Ten Key Intellectual Property Considerations for Artificial Intelligence Companies — A Guide for Florida AI Developers, Deployers and Investors

An artificial intelligence company is an intellectual property company. Its training data is licensed, scraped or created; its models, weights and prompts are trade secrets or open; its architectures and training methods may be patentable; its outputs may or may not be protected, and may or may not infringe; its name competes in the most crowded trademark field in a generation; and its customer contracts allocate all of these rights.

Over the past three years, a body of statutes, agency guidance and court decisions has begun to answer the questions those facts raise—not completely, but enough that an AI company can now plan. This post sets out the ten intellectual property considerations we regard as most important for companies that build or deploy AI systems, with particular attention to the law as it applies in Florida and the Eleventh Circuit.

For our related practice areas, see Patent Matters, Copyright Matters, Trademark Matters, Trade Secret Matters and Technology & Start-Up Services.

1. Training Data: Copyright, Fair Use and the Cost of Provenance

The threshold intellectual property question for any company that trains a model is whether it may use the works in its training corpus. The answer, as of September 2026, depends on what the works are, how they were obtained, and what the model does with them. Three district court decisions in 2025 framed the debate. In Bartz v. Anthropic (N.D. Cal. June 23, 2025), Judge Alsup held that training a large language model on lawfully acquired books was “quintessentially transformative” and a fair use, but that downloading millions of pirated books to build a permanent library was not—a distinction that produced a $1.5 billion class settlement, approved in July 2026, and spawned individual opt-out suits. In Kadrey v. Meta (N.D. Cal. June 25, 2025), Judge Chhabria granted summary judgment to Meta on fair use on the record before him while warning that a plaintiff who proved market dilution—a flood of AI-generated substitutes—might well prevail. And in Thomson Reuters v. Ross Intelligence (D. Del. Feb. 11, 2025), Judge Bibas rejected fair use where a legal-research startup used Westlaw headnotes to train a competing, non-generative search tool. The Third Circuit heard argument on June 11, 2026, and its decision, expected later this year, will be the first appellate ruling on AI training. The New York Times multidistrict litigation against OpenAI and Microsoft, the music publishers’ suits against Anthropic, and the studios’ suits against Midjourney remain pending, and courts abroad have diverged: Munich’s Regional Court held in GEMA v. OpenAI (Nov. 2025) that memorization of lyrics in a model infringes under German law, while the English High Court in Getty Images v. Stability AI (Nov. 2025) largely rejected Getty’s claims on the ground that a trained model is not itself an infringing copy.

The Copyright Office’s Part 3 report on generative AI training (pre-publication version, May 9, 2025) concluded that training is neither categorically fair use nor categorically infringing: uses that produce non-substitutive tools with guardrails are strongest, uses that generate outputs competing with the training works in their own markets are weakest, and the manner of acquisition matters.

The practical consequences for an AI company are these:

• Document the provenance of every dataset;
• Never use pirated or circumvented sources, which convert a defensible fair-use position into willful infringement;
• License where a licensing market exists (the news, music, image, book and code licensing markets have grown rapidly since 2024, and a functioning market strengthens the plaintiff’s fourth-factor argument against unlicensed use);
• Implement output guardrails against regurgitation; and
• Maintain records sufficient to answer the EU AI Act’s training-data summary requirement and California’s training-data transparency law, both now in force. Our earlier post on Bartz addresses the decision in detail.

2. Data Acquisition Beyond Copyright: Scraping, Terms of Service, Computer-Access Laws and Trade Secrets

Copyright is not the only law that governs how training data is obtained. Website terms of service that prohibit automated extraction are enforceable contracts in most courts when assent is properly obtained. The Ninth Circuit’s hiQ v. LinkedIn litigation (2022) held that scraping public pages does not violate the Computer Fraud and Abuse Act, but the case ended with hiQ liable for breach of LinkedIn’s terms. The Eleventh Circuit’s decision in Compulife Software v. Newman, 959 F.3d 1288 (11th Cir. 2020), holds that scraping a compilation of publicly accessible data at a scale no human could match, in violation of the site’s intended use, can be misappropriation of a trade secret by improper means under Florida’s Uniform Trade Secrets Act and the Defend Trade Secrets Act—and its 2024 sequel imposed joint and several liability on everyone involved in the scheme, including the contractor who wrote the scraper. Florida’s Computer Abuse and Data Recovery Act, § 668.801 et seq., provides a civil claim against unauthorized access to protected computers, and privacy statutes—the Illinois Biometric Information Privacy Act, the California Consumer Privacy Act, the Florida Digital Bill of Rights, the EU General Data Protection Regulation—regulate personal data in training sets independently of intellectual property.

