cryptocurrency regulation

---

How to use this toolkit

Four people walk into four different problems, and they are all the same problem. A founder wants to ship a generative-AI feature by the end of the quarter and needs to know whether the images it produces are hers to sell. A patent attorney fields a call from an inventor who says, a little sheepishly, that the model did most of the work. An in-house lawyer has a vendor's model-license on her desk with an indemnity clause that carves out the one risk she is most worried about. A privacy officer just learned that the company's new hiring tool scores candidates by analyzing recorded video interviews. None of them is facing a brand-new field of law. Each is facing an old body of law—copyright, patent, contract, privacy, tort—straining to decide how it applies to a machine that was never in the picture when the rules were written.

That is the organizing idea of this toolkit. Artificial intelligence did not create one new legal subject; it put simultaneous pressure on a dozen existing ones. The lawyer who does this work well is, before anything else, a good issue-spotter, because the hardest part is usually recognizing which doctrines a single product has quietly triggered. What follows is a map of the whole terrain and a staged walk through the decision points that matter, with each stage pointing to the primary authorities worth reading and the deeper UpLaw treatments that handle each subject at length. For the two book-length surveys that sit behind this roadmap, start with Key Legal Issues in Artificial Intelligence: An Overview and its fuller companion, Key Legal Issues in Artificial Intelligence: A Comprehensive Guide for Business and Counsel.

One caution belongs at the top, not buried in a footnote. AI law is among the least settled areas a practitioner can work in right now. Marquee copyright suits are mid-stream and producing contradictory district-court reasoning; the U.S. Copyright Office is publishing its guidance in installments; the USPTO revised its inventorship rules only in 2024; and legislatures at the state, federal, and foreign level are all drafting at once. Where this toolkit states a rule, treat it as the current best reading. Where it flags uncertainty, treat that as the real state of the world and verify before you rely.

The three-layer model: a frame for spotting the issue

Before drilling into any single doctrine, adopt a habit that makes the rest of the analysis cleaner. Separate every AI system into three layers, and figure out which layer your problem lives in.

The model is the trained system itself—the weights, the architecture, the fine-tuned parameters. It is the asset a developer builds and guards, and it is usually protected, if at all, as a trade secret rather than by copyright or patent. The inputs are everything that flows in: the training corpus assembled during development, and the prompts, documents, and data a user feeds in at run time. The outputs are what the machine emits—an image, a paragraph, a line of code, a risk score, a molecule.

Most AI questions resolve more crisply once you name the layer. A dispute over whether scraping a news archive to build a model was lawful is an input problem, governed by copyright and the law of unauthorized access. A dispute over who owns a generated logo is an output problem, governed by copyright's authorship requirement and the tool's terms of service. A dispute over whether the developer misappropriated a competitor's proprietary weights is a model problem, governed by trade-secret law. Then add the second axis—the actors. A developer trains the model, a deployer integrates it into a product or workflow, and an end user operates it. Liability and ownership routinely turn on who did what, and the EU AI Act, discussed below, formalizes exactly this developer-versus-deployer distinction into differentiated legal duties. Keep both axes in mind and the issues announce themselves.

Stage 1 — Map the terrain: which bodies of law AI stresses

AI counsel earns their keep at the spotting stage. A single deployment can implicate, in one breath, half a dozen regimes that each evolved for human conduct:

• Copyright governs both the training corpus (was the copying infringement, or fair use?) and the outputs (is a generated work protectable, and by whom?).
• Patent governs AI-assisted inventions—both whether the underlying algorithms are patent-eligible subject matter under Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014), and whether a human contributed enough to be named an inventor.
• Trade secret protects the model weights, the training pipeline, and the prompt-engineering know-how that a developer cannot or will not patent. UpLaw's Building Trade Secret Protection from the Ground Up: A Comprehensive Guide maps the reasonable-measures obligation that a model owner must satisfy.
• Privacy and data protection bound the personal and biometric data that flows into training sets and inferences, from Illinois's biometric statute to the GDPR.
• Tort and products liability answer for the harm a hallucinated answer, a biased score, or a mis-steered autonomous system causes.
• Consumer protection reaches deceptive or unfair AI conduct through the FTC's authority under Section 5 of the FTC Act, 15 U.S.C. § 45.
• Employment law governs automated hiring, scoring, and monitoring tools.
• Sector-specific regulation—in health, finance, and insurance—overlays everything with its own rules.
• Antitrust lurks wherever algorithmic pricing or data concentration raises collusion or monopolization concerns.

