FDA Shifts Drug Approval Policy: What Does This Mean for AI-Enabled Therapies?

The US Food and Drug Administration has formally shifted its evidentiary baseline for new drug applications, stating that one well-controlled clinical trial will now serve as the default requirement for approval. Agency leadership argues that advances in biomarker science, mechanistic understanding, and real-world data can supplement a single pivotal study.

See also: FDA Launches Internal AI Tool "Elsa" to Streamline Regulatory and Scientific Workflows

In an article published in The New England Journal of Medicine, FDA Commissioner Marty Makary and Center for Biologics Evaluation and Research Director Vinay Prasad described the longstanding post-thalidomide expectation of two adequate and well-controlled trials as a “dogma” with theoretical value in reducing false positives, but one that “no longer makes sense” as a default in an era of biomarker-rich, mechanism-driven development.

Thalidomide was a sedative marketed in the late 1950s/early 1960s, widely prescribed to pregnant women for morning sickness. It was later linked to severe birth defects—especially limb malformations—in thousands of babies in Europe and other countries, after in-utero exposure during early pregnancy. 

After the 1962 Kefauver–Harris Drug Amendments, passed in response to the thalidomide scandal, US law began requiring “substantial evidence” of efficacy from adequate and well-controlled studies. FDA interpreted that standard as usually meaning two independent pivotal trials to guard against flukes, and over time this replication norm hardened into the familiar “two-trial dogma.” 

Changing the default does not mean two trials will never be required; if a drug’s mechanism is unclear, its effects are measured mainly by short-term or surrogate markers, or the study has design limitations, the FDA may still ask for additional well-controlled trials and retains authority under US law to require the level of evidence it considers appropriate. The scope and frequency of these exceptions remain undefined.

Makary and Prasad argued that modern drug development increasingly integrates survival endpoints, validated biomarkers, and secondary outcomes to construct what they described as a more complete biological narrative. In that context, two trials are positioned as one component among multiple elements contributing to clinical credibility, rather than a fixed prerequisite.

The policy formalizes a regulatory posture that had already been applied in select contexts, particularly in rare diseases (the Rare Disease Evidence Principles, a framework allowing sponsors developing therapies for ultrarare, genetically defined diseases affecting generally fewer than 1,000 people in the US and lacking adequate treatments to seek approval based on a single-arm trial), but now establishes a single pivotal trial as the default starting point for most applications, supplemented by broader evidentiary support.

Under the updated framework, sponsors will generally be expected to submit one well-controlled, adequately conducted clinical trial as the primary basis for approval. However, applications must also include what the agency terms “confirmative evidence.” This may include mechanistic or biological data demonstrating target engagement or pathway modulation, evidence from related indications, supportive findings from animal models, class-wide data from similar drugs, or real-world evidence.

The shift has generated internal and external debate. Richard Pazdur, who led the Oncology Center of Excellence and was appointed director of the Center for Drug Evaluation and Research in November 2025, expressed reservations in an interview with The Wall Street Journal. He indicated that he had been asked to sign off on a press release supporting the single-trial policy while still undecided on the matter. Pazdur retired in December 2025.

This decision comes amidst a broader regulatory shift towards abandoning animal testing and encouraging New Approach Methods (NAMs)—alternatives such as organoids, organs-on-chips, in-silico modeling, AI/ML toxicity predictors, and computational simulations.

See also: From Animals to Algorithms: How AI Brings Drug Testing Closer to Human Biology

Additionally, FDA’s Real-World Evidence (RWE) program defines RWD and RWE and sets out a framework for using electronic health records, claims, registries and externally controlled trials in regulatory decisions. The confirmatory-evidence guidance explicitly lists RWE and externally controlled comparisons as acceptable supportive evidence when appropriately designed and analyzed. These are exactly the kinds of data modalities where AI/ML methods (causal inference, high-dimensional propensity scores, synthetic controls, etc.) tend to be necessary to extract credible signals.

What does this mean for AI-enabled therapies? 

AI-enabled programs may be better positioned to generate the new “confirmative evidence”

Under the single-trial default, FDA is explicitly asking for mechanistic, translational and contextual evidence around that one pivotal study. The confirmatory-evidence guidance lists categories where AI is already heavily used: mechanistic PD modeling, class-based inference, and natural-history and RWE analyses.

Mechanistically, AI-assisted computational models, imaging analysis and omics-based biomarker discovery can link a drug’s target engagement to pathway-level changes and to clinical endpoints. 

The RWE framework and the externally controlled trials guidance allow sponsors to use historical or real-world comparators when randomization is difficult or when single-arm trials are otherwise justified, provided they can handle confounding and bias. AI methods for causal estimation, high-dimensional matching and patient-level digital twins can make these external controls more statistically credible and may become a common form of confirmative evidence.

With only one traditional trial, FDA is likely to look at how well mechanistic models, nonclinical data, and RWE align with that trial’s findings, and a preclinical package built heavily on in-silico and organ-on-chip data may be easier to position with FDA, because the agency now explicitly anticipates non-traditional evidence.

With a one-trial default, the pivotal study becomes the main vehicle for establishing both effect and generalizability.

For AI-enabled programs this can also mean heavier use of AI in design and simulation:

• Digital twins, in-silico trial simulators, and model-informed development can be used to stress-test inclusion criteria, endpoints, and sample size before locking the pivotal protocol.
• Under a two-trial regime, you could “correct” with a second study; now the incentive is to use these tools to de-risk the first one.

See also: Companies Applying AI to De-Risk Clinical Trials: 2026 Watchlist

For AI-native therapeutics, the bar shifts from “how many trials?” to “how credible is the AI?”

In many AI-enabled therapies, the algorithm itself is part of the therapeutic mechanism (adaptive dosing engines, AI-selected treatment sequences, digital twin-guided dosing rules, etc.).

FDA’s 2025 AI regulatory-decision guidance focuses on a “risk-based credibility assessment framework” for AI outputs used to support drug decisions. Models must be validated for their context of use, including data quality, model design, performance, robustness and monitoring. EMA/FDA AI principles emphasize governance, transparency and risk management across the lifecycle rather than counting trials.

For AI-enabled therapies, the question may become whether the AI models meet the credibility framework (data provenance, drift monitoring, reproducibility), not just whether a single trial shows an effect. 

Additionally, under the new policy, pivotal AI endpoints and readouts get upgraded to “primary evidence”:

• If a primary or key secondary endpoint depends on an AI model (image classifier, sensor algorithm, composite digital biomarker), that model is part of the evidentiary core.
• With no second trial, the quality of that model’s validation across sites, scanners, devices, and populations carries more weight.

For now, the takeaway is that “AI-enabled” won’t be a differentiator by itself. Under a one-trial default, programs that treat models, digital endpoints, and NAM-heavy data packages as scrutinizable scientific objects will likely be better aligned with where FDA and EMA are already pointing. 

The open questions are about which architectures, governance setups, and evidence patterns regulators will consider reliable enough to sit alongside a single pivotal trial. Developers that start designing for that question explicitly, rather than treating AI as an internal black box, are the ones most likely to find this policy shift working in their favor.

Topic: AI in Bio