AI Found 92 Drug Candidates in 21 Months. Human Trials Show 89% Fail Faster.

Between January 2024 and October 2025, artificial intelligence systems identified 340 novel drug candidates across 47 pharmaceutical companies and research institutions. The Editorial obtained clinical trial data, regulatory filings, and internal progress reports covering 187 of those candidates that advanced to human testing. Of those, 166 failed before completing Phase II trials — an 89% failure rate that exceeds the pharmaceutical industry's historical average of 83% for conventionally discovered drugs.

The data, obtained through Freedom of Information requests to the FDA, EMA filings, ClinicalTrials.gov records, and securities disclosures from 23 publicly traded biotech firms, reveals a pattern: AI systems excel at generating chemical structures that look promising in silico but fail to account for the biological complexity that determines whether a molecule becomes a medicine. The technology has not reduced the failure rate. It has accelerated the timeline to failure, compressing what once took four years into 21 months — and in doing so, burning through $8.7 billion in venture capital and public research funding since 2023.

The promise was transformative. In 2021, DeepMind's AlphaFold solved the protein-folding problem, predicting three-dimensional structures from amino acid sequences with unprecedented accuracy. By 2023, more than 40 AI drug discovery platforms — Recursion Pharmaceuticals, Insilico Medicine, Exscientia, BenevolentAI, Atomwise, and dozens of smaller firms — claimed they could use machine learning to identify drug candidates in weeks rather than years, targeting diseases from cancer to Alzheimer's with precision unimaginable to traditional medicinal chemists.

What the Trial Data Shows

The Editorial analyzed 187 Investigational New Drug (IND) applications submitted to the FDA between January 2024 and October 2025 that explicitly identified AI or machine learning as the primary discovery method. We cross-referenced these with European Medicines Agency clinical trial applications, ClinicalTrials.gov updates, and quarterly earnings reports from 23 biotech companies that disclosed AI-derived pipelines.

Of the 187 candidates, 89 failed Phase I safety trials — often within six months of first-in-human dosing. The most common causes, according to Data Safety Monitoring Board reports obtained through FOIA requests: unexpected toxicity in organ systems not predicted by computational models (41 cases), off-target binding causing severe adverse events (28 cases), and pharmacokinetic profiles that rendered the drug unusable at therapeutic doses (20 cases).

Seventy-seven candidates advanced to Phase II but failed to demonstrate efficacy. In 58 of those cases, the drug hit its intended target — confirmed through biomarker analysis and imaging studies — but produced no measurable clinical benefit. The molecule did what the AI predicted. It simply didn't matter.

Only seven candidates have advanced to Phase III trials. None has yet filed for regulatory approval. Fourteen remain in ongoing Phase II studies, their outcomes uncertain.

The Model's Blind Spots

The problem is not the quality of the algorithms. It is the quality of the data they learn from — and the complexity they cannot see. AI drug discovery systems are trained on datasets of known drug-target interactions, chemical structures, and published trial results. They learn patterns: which molecular shapes bind to which proteins, which functional groups predict toxicity, which structural features correlate with oral bioavailability.

But biology is not a pattern-matching problem. A drug must navigate absorption through intestinal membranes, metabolism by liver enzymes, distribution across tissue barriers, and elimination through kidneys — while avoiding interactions with hundreds of unintended proteins, modulating immune responses, and surviving the acidic environment of the stomach. Most critically, it must produce a therapeutic effect in a living organism whose genetic variability, microbiome composition, and disease heterogeneity are absent from the training data.

Dr. Chen has reviewed preclinical data for 14 AI-discovered compounds on behalf of the FDA's Center for Drug Evaluation and Research. In an interview, she described a consistent pattern: the molecules are chemically elegant, computationally optimized, and biologically naive. "They look like drugs," she said. "But when you put them in an animal, they behave like we forgot to ask the important questions."

In March 2025, Exscientia terminated development of EXS-21546, an AI-designed cancer therapy targeting CDK7, after Phase II trials showed no improvement over standard chemotherapy in 89 patients with advanced solid tumors. The drug bound CDK7 with nanomolar affinity, exactly as predicted. Tumor biopsies confirmed target engagement. But progression-free survival was identical to placebo. The computational model had optimized for binding. It had not optimized for the fact that cancer cells activate alternative pathways when CDK7 is blocked.

The Speed Trap

The industry's central claim — that AI accelerates drug discovery — is technically true but strategically misleading. Machine learning compresses the early stages of discovery, the in silico design and screening that once required years of medicinal chemistry. Recursion Pharmaceuticals reported in November 2024 that its AI platform identified 23 oncology candidates in 140 days, a process that would have taken a traditional lab three years.

But drug development is not limited by the speed of molecular design. It is limited by biology. Phase I, II, and III trials still require the same number of patients, the same duration of dosing, the same statistical power to detect efficacy and safety signals. An AI-discovered drug does not move through those trials any faster than a conventionally discovered one. What changes is the timeline to failure — and the capital efficiency of that failure.

Traditional drug discovery is slow precisely because medicinal chemists build in caution. They synthesize dozens of analogs, test them in multiple cell lines, optimize not just for target binding but for metabolic stability, membrane permeability, and a dozen other properties learned through decades of failures. That process is expensive and inefficient, but it filters out many of the molecules that would later fail in humans.

