Who owns the output, and was the data clean?

Why this page is framed around coordination

Honesty about scope matters, so this is stated plainly up front. Patent prosecution, trademark filing, and the deep IP-portfolio work that AI companies need are the province of dedicated intellectual property counsel — specialists, often with technical degrees, who do that work daily. This practice does not pretend otherwise. What it provides is the governance-and-strategy layer that sits on top of and alongside that IP work: identifying where the IP and training-data exposure actually lives in an AI system, framing the right questions, and quarterbacking a coordinated effort with IP counsel and, where needed, outside and out-of-state specialists. The value is in the integration — connecting the IP analysis to the company's governance, contracts, privacy posture, and risk — done by someone who understands the technology because he architects and governs a production AI system of his own.

Ownership of AI output

When an AI system generates text, code, images, or other content, the question of who owns it — and whether anyone can own it — is genuinely unsettled and continues to develop; a significant part of that uncertainty traces to the U.S. Copyright Office's position that purely AI-generated material without sufficient human authorship is not registrable, the contours of which are still being worked out and should be verified as they develop. For a company whose product produces AI output, or whose employees create work with AI assistance, this is not academic: it affects what the company can claim, license, enforce, and sell. The practical work is making sure the company's terms of service, customer agreements, and internal policies address output ownership clearly and defensibly given the current state of the law — and revisiting them as the law evolves, which it is doing quickly. Getting this wrong surfaces at the worst time: when rights need to be enforced against a copyist, or transferred in a financing or acquisition, and the company discovers its position was never clear.

Training-data rights and provenance

The data a model is trained on is one of its largest and least-scrutinized legal exposures, and it is the issue most likely to surface painfully in diligence or litigation. The questions are concrete: where did the training data come from — licensed, purchased, user-contributed, or scraped — and did the rights to use it for training actually transfer? Does it contain personal data that triggers privacy obligations, or material whose use is contested? Provenance is difficult to reconstruct after the fact, so the discipline of documenting what a model was trained on, and on what legal basis, is far cheaper built in early than untangled later. This is a fast-moving area of law — the fair-use status of training models on copyrighted works is the subject of active, unresolved federal litigation — so the analysis must be current, and the harder questions are coordinated with dedicated IP counsel.

Trade secrets, confidentiality, and the AI stack

Beyond output and training data, an AI company's competitive value often lives in things best protected as trade secrets — model weights, training pipelines, prompts, fine-tuning methods, and proprietary datasets. Protecting these requires a higher level of discipline than ordinary confidential information: access controls, contractual restrictions, and governance that actually keeps the secret secret. Because Adam architected and governs these components in his own production AI system, he understands what is actually sensitive and where it leaks — which informs how the protective measures are designed, in coordination with IP counsel who handle the formal patent, trademark, and copyright filings where applicable and who advise on trade-secret protection and enforcement strategy.

A coordinated, multi-jurisdictional approach

IP and training-data exposure does not respect state lines, and neither does a publicly deployed AI product. As a practical risk posture, the work is built to the most demanding standards across the jurisdictions an AI system reaches — and where the company has international exposure, that includes foreign IP and data regimes — recognizing that which jurisdiction's law actually governs a given question is fact-specific. The firm works in cooperation with dedicated intellectual property counsel and with outside and local counsel in other jurisdictions as the matter requires, bringing in the right specialists rather than overreaching. What the company gets is an integrated strategy — IP, training data, governance, contracts, and privacy treated as one connected problem — led by someone who understands the underlying technology and is candid about which parts belong to which counsel.

What usually goes wrong

The defining failure is the training-data gap: building a product on data whose rights were never cleared or documented, then being unable to prove provenance when a diligence team or a plaintiff asks — a problem that can stall a financing or anchor a lawsuit. A close second is the company that treats output ownership casually, with terms that never resolve what it owns, until the moment it needs to enforce or transfer those rights and discovers the position was never clear. The third is the company that tries to handle sophisticated AI IP with a generalist and no coordinated IP counsel — or, conversely, with IP counsel who handle the filings but never connect them to the company's governance, contracts, and data practices, leaving the strategy fragmented.

Related

• AI for Startups & Product Counsel →
• HB 4875 — Digital Replica Rights →
• AI Vendor & Procurement Contract Review →