This summer, a drug whose biological target was nominated by one AI system and whose chemical structure was drafted by another entered Phase III. Insilico Medicine's rentosertib, an oral TNIK inhibitor for idiopathic pulmonary fibrosis, is being tested in a trial planned to enroll up to 320 patients at 47 sites in China over 52 weeks. If it completes Phase III, it would be the first drug originated largely by generative AI to complete that stage of clinical testing.
The verbs in that paragraph--nominated, drafted, originated--sound scientific, but they also raise a legal term of art: conception. Patent law asks who conceived the invention. Food and drug law asks whether a product remains the product that was cleared. Tort and insurance law ask who made the decision. Each question presumes a human answer. Bio-AI makes that presumption harder to maintain.
What the machines actually do
The legal questions track the system's architecture. Foundation models trained on biological data feed task-specific models that rank targets, generate molecules, or flag lesions on a scan. At the application end are systems that touch patients, from diagnostic software to the algorithms payers use to approve or deny care.
Two features set these systems apart. First, they are generative: they propose things --a target, a molecular structure, a draft care plan --rather than simply answer a question framed by a human. Insilico synthesized fewer than 80 machine-proposed molecules before nominating rentosertib. Second, they are adaptive: retrained on new data, a model may behave differently next quarter than it did when validated. Device law was built for artifacts that hold still. Adaptive systems complicate that frame.
The inventor problem
On inventorship, the rule is settled and blunt. Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), holds that only natural persons can be inventors. In November 2025, the U.S. Patent and Trademark Office (USPTO) rescinded its 2024 guidance on AI-assisted inventions and replaced it with a simpler rule: AI is a tool, like laboratory equipment or software, and ordinary conception doctrine applies, with no special test for AI.
Simple is not the same as easy. Conception is the formation in the inventor's mind of a definite and permanent idea of the complete and operative invention. Hybritech Inc. v. Monoclonal Antibodies, Inc., 802 F.2d 1367 (Fed. Cir. 1986). For a chemical compound, the inventor must be able to define it by structure and know how to make it. Amgen Inc. v. Chugai Pharmaceutical Co., 927 F.2d 1200 (Fed. Cir. 1991).
Consider the medicinal chemist who reviews a few hundred machine-generated structures and picks one. Did she form a definite idea of the molecule, or recognize one the machine had already formed? The tool analogy works when a human directs the search. It is less clear when the tool supplies the structure and the human supplies only the choice. Scientifically, the AI may have originated the molecule. Legally, only a human can conceive it. If the chemist's choice does not amount to conception, the AI cannot be named as the inventor instead. The invention may be left without a legally recognized inventor.
That question may surface years later in a deposition, when an accused infringer asks the scientist what the software contributed. Section 256 can correct a patent that names the wrong inventor. It cannot supply a human inventor if none contributed to conception.
The platforms face a different hurdle. Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025), cert. denied, No. 25-505 (U.S. Dec. 8, 2025), held that claims applying established machine learning to a new data environment, without claiming improvements to the model itself, are ineligible under section 101. Five months later, in Ex parte Desjardins, now precedential, the USPTO's Appeals Review Panel held that a genuine improvement in how a model trains can be patent-eligible. The decisions point to a practical distinction for patent portfolios: a claim to a molecule or a genuine advance in the model may stand on a different footing from a claim that applies established machine learning in a new setting. Other aspects of the model may instead be maintained as trade secrets.
The regulator's frozen frame
The Food and Drug Administration's (FDA) January 2025 draft guidance on AI in drug development expressly excludes drug discovery. For AI used in discovery, patent law is one of the more immediate legal issues. When AI is itself part of a medical product, FDA's device framework also comes into play.
A June 2026 Congressional Research Service report notes that the device framework evaluates a product at a point in time. The report cites about 1,450 AI-enabled devices authorized on that basis, most through the 510(k) pathway. Congress's response, in the Food and Drug Omnibus Reform Act of 2022, was the predetermined change control plan, or PCCP. Under a PCCP, the manufacturer specifies in advance how the model may change and how each change will be validated. FDA reviews the plan rather than requiring a separate review for each update covered by it.
A PCCP is, in effect, a contract with the regulator about the future behavior of a learning system. Its boundaries may shape the boundaries of the product's lawful operation, making the plan a legal exercise as well as an engineering exercise.
FDA has also adjusted the boundary for clinical decision support. Its January 2026 revision to the guidance extends enforcement discretion to software that offers a single clinically appropriate recommendation, once a near-automatic trigger for device status. The statutory condition remains: to stay outside device regulation, the software must allow the clinician to independently review the basis for its recommendation.
