Chapter 1. Who Decides, by When, and What Reopens It

The texts answer the paper's three questions unevenly. They say exactly who decides, they set dates only for the decider's own procedures, and they write the conditions for looking again in three different vocabularies: the law's, the engineer's and the data scientist's. This chapter first says how the texts were chosen and read, then takes the questions in turn.

1.1Data and method

The paper reads eleven texts, chosen by one rule: every text saved for this study whose licence allows it to be quoted, that is, every work of the United States government (which is not subject to copyright) and every article published under the Creative Commons Attribution licence. The list, with the licence and the evidence for it, is kept beside the texts in a manifest. A twelfth item there, the European Medicines Agency's reflection paper on artificial intelligence in the lifecycle of medicines, is listed only as a bibliographic record because no open licence for its text was found, and the paper does not quote or use it.

Seven texts are government works. One is two sections of the 21st Century Cures Act of 2016: section 3011, which created a process to qualify drug development tools, and section 3022, on real world evidence. One is section 3209 of the Consolidated Appropriations Act, 2023, known as the FDA Modernization Act 2.0, with the short section 3507 that follows it. One is the bill for an FDA Modernization Act 3.0 as introduced in the Senate on February 3, 2025; the paper read only that text and says nothing about what became of it. Four come from the Food and Drug Administration: the Federal Register notice of its draft guidance, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products; the guidance's landing page; its topic page Artificial Intelligence for Drug Development; and its announcement FDA Announces Plan to Phase Out Animal Testing Requirement for Monoclonal Antibodies and Other Drugs. The draft guidance itself was not saved, so the paper describes it only as the notice and the landing page describe it.

Four texts are articles under CC BY 4.0, each read in full from PubMed Central: those of Viceconti and others, Bender and Cortes-Ciriano, Qureshi and others, and Zushin and others, introduced above. Other works are cited only from their abstracts, as the bibliography lists them, and nothing is said of them beyond what an abstract says.

Each text was read for three things: the person or body it names as deciding whether a computed result may serve as evidence; any date or period it sets; and any condition it writes down under which a decision should be looked at again. Every block quote is a passage of a saved text, word for word, and its source line names the file. Every number in the prose stands in a saved text and is listed, with its file and place, in a numbers file. Passages that name a particular model were not quoted. Statements about the future are given as their authors' statements.

The texts are not a sample of anything. They are the documents through which one country's legislature and regulator have spoken about the question in the last decade, with four scholarly voices beside them. What they show is how the question is framed in those documents, not how any application was decided.

1.2Who decides

The statutes give the decision to one office and keep it there. The Cures Act did not mention computer models; it created a general process by which a tool for developing drugs is qualified for a stated use. Its first step belongs to whoever wants the tool accepted, and every decision after it to the Secretary of Health and Human Services, whose department includes the FDA:

What follows is a series of gates. The Secretary accepts or declines the letter, then a qualification plan, then a full qualification package, and only then decides the matter itself:

Two phrases in that sentence carry the weight. The first is scientific merit: the test is the quality of the evidence, not the policy preference of the moment. The second is context of use, which the act defines as the circumstances under which a tool is to be used in drug development and regulatory review. A tool is never qualified in general. It is qualified for a stated use, and once qualified for it, any developer may rely on it there. The act's list of what counts as a drug development tool names biomarkers and clinical outcome assessments and then leaves a door open:

This study reads the door as the place where a computational model could enter the qualification process, and notes that the same sentence puts the key in the Secretary's hand: a method is a tool when the Secretary determines that it aids. The act also keeps the bar where it was. Its rule of construction says that nothing in the section shall be construed

The same words close the section on real world evidence. Whatever new kinds of evidence the act admits, it admits them under the old standard.

The FDA Modernization Act 2.0 is often read as the moment the law let models in, and in one sense it was. It struck the words that tied the investigational stage to animals and put a definition in their place:

The definition then lists what such a test may include. Computer modelling is third on the list and animal tests are fifth:

Read as a decision, the amendment widens a category and decides nothing within it. It does not say that a computed result replaces an animal study, or when; it says that both are nonclinical tests, and it says so with the permissive word may. Who weighs one against the other is left where the Cures Act put it. Zushin and others read the change more strongly:

The sentence is careful in its own way: what the alternatives can be used for is to seek, and what is sought must still be granted. The difference between the statute's wording and the article's is small but it is the whole of this paper's first question. The law opened a category; the article describes a request; neither describes an acceptance.

The bill for an FDA Modernization Act 3.0, introduced in 2025, after that amendment became law, would go a step further. It would require the agency to rewrite its own regulations so that they speak of nonclinical rather than animal tests:

The bill would change the words the agency's rules use. It would not change who applies them.

Inside the agency, the texts name further deciders. The draft guidance on artificial intelligence, as its notice describes it, is issued by the Center for Drug Evaluation and Research with other centres and offices of the agency, and it asks sponsors to plan, gather and document the information that establishes a model's credibility, with the stringency of that work tailored to the model's risk. Its notice defines credibility in terms that echo the Cures Act's context of use:

The topic page adds a body the statutes do not mention, a council the drug centre established in 2024:

The page describes the council as a body that coordinates and oversees the centre's AI activities; it does not say that the council decides whether a given model's output is evidence. Its listed duty to promote consistency in how AI is considered when evaluating safety, effectiveness and quality is the closest any text comes to a body that would see the same question asked by many requestors. The 2025 announcement adds a further layer, the agency's announced plan, and a partner for validation, an interagency committee on alternative methods. In none of these texts does the decision move outside the agency. Viceconti and others, writing for scientists rather than regulators, say the same from the other side:

The answer to the first question is therefore clear in every text: the regulator decides, for one use at a time, on evidence a requestor brings, against a standard of evidence that the Cures Act declares unaltered and that the other sections read do not touch. What the texts leave open is how the regulator weighs a model against an animal when both are nonclinical tests. None of them gives a rule for that.