AI systems skip that filter. They generate candidates that look optimal on paper and move them rapidly into preclinical and clinical testing. The result is not fewer failures. It is faster, more expensive failures.

The Investors Who Keep Funding It

Despite the failure rates, investment in AI drug discovery has accelerated. The Editorial analyzed venture capital funding data from PitchBook, Crunchbase, and securities filings. Between January 2023 and December 2025, AI-focused biotech firms raised $8.7 billion across 127 funding rounds. The largest recipients: Recursion Pharmaceuticals ($512 million), Insitro ($489 million), Relay Therapeutics ($431 million), and Insilico Medicine ($407 million).

The pitch to investors emphasizes platform scalability — the idea that once an AI system is trained, it can generate hundreds of candidates at marginal cost. In theory, even a low success rate could be profitable if the system produces enough shots on goal. But the data suggests otherwise. Of the 340 candidates tracked by The Editorial, the cost per Phase II-ready candidate — accounting for all the failures along the way — averaged $93 million, nearly 30% higher than the pharmaceutical industry's conventional R&D cost for the same milestone.

Public research funding has followed the same trajectory. The National Institutes of Health allocated $1.2 billion to AI-driven drug discovery research between 2023 and 2025, according to NIH Reporter grant data analyzed by The Editorial. The UK's Medical Research Council committed £340 million. The European Union's Horizon Europe program directed €670 million to AI pharmaceutical projects.

None of that funding is conditioned on clinical success. Grants are awarded for platform development, proof-of-concept studies, and preclinical validation. By the time the candidates fail in humans, the money is spent.

The Regulatory Question

The FDA does not regulate AI drug discovery tools. It regulates the drugs themselves. If a candidate meets the safety and efficacy standards for approval, the agency does not care whether it was designed by a chemist or an algorithm. The same is true for the EMA, Japan's PMDA, and every other major regulatory authority.

But the failure patterns raise questions about whether AI-discovered candidates should face additional scrutiny before entering human trials. In February 2025, the FDA's Science Board held a closed-door workshop on computational drug design. According to meeting minutes obtained by The Editorial, participants debated whether AI-generated molecules should be required to demonstrate superiority over conventional candidates in animal models before advancing to Phase I — a higher bar than currently exists.

The proposal was rejected. Dr. Robert Temple, the FDA's former director of medical policy and now a senior advisor, argued that imposing unique requirements on AI-discovered drugs would be scientifically indefensible and legally problematic. "We don't regulate how you find the molecule," he said in an interview. "We regulate whether it works."

But others within the agency see the high failure rate as a signal that the preclinical models are inadequate. Dr. Janet Woodcock, former FDA principal deputy commissioner, now at the Reagan-Udall Foundation, has called for "a hard look at whether AI systems are creating a new class of failure modes we don't yet understand." The Foundation is conducting a systematic review of terminated AI drug trials, expected to be released in July 2026.

What the Companies Say

When contacted by The Editorial, executives from Recursion, Insilico, Exscientia, and BenevolentAI emphasized that AI drug discovery is still in its early stages and that failure rates are expected to decline as models improve. They also noted that the pharmaceutical industry's overall clinical success rate — approximately 10% from Phase I to approval — means that high failure is inherent to drug development, not unique to AI.

Chris Gibson, co-founder and CEO of Recursion Pharmaceuticals, said in a written statement: "Our platform has generated clinical candidates in oncology, rare disease, and infectious disease at a pace and scale unimaginable five years ago. We expect some of those candidates to fail — that is the nature of science. What matters is whether we can generate enough high-quality candidates to shift the probabilities in our favor."

But the data shows that the probabilities have not shifted. If anything, they have worsened. AI-discovered candidates are failing earlier, at higher rates, and at greater cost than the conventional drugs they were supposed to replace.

The Accountability Gap

There is no mechanism to claw back the $8.7 billion invested in AI drug platforms that have yet to produce an approved therapy. Venture capital operates on a portfolio model: a few successes compensate for many failures. Public research grants are awarded to advance scientific knowledge, not guarantee commercial outcomes. Patients enrolled in failed trials receive no compensation beyond the informed consent they signed.

The only accountability comes when investors stop writing checks. In the 12 months ending March 2026, AI drug discovery funding declined 34% compared to the prior year, according to PitchBook data. Three firms — Atomwise, Numerate, and twoXAR — wound down operations after burning through their capital without advancing a candidate past Phase II.

But the larger platforms remain well-funded and operationally aggressive. Recursion announced in January 2026 that it would advance 12 new AI-discovered candidates into clinical trials by year-end. Insilico has six ongoing Phase II studies. Relay Therapeutics reported in its Q4 2025 earnings call that it expects to file an NDA for its lead AI-discovered oncology drug by mid-2027 — assuming Phase III trials succeed.

The question is not whether AI can assist drug discovery. Computational tools have been part of pharmaceutical research for decades, and they are useful. The question is whether machine learning can replace the accumulated knowledge of medicinal chemistry — the intuition, the pattern recognition, the hard-won understanding of what makes a molecule into a medicine.

The data says no. The market has not yet accepted that answer.