An AI company should therefore maintain a data-acquisition policy that reviews each source for terms of service, technical access controls, personal data and trade secret indicators, records the basis for use, honors robots.txt and the emerging machine-readable opt-out standards that the EU AI Act makes legally significant, and treats data obtained by a vendor as carrying the vendor’s risk—with contractual representations and indemnities to match. The same analysis, in reverse, protects the company’s own data: terms of service prohibiting extraction, rate limits, authentication and bot detection are what make a Compulife claim available against competitors who scrape the company’s products.

3. Ownership of Outputs: What AI-Generated Content Can Be Protected

Under Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. Mar. 18, 2025), cert. denied (Mar. 2, 2026), a work generated autonomously by an AI system has no author under the Copyright Act and cannot be registered. Under the Copyright Office’s Part 2 report (Jan. 29, 2025) and its registration guidance, works that combine human authorship with AI-generated material are registrable to the extent of the human contribution—the selection, arrangement, modification or creative control that a human exercised—but prompts alone, however detailed, do not make the prompter an author of the output. Allen v. Perlmutter (D. Colo.), challenging the refusal to register a prompt-generated image, remains pending. The same principle governs patents: under Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), an AI system cannot be an inventor, and the USPTO’s inventorship guidance, first issued in February 2024 and revised in November 2025, requires that a natural person have made a significant contribution to the conception of each claim, whether or not an AI tool assisted.

For an AI company the consequences run in two directions. Its own products—model outputs offered to customers—may not be protectable by copyright at all, which affects how the company describes what it delivers, what it warrants, and what its customers can enforce against copyists. And the company’s internal use of AI to write code, draft documentation, design interfaces and generate marketing content may leave those assets without copyright protection unless human contribution is documented. The response is procedural: record the human creative decisions in the workflow, register the human-authored components, rely on trade secret protection and contract for what copyright will not protect, and describe AI involvement accurately in registration applications, which the Copyright Office requires and which, if omitted, can invalidate the registration.

4. Output Liability: Infringing, Defamatory and Replicating Outputs

The complement to the ownership question is whether outputs expose the company to liability. Three theories predominate. The first is copyright infringement by output—a model that reproduces a training work, or produces a substantially similar derivative, infringes regardless of whether training was fair use, and the New York Times, music-publisher, GEMA and studio cases all rest in part on demonstrated regurgitation. The second is trademark and false endorsement—outputs that reproduce logos, characters or brand names, and models that generate content “in the style of” an identifiable brand, draw Lanham Act and dilution claims. The third is the digital replica—voices, likenesses and personas. The Copyright Office’s Part 1 report (July 31, 2024) recommended a federal digital-replica right; the NO FAKES Act, reintroduced in 2025 and reported by the Senate Judiciary Committee in June 2026, would create one; the TAKE IT DOWN Act (May 2025) already criminalizes non-consensual intimate deepfakes and requires platform takedown within 48 hours; and state law—Florida’s right of publicity statute, § 540.08, its 2024 criminal statute on altered sexual depictions, § 836.13, and Tennessee’s ELVIS Act—reaches unauthorized voice and likeness replication now. Lehrman v. Lovo (S.D.N.Y. 2025) allowed voice actors’ state-law publicity and contract claims against a voice-cloning company to proceed even as it dismissed the copyright claims.

Output liability is managed through engineering and contract together: content filters and provenance watermarking; refusal behaviors for named artists, brands and living persons, retrieval controls that keep licensed and unlicensed material separate, terms of use that prohibit infringing and replicating uses and allocate responsibility to the user for prompts, and, for enterprise customers, the copyright indemnities that the major platforms now offer—which are conditioned on the customer’s use of the platform’s guardrails and which an AI company should be prepared to match or explain. Our earlier post on digital replicas sets out nine recommendations for companies using generative models to create content.

5. Patents: Eligibility, Inventorship and the Choice Between Patents and Secrecy

AI inventions are patentable, and the environment for them improved markedly in late 2025. In Ex parte Desjardins (USPTO Appeals Review Panel, Sept. 26, 2025; designated precedential Nov. 4, 2025), the Office held that claims to a method of training a machine-learning model across sequential tasks—computing importance measures for parameters and adjusting them to preserve prior learning—were patent-eligible under § 101 because they improved how the model itself operated, and it instructed examiners in a December 2025 memorandum to credit technological improvements described in the specification rather than dismissing claim elements as generic computer components. The contrast with the Federal Circuit’s Rady v. Boston Consulting Group (Mar. 27, 2024), which invalidated a patent that simply recorded gemstone signatures on a blockchain, defines the line: claims that improve the model, the training process, the data structure, the inference architecture or the hardware are eligible, but claims that apply a conventional model to a business problem are not.