Practical Law's overview of the field organizes the exposure this same way, running from commercial contracts and products liability through privacy, IP, employment, and antitrust. The lesson for the practitioner is procedural rather than doctrinal: do not analyze an AI matter as an "AI question." Decompose it into the specific legacy doctrines it touches, run each to ground, and only then reassemble the picture. For the historical context—how courts have repeatedly bent old doctrine around new computing technology, from software copyright through the CFAA to AI—see The Evolution of Computer Law: From Software Copyright to Artificial Intelligence.

Stage 2 — Who owns what the machine makes

Ask a founder who owns the image a text-to-image tool just generated and the confident answer is usually "we do—we paid for it." The honest answer is stranger: in the United States, purely machine-generated output may be owned by no one at all, because copyright attaches only to works of human authorship.

That principle is old, not new. The Supreme Court grounded copyright in human creativity as far back as Burrow-Giles Lithographic Co. v. Sarony, 111 U.S. 53 (1884), which upheld protection for a photograph precisely because a human had made the creative choices of pose, lighting, and arrangement. Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991), reaffirmed that originality—"the touchstone of copyright"—requires at least a modicum of human creative judgment, not mere effort or mechanical compilation. The Ninth Circuit even confirmed, in the famous "monkey selfie" case, that a non-human cannot hold a copyright: Naruto v. Slater, 888 F.3d 418 (9th Cir. 2018). Apply that lineage to a diffusion model and the result follows. Where a human merely typed a short prompt and accepted whatever appeared, there is little human authorship to protect. Where a human iteratively selected, arranged, masked, and edited the machine's output, those human contributions may be protectable even though the raw generation is not.

The Copyright Office has translated this into registration policy. Its March 2023 statement on works containing AI-generated material, 88 Fed. Reg. 16190 (Mar. 16, 2023), instructs applicants to disclaim the AI-generated portions and to claim only the human-authored contributions—the selection, coordination, and arrangement a person supplied. Part 2 of the Office's ongoing report series, Copyright and Artificial Intelligence (copyright.gov/ai), develops the human-authorship analysis in detail. The upshot for the ownership question is that copyright protection in AI output is a matter of degree, measured by the density of human creative control over the specific result.

Because copyright can be thin or absent, the second track—contract—usually does the real work. Virtually every commercial AI tool's terms of service allocate whatever rights exist between the provider and the user, assign output to the user (or license it), and disclaim warranties about it. Read those terms before you assume you own the output, and certainly before you assume you may use it commercially or exclusively. Ownership of AI inventions, as distinct from creative works, follows a parallel logic and is treated in depth in Who Owns AI-Generated Inventions? Machine Creativity and Patent Ownership.

Worked example. A boutique branding agency runs a text-to-image model to generate roughly two hundred candidate marks for a coffee brand, narrows them to one, and then spends a week in a vector editor redrawing the cup, recoloring the palette, and hand-kerning the wordmark. The raw generation is almost certainly uncopyrightable machine output. The agency's substantial, documented human reworking is a protectable contribution in those human-authored elements—and the registration should claim exactly that, disclaiming the AI base under the 2023 guidance. Critically, the client engagement letter should assign whatever rights exist by contract and treat the deliverable as a work made for hire under 17 U.S.C. § 201(b), rather than trusting copyright's defaults to hand the client anything at all. If the agency skips that step, it may deliver a logo that neither party can stop a competitor from copying.

Stage 3 — Can a machine be an inventor or author? The threshold "no"

Every AI ownership conversation eventually hits the foundational question: can a non-human qualify as the inventor or author in the first place? U.S. courts and agencies have answered consistently and firmly. No.

On the patent side, the Federal Circuit held in Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), cert. denied, that an "inventor" under the Patent Act must be a natural person, so the AI system known as DABUS could not be named as an inventor on a patent application. The court read the statute's text closely—the Patent Act repeatedly refers to inventors as "individuals" and uses personal pronouns, see 35 U.S.C. §§ 100(f), 115—and found no room for a machine. The district court had reached the same conclusion in Thaler v. Hirshfield, 2021 WL 3934803 (E.D. Va. Sept. 2, 2021). On the copyright side, the D.C. Circuit affirmed in 2025 that a work generated autonomously by an AI, with no human author, is ineligible for registration; the district court's decision below is reported at Thaler v. Perlmutter, 687 F. Supp. 3d 140 (D.D.C. 2023). Both holdings track the same human-authorship and human-inventorship spine that runs from Burrow-Giles through Naruto.

Read those holdings narrowly, because their reach is narrower than the headlines suggest. Neither case says that AI-assisted inventions are unpatentable, or that AI-assisted works are uncopyrightable. Each says only that the named inventor or author must be a human who made the kind of contribution the relevant law requires. That leaves the live question exactly where practitioners now spend their time: how much human contribution is enough?