One more distinction belongs here because it changes the object of the question. The draft guidance does not cover the use that gives in-silico drug discovery its name:

A model that proposes a molecule is, on this study's reading, asking nothing of a regulator. Only when a computed result is offered as information about safety, effectiveness or quality does it become evidence that someone must accept or refuse. The texts thus divide the field in two: discovery, which the draft guidance leaves out, and evidence, where a regulator's decision is needed. This study's question lives entirely in the second half.

1.3By when

The texts are full of dates, and almost all of them bind the agency's paperwork rather than any act of acceptance. The Cures Act set a sequence for its qualification process: a public meeting within two years, a draft taxonomy of biomarkers within two years, draft guidance within three, and a report within five. Its words on the guidance give both a deadline and a hand-off:

What the guidance must contain includes the timing of the reviews themselves. The statute does not set it; it asks the guidance to:

So the one clock that would bear on a requestor waiting for an answer is left for the agency to set for itself. The section on real world evidence follows the same pattern, with a framework within two years, draft guidance within five, and revised or final guidance within 18 months of the comment period's close. Every one of these is a deadline for writing down how decisions will be made.

The FDA Modernization Act 3.0 would add one more, shorter and sharper:

It would also let that rule take effect at once, without the agency having to show good cause for skipping the usual notice. Even here the deadline falls on a change of words in the regulations, not on the acceptance of any model.

The draft guidance carried its own date, for the public rather than the agency: comments by April 7, 2025, so that the agency could consider them before it begins work on the final version. The notice says what the agency drew on and how much of it there was:

The same sentence goes on to the agency's experience with over 300 submissions with AI and machine learning components across all phases of drug development. The topic page, describing the same draft guidance later, counts the experience differently, as over 500 submissions with AI components from 2016 to 2023. The two figures need not conflict, since each is a floor and the texts give no common period; but neither text says how many of those submissions offered a computed result in place of animal or clinical evidence, which is the decision this paper is about.

The 2025 announcement is the only text that speaks of acceptance happening now:

Encouraged is not accepted, and the rest of the announcement keeps its tenses in the future: the agency will work to update its guidelines, companies may receive streamlined review, a pilot programme is aimed at over the coming year, and policy changes are expected to roll out in phases. Its most concrete promise about models is a forecast, given here in the announcement's own words:

The answer to the second question is that the texts set dates for rules, guidance, frameworks, comment periods and pilots, and none for a decision to accept a computed result. The timing of that decision is either left to guidance the agency writes for itself or tied to the pace of a sponsor's own application.

1.4What would reopen the decision

If acceptance is decided one use at a time, the conditions for looking again matter as much as the conditions for saying yes. The texts write three kinds.

The first is legal. The Cures Act lets the Secretary undo a qualification and says on what ground:

The trigger is new information about the basis, and the remedy is tied to the same context of use for which the tool was qualified. The act adds a meeting on request for the requestor, and a duty to publish every rescission or modification. It does not say who must go looking for the new information, or how often. The draft guidance is reopenable in a weaker sense: its notice says that, once final, it will represent the agency's current thinking and will not bind the agency or the public. Current thinking is by its nature something that can change.

The second kind is the engineer's. Viceconti and others make the strength of the checking depend on what rests on the model, and they define that dependence in two parts:

This turns the paper's question into a sliding scale. A model that only adds weight to animal and clinical data carries less influence than one that replaces them, and the checking required rises with it. A decision to let a model stand in for an animal study raises its influence by definition, so, on this study's reading, it is also a decision to demand more of it. Their framework also names the conditions under which an accepted model should be doubted. One is the range of the evidence:

Another is the kind of model. For models built from data alone, they write, credibility has to be assessed using different approaches from those suited to mechanistic models:

This study reads that as meaning that the acceptance of such a model is never final but holds until the next test. The authors note an FDA proposal for software as a medical device that suggested such models should be continuously tested, to avoid issues such as context drift. This study reads the reopening condition for a learned model, in that proposal, as a schedule rather than an event.

The third kind is the data scientist's, and it is the hardest to observe. Bender and Cortes-Ciriano argue that a model can perform well on the data it was tested on and still fail the question that matters, because the proxy it was trained on does not map onto what happens in a living body:

Their condition for looking again is not new information about a single tool but a gap between past and future performance, which they say current data sets make hard to measure: performance has to be evaluated so that retrospective results translate into prospective ones. Qureshi and others write the same doubt as a question about what follows a computed design:

The three kinds of condition sit at different distances from the decider. The legal one is in the decider's own statute but waits for someone to bring the new information. The engineer's is built into the act of acceptance, since the rigour required is set by the model's influence, but it depends on the context of use having been stated correctly; Viceconti and others warn that a model can be credible for a stated use and still have little clinical applicability if that use was misunderstood. The data scientist's condition is the most fundamental and the least observable: it says that the evidence on which acceptance rests may not predict what it is taken to predict, and that, in the authors' words, data of the right types are not currently available.

The 2025 announcement contains the one reopening clause that applies to the agency's own plan rather than to a tool:

It promises that the plan will be revised by what a pilot finds. It does not say what finding would slow or reverse it.