Inventorship remains a constraint. Because an AI system cannot be an inventor, each claim must reflect a significant contribution by a named natural person. Where AI tools were used in the inventive process, the company should document the human contribution to conception—the problem framed, the architecture chosen, the results recognized and refined—contemporaneously, and should disclose AI use where the USPTO’s duty-of-candor rules require it. Finally, the decision to patent must be weighed against secrecy. A patent discloses the invention and, for many AI advances, will be designed around or superseded within its term; a trade secret protects only what stays secret, which for a deployed model may be difficult. Most AI companies patent architectures, training methods and hardware-adjacent inventions that are observable or reverse-engineerable, and hold weights, training data, fine-tuning recipes, evaluation sets and prompt libraries as trade secrets.

6. Trade Secrets: Protecting Models, Weights, Data and Prompts

For most AI companies the most valuable intellectual property is a trade secret: the trained weights, the curated and labeled training data, the reinforcement-learning and fine-tuning datasets, the system prompts and orchestration logic, the evaluation benchmarks, and the know-how of the engineering team. All are protectable under the Defend Trade Secrets Act and the Florida Uniform Trade Secrets Act if they derive value from secrecy and the company takes reasonable measures to protect them—and Florida courts, as Yellowfin Yachts v. Barker Boatworks, 898 F.3d 1279 (11th Cir. 2018), shows, examine those measures closely.

An AI company should maintain a trade secret register that identifies each asset, control access to weights and data through role-based permissions, logging and secure enclaves, execute confidentiality and invention-assignment agreements with every employee and contractor, including the DTSA whistleblower notice that preserves exemplary damages and fees, restrict what may be entered into third-party AI tools, since prompts submitted to an external model may be retained and used for training under that provider’s terms, and address departures with return of information, revocation of access and written reminders.

Florida offers two advantages here. Because Florida does not recognize inevitable disclosure, the departure of a key researcher to a competitor cannot be enjoined on the basis of what she knows. However, Florida’s restrictive covenant statute, § 542.335, and the 2025 CHOICE Act permit non-competition and garden-leave agreements, supported by the company’s trade secrets, that are enforceable to a degree unavailable in California and constrained in Illinois and elsewhere. And the decision whether to release model weights openly is a trade secret decision that cannot be reversed: an open-weights release surrenders secrecy in the weights permanently and should be made only with a clear licensing strategy and after separating the assets—data, recipes, evaluation sets—that will remain proprietary. See Restrictive Covenants & the CHOICE Act.

7. Open Source and Model Licenses: Compliance in Both Directions

AI systems are built on open-source software, open datasets and, increasingly, open-weights models released under licenses that are not open source in the traditional sense. Each carries obligations. Copyleft licenses (GPL, AGPL) in inference servers, training frameworks or embedded components can require disclosure of the company’s own source code when the software is distributed or, under the AGPL, offered as a network service. Dataset licenses frequently restrict commercial use, require attribution or prohibit use for training. Model licenses—Meta’s Llama community license, the various Responsible AI Licenses, Creative ML OpenRAIL, and the bespoke licenses attached to many foundation models—impose acceptable-use restrictions, user thresholds, attribution requirements, and in some cases prohibitions on using outputs to train competing models. Non-compliance is copyright infringement and breach of contract, and it surfaces in every financing and acquisition diligence process.

An AI company should maintain a software bill of materials that extends to models and datasets—an “AI bill of materials”—recording the license of every component and the obligations it triggers, review each open model’s license before fine-tuning or deployment, avoid AGPL and similar components in proprietary services unless the company intends to comply;,and, when the company releases its own models, choose and draft a license that reflects its strategy for community use, commercial licensing and competitive restriction.

8. Trademarks: Naming and Branding in a Crowded Field

The rush to brand AI products has produced the most congested trademark landscape since the dot-com era. The USPTO treats “AI” and “.ai” as descriptive or generic when applied to AI products, so a mark’s distinctiveness must come from the rest of it. Hundreds of applications combine common words with “AI,” and the resulting refusals and oppositions are routine. OpenAI v. Open Artificial Intelligence, Inc. (N.D. Cal.; preliminary injunction affirmed by the Ninth Circuit) illustrates the stakes: the company obtained an injunction against a prior registrant of “Open AI” only after proving that the registrant’s use was not bona fide and that OpenAI’s mark had acquired distinctiveness. Model names, product names, agent names and the names of features are each potential marks, and each should be cleared—against the federal register, state registers including Florida’s, domain names including .ai, app stores and the model repositories where names propagate—before launch, and registered in Classes 9 (software), 42 (software-as-a-service and AI development services) and, where relevant, 35 and 41. Trademark law also governs outputs and marketing: models should not use third-party marks as source identifiers for their own outputs, comparative advertising against named competitors must be substantiated, and keyword advertising on competitors’ marks is subject to the Eleventh Circuit’s 2026 decision in Deltona Transformer v. NOCO, addressed in our earlier post. See Trademark Matters.