For patents, the USPTO answered in its 2024 rulemaking, Inventorship Guidance for AI-Assisted Inventions, 89 Fed. Reg. 10043 (Feb. 13, 2024). A natural person must have made a "significant contribution" to the claimed invention, measured using the joint-inventorship factors from Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998). Merely owning the AI, merely recognizing a problem, or merely running the system and appreciating its output does not, by itself, make a person an inventor. Conversely, a person who conceives a specific solution using AI as a sophisticated tool, or who reduces an AI-suggested idea to a concrete, patentable form through inventive judgment, can be a proper inventor. The guidance walks through hypothetical fact patterns to show where the line falls, and PTO practice will keep refining it. The comparative dimension—how the EPO, UK, and major Asian patent offices treat the same question—is developed in AI and Inventorship: A Global Comparison of Machine Contributions to Patents.

Worked example. A medicinal chemist prompts a generative model to propose novel kinase-inhibitor scaffolds. The model returns forty candidate structures. The chemist, drawing on years of judgment about metabolic stability and synthetic accessibility, selects three, designs a synthetic route, runs assays, and iterates one candidate into a compound that actually binds the target and survives liver-microsome testing. Under the 2024 guidance, the chemist's selection, design, and reduction to practice look like a significant Pannu contribution—the patent should name the chemist, and only the chemist, as inventor. Naming "DABUS" or the model as a co-inventor would be both legally impossible after Thaler v. Vidal and a self-inflicted validity problem. But note the flip side: if the chemist had simply accepted the model's top-ranked structure without inventive judgment, there may be no valid human inventor at all—and, under current law, nothing patentable. That gap is one of the genuinely unresolved frontiers of AI patent practice, and it is why documentation of the human's conception is now a core prosecution discipline, covered in Prosecuting a Patent Application: A Roadmap and Research Guide.

Stage 4 — Training-data copyright: the fair-use war

The most financially consequential AI litigation on the docket has almost nothing to do with outputs and everything to do with inputs. To build a large model, a developer copies enormous quantities of text, images, code, or audio—much of it copyrighted—into a training pipeline. Rights holders say that copying is straightforward infringement of the reproduction right under 17 U.S.C. § 106. Developers respond that the copying is a transformative fair use under 17 U.S.C. § 107, because the model learns statistical relationships rather than storing or republishing expression. Whoever is right will reset the economics of the entire industry, and as of this writing the question is genuinely open.

The battleground is the familiar four-factor fair-use test, applied to unfamiliar facts. Factor one asks about the purpose and character of the use, and after Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023), a defendant can no longer wave the word "transformative" and win; the Court insisted that a new purpose must be weighed against the copying's commercial character and its degree of substitution. Developers lean on the line of cases that blessed mass, functional copying: Authors Guild v. Google, Inc., 804 F.3d 202 (2d Cir. 2015), which held Google's scanning of millions of books to build a searchable index transformative; its companion, Authors Guild v. HathiTrust, 755 F.3d 87 (2d Cir. 2014); and Google LLC v. Oracle America, Inc., 141 S. Ct. 1183 (2021), which treated the reuse of functional software interfaces as fair. The transformative-use tradition traces to Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994), and Sony Corp. of America v. Universal City Studios, Inc., 464 U.S. 417 (1984). Rights holders counter with factor four—market harm—and a pointed factual argument: if a model can be prompted to emit near-verbatim reproductions of training works ("memorization"), the fair-use defense weakens, and a licensing market for training data now plainly exists, so unlicensed copying displaces it.

Early rulings split, and they split on their facts. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. (D. Del. 2025) rejected a fair-use defense where a defendant copied headnotes to build a directly competing legal-research tool—non-generative, and a close market substitute. In the generative context, two San Francisco judges reached partly opposite conclusions within weeks of each other in mid-2025: in Bartz v. Anthropic PBC (N.D. Cal. 2025), the court treated training a large language model on books as transformative in principle, while holding that building a permanent library from pirated copies was not excused by the eventual fair use; in Kadrey v. Meta Platforms, Inc. (N.D. Cal. 2025), the court granted the developer summary judgment on the record before it while warning that a better-developed showing of market dilution might have changed the outcome. The bellwethers everyone is watching—The New York Times Co. v. Microsoft Corp. (S.D.N.Y.), Getty Images (US), Inc. v. Stability AI, Inc. (D. Del.), and Andersen v. Stability AI Ltd. (N.D. Cal.)—turn heavily on whether outputs reproduce protected expression and whether the underlying copies were lawfully obtained. A distinct but related theory travels under the copyright-management-information provisions of the DMCA, 17 U.S.C. § 1202, where plaintiffs allege that training stripped author and rights metadata. The Copyright Office addresses the training stage directly in Part 3 of its AI report. UpLaw tracks the whole docket in Generative AI on Trial: Copyright Infringement Claims After the New York Times and Getty Suits.