That is not a demand for public disclosure of source code. Information a company submits to FDA stays confidential. The harder question is whether a developer can explain the basis for a model's output, rather than simply keep the model proprietary. FDA does not appear to have authorized a generative AI device as of the date of this article, although several generative tools have received breakthrough designation. That designation is distinct from device authorization.
Who decided?
The third gate is accountability. In many drug and device liability settings, the learned intermediary doctrine places the physician between the manufacturer and the patient: the manufacturer warns the physician, and the physician exercises judgment for the patient. The doctrine depends on the physician's role as a decisionmaker, not simply as a conduit for a recommendation.
When software offers one answer and the clinician has 90 seconds, it may be harder to assess how much independent judgment the clinician exercised. FDA's revised guidance itself recognizes that time-critical software may not allow independent review. That raises questions about whether a physician who accepts an opaque output has acted as a learned intermediary at all, and how responsibility may be allocated among the physician, the developer and the manufacturer.
Europe is moving toward a different allocation. Its revised Product Liability Directive, which member states must implement by Dec. 9, 2026, explicitly classifies software, including AI, as a product subject to strict liability. It also allows courts, under specified conditions, to presume defect or causation when technical complexity makes proof excessively difficult.
The payer side previews the dispute. In Estate of Lokken v. UnitedHealth Group (D. Minn.), plaintiffs allege that UnitedHealthcare used an algorithm, nH Predict, in place of physicians to end post-acute care for Medicare Advantage members. UnitedHealth denies that the tool makes coverage decisions. In February 2025, the court allowed contract claims to proceed and dismissed the other claims as preempted by the Medicare Act. In March 2026, a magistrate judge ordered discovery reaching back to 2017 but denied production of the source code as irrelevant to the surviving claims.
The case illustrates how the claims that remain can shape access to an algorithm's internal workings. The court allowed discovery into how the algorithm was used but did not order production of its source code.
California moved early. SB 1120, the Physicians Make Decisions Act, effective Jan. 1, 2025, restricts health plans and insurers from allowing AI to supplant clinical judgment, requires a licensed clinician to make medical-necessity denials, and permits regulatory audits of the tools. At least five other states, including Arizona, Maryland and Nebraska, enacted human-review requirements in 2025.
A patchwork, and a principle
The federal government is challenging some state-level approaches. A December 2025 executive order created an AI Litigation Task Force, which began work in January 2026 to challenge state AI laws. But an executive order, by itself, does not establish that a state statute is preempted. Separate federal preemption rules may limit how state laws apply to particular plans. Laws like SB 1120 may not reach Medicare Advantage plans, where federal standards can preempt state law, or self-funded employer plans governed by the Employee Retirement Income Security Act (ERISA). Federal preemption was the basis for dismissing most of the claims in Lokken.
The European Union is on its own track. For device companies, its framework adds a separate compliance layer. At this year's GT-hosted European Life Sciences Summit in Berlin, participants repeatedly returned to the collision between the AI Act and the Medical Device and In Vitro Diagnostic Regulations. An AI system used in a medical device that requires notified-body review under the MDR or IVDR is classified as high-risk under the AI Act. One product may therefore be subject to both regimes.
The Digital Omnibus amendments, in force since July 27, 2026, postponed those obligations until Aug. 2, 2028, and took industrial machinery out of the AI Act's direct reach. Medical devices remained covered. Delayed, not dropped.
A more interesting point of convergence is change control. Like the FDA, the European Union recognizes in its AI Act that a learning system may change: modifications the provider specifies in advance at the initial conformity assessment are not treated as substantial modifications requiring a new assessment. Regulators on both sides of the Atlantic are beginning to define a product by its permitted future, not just its present state. How a notified body will treat a planned model update under the MDR remains unsettled.
What to consider doing now
Three practical questions follow. First, what records show the human contribution to conception: who framed the problem, constrained the search and chose the candidate--and why?
Second, how will the company approach patent protection and trade-secret protection? FDA, state regulators and courts may seek an account of how a model reaches its outputs without requiring public disclosure of the model itself.
Third, how do collaboration and license agreements address inventorship risk, authority over model changes and liability for algorithmic error? An agreement that is silent on those points leaves the allocation unresolved.
The law has never needed machines to be persons. It needs persons to be answerable. The inventor's oath, the device clearance and the physician's signature are all ways of naming the human who will answer for what happens next. Bio-AI does not end that need. It makes the name harder to find.