9. Contracts: Allocating Rights with Customers, Vendors, Data Licensors and Employees

Because so much of AI intellectual property is either unprotectable or uncertain, contract does the work that property law cannot. Four sets of agreements matter most.

Customer terms should state who owns inputs and outputs (most enterprise terms assign outputs to the customer, subject to the reality that they may not be copyrightable), whether and how customer data will be used to train or improve models (the answer is now a competitive term, and a misstatement is a FDUTPA and FTC Act problem), what the company warrants about outputs, what indemnity it offers for third-party intellectual property claims and on what conditions, and what uses are prohibited.

Vendor and API terms—the terms under which the company itself consumes foundation models, data and tools—should be reviewed for retention and training rights in the company’s prompts and data, for restrictions on building competing models, for the scope of the vendor’s indemnity, and for the vendor’s own license compliance.

Data licensing agreements, now a mature market for news, music, images, books, code and specialized corpora, should define the licensed uses (training, fine-tuning, retrieval, display), the term and the survival of trained models after termination, exclusivity, provenance warranties and audit rights.

Employee and contractor agreements should assign inventions and works made with AI assistance, address inventorship documentation, restrict the use of external AI tools on company information, and include the DTSA notice and the restrictive covenants that Florida law permits. Our earlier post on technology start-up services addresses these agreements in more detail.

10. The Regulatory Overlay: Consumer Protection, Transparency and Florida Enforcement

Intellectual property rights in AI are increasingly enforced through consumer-protection and transparency law rather than through infringement suits alone. The State of Florida’s June 1, 2026 action against OpenAI under the Florida Deceptive and Unfair Trade Practices Act alleges that marketing an AI system as safe and reliable while failing to disclose known limitations is a deceptive practice, and the Attorney General’s 2025–2026 actions against social media and streaming platforms rest on similar theories about data collection and product design. The EU AI Act’s obligations for general-purpose model providers, in force since August 2, 2025, require a copyright compliance policy that honors machine-readable opt-outs and a public summary of training content; California’s training-data transparency statute (AB 2013), effective January 1, 2026, requires developers to publish documentation of the datasets used to train models offered in California; Colorado’s AI Act, delayed to June 30, 2026, imposes duties on high-risk systems; and the TAKE IT DOWN Act’s platform-removal obligations took effect in May 2026.

Each of these regimes makes the provenance records, licensing documentation and guardrails described above a legal requirement rather than a best practice, and each is enforceable against a Florida company that offers its products in the relevant jurisdiction. See FDUTPA.

A Checklist for Florida AI Companies:

• Adopt a written data-acquisition and provenance policy, record the source, license or legal basis for every dataset, and exclude pirated and circumvented sources.
• License training content where a market exists, and document the fair-use analysis where it does not.
• Implement and document output guardrails against regurgitation, brand and character reproduction, and voice and likeness replication.
• Record human creative contribution in AI-assisted work, register the human-authored components, disclose AI use in registration applications.
• Document human conception in AI-assisted inventions, file on architectures, training methods and hardware-adjacent improvements under the Desjardins framework, and hold weights, data and recipes as trade secrets.
• Maintain a trade secret register, control access to weights and data, and execute confidentiality, invention-assignment and, where appropriate, § 542.335 and CHOICE Act agreements with the DTSA notice.
• Keep an AI bill of materials, comply with open-source, dataset and model licenses, and choose deliberately whether and how to release models.
• Clear and register product and model names in Classes 9 and 42, avoid marks that are merely “AI” plus a descriptive term, and monitor the field.
• Align customer, vendor, data-licensing and employment agreements so that ownership of inputs and outputs, training rights, indemnities and restrictions are consistent across the stack.
• Build the compliance records—training summaries, transparency documentation, safety substantiation—that FDUTPA, the EU AI Act, California, Colorado and the TAKE IT DOWN Act now require, and treat marketing claims about safety, accuracy and data use as regulated representations.

Conclusion

The law of intellectual property in artificial intelligence is being written case by case, but its outline is now visible. Training on lawfully acquired data with guardrails is defensible, but training on pirated data is not. Outputs belong to no one unless a human made them, and may infringe regardless. AI inventions are patentable when they improve the technology, and AI inventors do not exist. Weights and data are trade secrets only if guarded. Licenses bind in both directions. Names must be cleared. Contracts allocate what property law leaves open. And regulators—including Florida’s Attorney General—enforce accuracy and transparency about all of it. An AI company that builds its intellectual property program around these ten considerations will be positioned to raise capital, to license, to enforce and to defend.

If your company would like an intellectual property audit of its data, models, agreements and branding, or is facing a claim or a regulatory inquiry relating to an AI product, please visit our Technology & Start-Up Services, Patent Matters, Copyright Matters and Trade Secret Matters pages, or contact us at 305-374-8303.