There is a prior question the fair-use debate often skips: how the developer acquired the data. Much training data is scraped, and scraping sits at the intersection of the Computer Fraud and Abuse Act, breach-of-contract terms of service, and copyright. After hiQ Labs, Inc. v. LinkedIn Corp., 31 F.4th 1180 (9th Cir. 2022), and the Supreme Court's narrowing of the CFAA in Van Buren v. United States, 141 S. Ct. 1648 (2021), scraping publicly available data is less likely to be a federal computer-crime, but it can still breach a site's terms and can still infringe if it copies protected expression. That tangle is mapped in Web Scraping and the Law After hiQ v. LinkedIn: Where the CFAA, Contract, and Copyright Collide.

Worked example. A startup fine-tunes an open-weights image model on a corpus it assembled by scraping a stock-photography site, ignoring the site's terms that prohibit automated harvesting. Its product later generates an image that a user prompts toward a specific, recognizable photograph in the training set, and the output is strikingly similar. The startup faces a stacked set of claims: direct copyright infringement in the input copying; a § 1202 claim if it stripped the photographer's embedded rights metadata; a breach-of-contract claim under the site's terms; and an output-infringement claim on the strikingly similar image. Its fair-use defense is now fighting uphill on the Warhol factors, and the "pirated corpus" wrinkle from Bartz means even a favorable transformative-use ruling on training may not save the piece of the case built on unlawfully obtained copies. The lesson many developers are learning the expensive way: provenance of training data is now a first-order legal question, not a data-engineering detail.

Stage 5 — Biometrics and privacy as hard constraints on the data

A model is only as lawful as the data it eats, and where that data includes faces, voices, fingerprints, gait, or other bodily identifiers, biometric-privacy law becomes a hard constraint that bites at the development stage, long before deployment.

Illinois's Biometric Information Privacy Act (BIPA), 740 ILCS 14, is the sharpest instrument in the field, for one structural reason: it gives individuals a private right of action with liquidated statutory damages. Before collecting a biometric identifier, a private entity must provide written notice of the purpose and retention period and obtain informed written consent, 740 ILCS 14/15(b); it must publish a retention-and-destruction schedule, 740 ILCS 14/15(a); and it may not sell or profit from biometric data, 740 ILCS 14/15(c). The Illinois Supreme Court removed the two escape hatches defendants most wanted. In Rosenbach v. Six Flags Entertainment Corp., 2019 IL 123186, it held that a plaintiff need not plead any actual injury beyond the statutory violation itself to sue. And in Cothron v. White Castle System, Inc., 2023 IL 128004, it held that a separate claim accrues with every unconsented scan—an accrual rule that turns routine, repeated biometric interactions into potentially staggering exposure and has driven nine-figure class settlements. The reach is not limited to Illinois residents' physical presence: in the litigation over face-template extraction from uploaded photos, illustrated by cases like Rivera v. Google Inc., 238 F. Supp. 3d 1088 (N.D. Ill. 2017), courts have allowed BIPA claims premised on algorithmic faceprinting to proceed.

Other states regulate biometrics with softer teeth. Texas's Capture or Use of Biometric Identifier Act (CUBI), Tex. Bus. & Com. Code Ann. § 503.001, and Washington's biometric statute, RCW §§ 19.375.010–.040, impose comparable consent and retention duties but channel enforcement through the attorney general rather than a private right of action. Meanwhile, the comprehensive state privacy laws—California's CCPA/CPRA, Cal. Civ. Code § 1798.100 et seq., along with the statutes in Colorado, Connecticut, Virginia, and Utah—fold biometric information into "sensitive" personal data with heightened obligations, and several now regulate automated decision-making and profiling directly. Where EU personal data is involved, the GDPR adds a lawful-basis requirement, treats biometric data used for identification as a special category under Article 9, and, in Article 22, constrains solely automated decisions that produce legal or similarly significant effects. Cross-border training or inference triggers the transfer machinery covered in Moving Personal Data Out of the EU After Schrems II: SCCs and Transfer Impact Assessments.

The operational takeaways are concrete. Inventory every category of personal and biometric data flowing into training and inference. Secure a lawful basis and any required consent before collection, not after. Minimize what you collect and set enforceable retention limits—the same data-minimization discipline covered in Data Minimization and Storage Limitation: Curbing the Over-Retention of Personal Data. And document all of it, because in a BIPA class action the consent records are the ballgame. The full biometric picture for AI builders—statute by statute, with EU AI Act overlays—is developed in Biometric Privacy Law and the AI Systems That Read Faces and Voices.

Worked example. A retailer deploys a shrink-reduction system that scans the faces of everyone entering its Illinois stores against a watchlist. Even if the vendor built and hosts the model, the retailer is the entity "collecting" biometric identifiers from Illinois residents. Without prior written notice and consent—difficult to obtain from walk-in shoppers—and a published retention schedule, each scan is a discrete BIPA violation under Cothron, and under Rosenbach no shopper needs to show a dime of actual harm to sue. Multiply the per-violation statutory floor by daily foot traffic across a year, and the litigation exposure can dwarf the value of the loss-prevention the system was meant to deliver. The right move is upstream: a data-protection impact assessment and a consent architecture built before the cameras turn on, as part of the program described in Standing Up a Privacy Compliance Program: A Practical Checklist.

Stage 6 — When AI causes harm: negligence, defect, and the duty to warn

Ownership and privacy dominate the headlines, but the exposure that keeps some general counsel up at night is simpler and older: what happens when the machine gets it wrong and someone is hurt or loses money? Here AI slots into the mature framework of tort and products liability, with a few genuinely hard adaptation questions.

Traditional theories apply first. A developer or deployer may face negligence claims for failing to design, test, monitor, or warn about a foreseeable risk, and strict products liability where an AI-enabled product is sold in a defective condition "unreasonably dangerous" to the user, the standard drawn from Section 402A of the Restatement (Second) of Torts and adopted in most states. Product-defect law recognizes design defects, manufacturing defects, and failure-to-warn claims, and a single AI flaw can support several theories at once. Practical Law's survey of AI tort liability walks through the classic fact pattern that most resembles an AI case: plaintiffs who sued the makers of GPS navigation devices after a driver, following the device's directions, drove a bus into a low overpass, alleging the product was defectively designed because it failed to account for vehicle height and failed to warn. Swap the navigation device for a recommendation engine, a diagnostic model, or a driver-assistance system and the doctrinal questions are the same—only harder, because the model's behavior is probabilistic and its internal reasoning is opaque.

Several adaptation problems recur. Foreseeability and the duty of care are strained when a model produces outputs its own designers cannot fully predict; plaintiffs argue that unpredictability is itself a design defect, while defendants argue it breaks the causal chain. The learned-intermediary doctrine, which in some contexts shifts the duty to warn from a manufacturer to a professional who stands between the product and the end user, is being tested where a physician or lawyer relies on an AI tool—does a warning to the professional discharge the developer's duty, or is the end user's reliance too direct? Hallucination creates a distinctive failure mode: a model that confidently fabricates a citation, a dosage, or a fact can cause harm through information rather than a physical defect, pushing at the line between product liability and negligent misrepresentation. And algorithmic bias generates a parallel track of exposure under civil-rights and consumer-protection law, including the FTC's authority to police unfair or deceptive practices under 15 U.S.C. § 45 and the Fair Credit Reporting Act's constraints on automated adverse decisions.

Worked example. A clinic licenses a diagnostic-support model that flags likely conditions from imaging. The model, trained on a non-representative dataset, systematically under-flags a condition in a demographic subgroup, and a patient is injured by a missed diagnosis. The plaintiff can plausibly plead a design-defect theory (the training data made the product unreasonably dangerous for a foreseeable population), a failure-to-warn theory (the vendor never disclosed the dataset's demographic gaps), and a negligence theory against the clinic for over-relying on an unvalidated tool. The vendor will invoke the learned-intermediary doctrine, arguing the treating physician was the responsible decision-maker; the plaintiff will respond that the tool was marketed to substitute for, not merely inform, clinical judgment. None of this is science fiction—it is § 402A and failure-to-warn doctrine wearing new clothes, and it is why AI vendors increasingly wrap their products in the warnings, use-restrictions, and human-in-the-loop requirements that Stage 8's governance program is designed to enforce.

Stage 7 — Contracting for AI: allocating the risk that doctrine leaves open

Because so much of AI law is unsettled, the contract is where risk actually gets allocated—well or badly. Whether a client is building a model, buying access to one, or embedding a vendor's AI into its own product, a competent AI agreement answers a predictable set of questions. Most of them are familiar from technology-transaction practice; adapt an existing playbook rather than starting from a blank page, and lean on UpLaw's Software License Agreements: Key Terms and How to Negotiate Them and the provision-by-provision Reviewing a Software License Agreement: A Practical Checklist.

Ownership across all three layers. Spell out who owns the inputs, the fine-tuning data, the model improvements, and the outputs. Given Stage 2's authorship uncertainty, do not assume copyright will silently assign output to the customer—assign it expressly, and address derivative rights in any customer data used to fine-tune the model.

Data-use and training restrictions. The single most important clause for a customer feeding confidential data into a vendor's tool is a prohibition on the vendor training its models on the customer's inputs unless the customer affirmatively opts in. Without it, a company can leak trade secrets or personal data into a shared model with a single careless prompt.

IP infringement indemnity. This is the clause the training-data war has made central. The customer wants the vendor to indemnify against third-party claims that the model's training data or its outputs infringe copyright. Vendors resist, and Practical Law's analysis of AI licensing flags the exceptions they typically carve out—infringement arising from the customer's modifications, from combining the tool with other products, or from the customer's own inputs. Negotiate those carve-outs deliberately; a broad "customer inputs" exception can swallow the indemnity whole when the customer's ordinary prompts are the very thing that produced the infringing output.

Warranties and disclaimers. Address accuracy, and be realistic that vendors will disclaim it heavily given hallucination risk; address non-infringement; and, increasingly, address lawful sourcing of training data and bias/non-discrimination in outputs. A warranty that the training corpus was lawfully obtained is now worth negotiating hard, because it converts Stage 4's litigation risk into a contractual allocation.

Confidentiality and security. Prompts and proprietary datasets are trade secrets; protect them with confidentiality terms that survive termination, using the drafting approach in Non-Disclosure Agreements for Technology Deals: Drafting for Enforceability.

Liability caps and open-source model terms. Cap liability with an eye to the outsized, hard-to-predict harm AI outputs can cause, and carve out the indemnity and confidentiality breaches from the cap. Where the deal uses open-weights or open-source models, run the licenses through open-source diligence, because model and dataset licenses increasingly carry copyleft-style and use-restriction terms; UpLaw's Open Source Software Licenses, Compliance, and Risk: A Comprehensive Guide explains the traps. For iterative, sprint-based AI builds where scope is inherently fluid, structure the deal around the acceptance and IP-ownership approach in Contracting for Agile Software Development: Structuring Iterative Deals.

Worked example. A fintech licenses a vendor's LLM to summarize customer support transcripts. The signed order form is silent on training and contains a broad indemnity carve-out for "outputs derived from Customer Data." Six months later, a summary reproduces a chunk of a copyrighted knowledge-base article the fintech had pasted into a prompt, and the article's owner sues. The vendor points to the carve-out—the output was "derived from Customer Data"—and declines to indemnify. Worse, the fintech discovers the vendor's default settings retained and trained on its prompts, meaning confidential customer data may now sit inside a shared model. Two clauses, negotiated at signing, would have prevented both problems: an opt-out-by-default training restriction and a narrower carve-out limited to the content the customer supplied rather than anything "derived from" it. This is why AI contract review is less about boilerplate and more about mapping the three layers onto specific risk allocations, a discipline UpLaw treats across the Software and Open Source Licensing: An IP-to-Compliance Toolkit.

Stage 8 — Governance: building a program that makes defensible decisions

The final internal discipline is governance—standing up a program so the organization makes consistent, documented, defensible decisions about AI instead of ad hoc ones. A workable program is not a policy binder that gathers dust; it is a living operating system for AI risk, and it connects directly to the privacy program of Stage 5, the contracting playbook of Stage 7, and the incident-response function that protects trade secrets and personal data.

A mature AI governance program does the following work:

1. Inventory. Maintain a register of every place AI is used or built across the organization, including shadow use of third-party tools by individual teams. You cannot govern what you cannot see.
2. Risk classification. Tier each use by stakes and by the regime it triggers—a marketing-copy assistant is not a hiring tool is not a medical-diagnostic system. This tiering mirrors the EU AI Act's risk pyramid and lets you apply proportionate controls.
3. Acceptable-use policy. Set clear rules for permitted and prohibited uses, above all a bright-line prohibition on feeding confidential, privileged, or personal data into third-party tools that do not contractually forbid training on it.
4. Human oversight. Require meaningful human review for high-stakes outputs, and design the review so it is real rather than a rubber stamp—the "human-in-the-loop" that both tort law and emerging regulation expect.
5. Bias, explainability, and testing. Test high-impact models for disparate performance across protected groups, document the testing, and preserve enough explainability to defend the decision later.
6. Vendor and data diligence. Diligence the provenance of vendors' training data and the terms of their licenses, folding the Stage 7 contract review into procurement.
7. Incident response. Plan for the AI-specific failure modes—a leaked trade secret via a prompt, a discriminatory output, a hallucination that reaches a customer—and integrate them into the breach playbook covered in Protecting Trade Secrets During a Data Breach: Building IP Into Incident Response.

Governance is also where external regulation lands, and it is arriving fast. The NIST AI Risk Management Framework (AI 100-1) and its Generative AI Profile (NIST AI 600-1) supply the leading U.S. voluntary vocabulary for categorizing and managing AI risk, and they map cleanly onto the program above. The EU AI Act, Regulation (EU) 2024/1689, introduces risk-tiered obligations—prohibited practices, high-risk-system duties, and transparency requirements—with extraterritorial reach that captures many U.S. developers and deployers whose systems touch the EU market. Domestically, the states are moving: the Colorado Artificial Intelligence Act (SB 24-205) imposes duties on developers and deployers of high-risk systems to guard against algorithmic discrimination; New York City's Local Law 144 requires bias audits of automated employment-decision tools; and the Illinois Artificial Intelligence Video Interview Act, 820 ILCS 42, requires notice and consent before AI analyzes recorded job interviews. Regulators are also applying existing law—the FTC under Section 5, the EEOC to automated hiring, and financial regulators to algorithmic underwriting—without waiting for AI-specific statutes. A living governance program lets an organization absorb these changes incrementally instead of scrambling each time a new rule lands. For the privacy backbone the AI program plugs into, see Building a Privacy Compliance Program: Governance, Data Mapping, and the Legal Patchwork and the full Privacy and Data Protection Toolkit.

Worked example. A mid-sized insurer's underwriting team quietly adopts a third-party model to pre-score applications, and a compliance officer discovers it only after a declined applicant complains. The model was never inventoried, never risk-classified, never bias-tested, and the vendor contract permits training on the insurer's submitted data. Under the Colorado AI Act, this is a high-risk system deployed without the required diligence and consumer disclosures; under state insurance regulation, it may constitute unfair discrimination; and the training-permission clause may have leaked policyholder data into a shared model. Every one of those failures is a governance failure that the seven-step program would have caught at the inventory or procurement stage. The cost of retrofitting compliance after a complaint always exceeds the cost of the program that would have prevented it.

Stage 9 — Lessons from adjacent frontiers: drones and crypto

AI is not the first technology to sprint past its statute, and the two most instructive recent examples—drones and cryptocurrency—teach the same lesson about how the legal system actually absorbs disruption. It rarely waits for a bespoke statute. It applies existing doctrines by analogy first, tolerates the resulting friction, and clarifies later. Reading how that played out sharpens judgment about where AI law is heading.

Drones forced regulators to reconcile a cheap flying camera with airspace safety, privacy, trespass, and property law. The FAA claimed jurisdiction over all drones as "aircraft" under 49 U.S.C. § 40102, built a commercial-operations regime in the Part 107 rules, 14 C.F.R. §§ 107.1–107.205, and layered on a Remote ID identification requirement. But the FAA's safety authority left the privacy questions—can your neighbor's drone hover outside your second-floor window?—to a patchwork of state statutes, common-law intrusion and trespass claims, and Fourth Amendment doctrine borrowed from the aerial-surveillance cases California v. Ciraolo, 476 U.S. 207 (1986), and Florida v. Riley, 488 U.S. 445 (1989), now pressed hard by the digital-privacy reasoning of Carpenter v. United States, 138 S. Ct. 2206 (2018). The result is a layered federal-safety-plus-state-privacy regime that any AI lawyer worried about surveillance systems will find familiar. UpLaw maps it in How Drones Are Regulated in the United States: A Practical Overview.

Cryptocurrency forced securities, commodities, money-transmission, and tax law onto a technology none of them anticipated, and regulators stretched old categories to reach it. The SEC applied the Howey investment-contract test from SEC v. W.J. Howey Co., 328 U.S. 293 (1946), to token sales; the CFTC asserted commodity jurisdiction over Bitcoin and Ether; FinCEN applied money-transmission rules; and the IRS taxed crypto as property. The friction—overlapping agencies, inconsistent enforcement, and litigation over whether a given token is a security—looks a great deal like the multi-agency scramble now forming around AI. Abroad, the EU eventually did legislate comprehensively, with the Markets in Crypto-Assets Regulation (MiCA), Regulation (EU) 2023/1114—the same "analogize domestically, legislate comprehensively in Brussels" pattern the EU AI Act repeats. The comparative picture is in How Cryptocurrency Is Regulated Around the World: A Comparative Overview.

Two more AI-adjacent frontiers deserve a flag. First, platform liability for AI-generated content: whether Section 230 of the Communications Decency Act, 47 U.S.C. § 230(c)(1), immunizes a platform for content its own model generated is contested, because the statute protects platforms from liability for third-party content, and a model's output may be the platform's own—an open question analyzed in Section 230, the IP Carve-Out, and Platform Liability for User-Generated Infringement. Second, the right of publicity and the wave of new anti-deepfake and voice-cloning statutes give performers and ordinary people a claim against AI systems that clone their voice or likeness; the doctrine and the new legislation are covered in The Right of Publicity: Controlling Commercial Use of Name, Image, and Likeness. The through-line across all of these frontiers is the same: when a technology is new, judgment about which existing doctrine a court will reach for is worth more than confidence about a statute that may not exist yet.

The AI counsel's first-pass audit

When an AI matter lands on your desk, the following working checklist forces the issue-spotting the rest of this toolkit is built on. It is deliberately doctrine-agnostic; run it before you decide which specialist chapters to open.

1. Classify the layer. Is the problem about the model, the inputs, or the outputs? Name it before analyzing it.
2. Identify the actors. Is your client the developer, the deployer, or the end user? Ownership and liability shift with the role.
3. Trace the data provenance. Where did the training and prompt data come from, was it lawfully obtained, and does it contain personal or biometric information? (Stages 4 and 5.)
4. Test output ownership. How much human authorship shaped the specific output, and what do the tool's terms of service assign? (Stage 2.)
5. Screen inventorship and eligibility. If a patent is in play, did a human make a significant Pannu contribution, and is the subject matter Alice-eligible? (Stage 3.)
6. Run the privacy gate. BIPA and its siblings, comprehensive state laws, and GDPR obligations—consent, minimization, retention, automated-decision rules. (Stage 5.)
7. Map the harm exposure. What tort, products-liability, and consumer-protection theories does a bad output invite, and who bears them? (Stage 6.)
8. Read the contract for the three-layer allocation. Ownership, training restrictions, IP indemnity and its carve-outs, warranties on lawful sourcing, confidentiality, and the liability cap. (Stage 7.)
9. Check the governance posture. Is the use inventoried, risk-tiered, human-supervised, and bias-tested, and which regulations (EU AI Act, Colorado AI Act, NYC LL 144, sector rules) apply? (Stage 8.)
10. Flag the moving ground. Note explicitly which conclusions rest on unsettled law, and advise with the humility the subject requires.

A last word on posture. The temptation in a fast-moving field is to sound more certain than the law allows. Resist it. The best AI counsel distinguishes cleanly between what is settled (a machine cannot be a named inventor or an author; BIPA requires consent), what is contested (training-data fair use), and what is genuinely unknowable until courts and legislatures act (how much human contribution rescues an AI-assisted invention). Say which is which. Clients can plan around uncertainty they understand; they cannot plan around false confidence.

Related UpLaw Resources

Related Guides

• Key Legal Issues in Artificial Intelligence: A Comprehensive Guide for Business and Counsel
• Building a Privacy Compliance Program: Governance, Data Mapping, and the Legal Patchwork
• Data Minimization and Storage Limitation: Curbing the Over-Retention of Personal Data
• Anatomy of a Software License Agreement: A Comprehensive Guide
• What an IP Case Is Worth: A Guide to Intellectual Property Damages

Related Articles

• Key Legal Issues in Artificial Intelligence: An Overview
• The Evolution of Computer Law: From Software Copyright to Artificial Intelligence
• Who Owns AI-Generated Inventions? Machine Creativity and Patent Ownership
• AI and Inventorship: A Global Comparison of Machine Contributions to Patents
• Generative AI on Trial: Copyright Infringement Claims After the New York Times and Getty Suits
• Web Scraping and the Law After hiQ v. LinkedIn: Where the CFAA, Contract, and Copyright Collide
• Biometric Privacy Law and the AI Systems That Read Faces and Voices
• Moving Personal Data Out of the EU After Schrems II: SCCs and Transfer Impact Assessments
• Software License Agreements: Key Terms and How to Negotiate Them
• Non-Disclosure Agreements for Technology Deals: Drafting for Enforceability
• Protecting Software and Business-Method Inventions After Alice: A Section 101 Strategy Guide
• Open Source Software Licenses, Compliance, and Risk: A Comprehensive Guide
• Protecting Trade Secrets During a Data Breach: Building IP Into Incident Response
• How Drones Are Regulated in the United States: A Practical Overview
• How Cryptocurrency Is Regulated Around the World: A Comparative Overview
• Section 230, the IP Carve-Out, and Platform Liability for User-Generated Infringement
• The Right of Publicity: Controlling Commercial Use of Name, Image, and Likeness
• Contracting for Agile Software Development: Structuring Iterative Deals
• Building Trade Secret Protection from the Ground Up: A Comprehensive Guide

Related Checklists

• Standing Up a Privacy Compliance Program: A Practical Checklist
• Reviewing a Software License Agreement: A Practical Checklist

Related Toolkits

• The Privacy and Data Protection Toolkit: Building and Operating a Compliance Program
• Software and Open Source Licensing: An IP-to-Compliance Toolkit
• Prosecuting a Patent Application: A Roadmap and Research Guide
• The IP Transactions Toolkit: A Roadmap for Licensing, Assignments, and Diligence

---

This toolkit is general legal information from UpLaw (uplaw.us), not legal advice, and no attorney-client relationship arises from reading it. AI and emerging-technology law is changing rapidly and is frequently unsettled; verify the current state of statutes, cases, agency guidance, and regulations at the official sources before relying on anything here.