Artificial intelligence has moved drug discovery from hype to measurable clinical reality, with 117 AI-enabled therapeutic candidates now in human trials across 63 companies as of July 2026. Yet the field is being humbled by the same hard truths that have always governed pharmaceutical development: most drugs fail, even the ones designed by machine learning. Only 51 percent of these AI-discovered assets have completed Phase 1 testing, and just 6.8 percent have advanced to Phase 2, according to a peer-reviewed analysis presented at the American Society of Clinical Oncology (ASCO) in 2026.

The symbolic starting point for AI drug discovery is often marked as January 2020, when Exscientia and Sumitomo Pharma announced that a compound called DSP-1181 had entered clinical trials in just 12 months, compared to an industry average of 4.5 years for the exploratory research phase. That milestone made headlines as a potential game-changer. By 2026, however, DSP-1181 was quietly discontinued in 2022, and no AI-discovered drug has yet received full FDA approval.

Which AI-Designed Drugs Are Closest to Patients?

The most advanced candidate is rentosertib, developed by Insilico Medicine for idiopathic pulmonary fibrosis, a progressive lung disease. In Phase 2a trials published in Nature Medicine, rentosertib showed a mean improvement of 98.4 milliliters in forced vital capacity, a key measure of lung function, compared to a decline of 20.3 milliliters in the placebo group over 12 weeks. Insilico announced and registered a 320-patient Phase 3 study on July 7, 2026, with an estimated start date of August 30, 2026.

Just three weeks later, on July 29, 2026, Insilico received its first FDA Fast Track Designation for ISM6331, a pan-TEAD inhibitor for advanced mesothelioma, a rare and aggressive cancer. Fast Track status accelerates the FDA review process for drugs addressing serious conditions with unmet medical needs.

Generate:Biomedicines' antibody GB-0895 is also advancing rapidly. The company began Phase 3 trials for severe asthma on December 3, 2025, making it one of the earliest AI-designed drugs to reach late-stage testing. Recursion Pharmaceuticals, which absorbed the AI-native biotech Exscientia in a merger that closed in November 2024, is running a pipeline that includes REC-4881 for familial adenomatous polyposis, a genetic condition causing hundreds of polyps in the colon, which showed a 43 to 53 percent reduction in polyp burden in Phase 2 trials.

Isomorphic Labs, the Alphabet and DeepMind spinout built on AlphaFold protein-folding technology and backed by a record $2.1 billion Series B round in May 2026, has not yet disclosed a clinical candidate or dosed a patient as of mid-2026. CEO Demis Hassabis pushed his own guidance for clinical trials from an earlier end-of-2025 target to end of 2026.

Why Are AI-Discovered Drugs Failing at the Same Rate as Traditional Drugs?

The field has absorbed several high-profile setbacks that underscore the reality of drug development. BenevolentAI's BEN-2293 missed its efficacy targets in April 2023, triggering 180 layoffs and a $56 million cost reduction. Schrödinger halted SGR-2921 in August 2025 after two treatment-related deaths. These failures reveal a sobering truth: AI excels at finding promising molecules in the lab, but predicting how those molecules will behave in human bodies remains extraordinarily difficult.

A 2024 analysis by Boston Consulting Group found that AI-discovered molecules had an 80 to 90 percent success rate in Phase 1 trials, but only roughly 40 percent in Phase 2, comparable to historic industry averages. The 2025 Nature Medicine paper on rentosertib explicitly cautioned that "AI-discovered drugs have experienced similar levels of phase 2 trial failure as non-AI-discovered drugs" and that none had yet progressed through Phase 3 at the time of writing.

The core challenge is that machine learning models are trained on historical data about molecular properties and past drug trials. They can predict which compounds are likely to bind to a target protein or show activity in a test tube. But predicting how a drug will move through the human body, whether it will cause side effects, and whether it will actually help patients requires understanding biology at a level that current AI systems have not yet mastered.

How to Evaluate AI Drug Discovery Claims

As the field matures, distinguishing genuine progress from marketing narrative has become essential. Here are the key metrics to watch:

• Phase Advancement: Focus on drugs that have completed Phase 2 trials and entered Phase 3, not companies making broad claims about their platform technology or the number of molecules in their pipeline.
• Published Data: Look for results published in peer-reviewed journals, not just press releases. Peer review provides independent scrutiny of trial design and statistical analysis.
• Regulatory Milestones: Fast Track Designations, Breakthrough Therapy Designations, and other FDA recognitions indicate that regulators believe a drug addresses an unmet medical need, but they do not guarantee approval.
• Honest Failure Reporting: Companies that openly discuss discontinued programs and Phase 2 failures are more credible than those that only highlight successes.
• Funding and Runway: Check whether a company has sufficient capital to complete Phase 2 and Phase 3 trials, which can cost hundreds of millions of dollars and take years.

The global AI-in-drug-discovery market itself is estimated at roughly $2.3 to $2.6 billion in 2025, rising to $2.9 to $3.3 billion in 2026, with several research firms projecting a compound annual growth rate near 25 percent through the early 2030s. That growth reflects genuine investor confidence, but it also reflects the reality that AI is a tool, not a magic wand.

The traditional drug development process costs roughly $2 to $3 billion and takes 10 to 15 years to bring a single new drug to market, according to a 2025 Nature Medicine analysis. AI-native biotechs have promised to compress those timelines and reduce costs by automating target discovery, molecule design, and preclinical testing. Some progress is real: Exscientia's original claim of completing exploratory research in 12 months instead of 4.5 years was not false, even though DSP-1181 ultimately failed. But the promise of AI has always been about speed and cost in the early stages. The real test comes in Phase 2 and Phase 3, where the complexity of human biology and the heterogeneity of patient populations create challenges that no algorithm has yet fully solved.

By mid-2026, the field has moved past the era of pure speculation. Insilico Medicine, Recursion Pharmaceuticals, Generate:Biomedicines, and others are generating verifiable regulatory and clinical milestones. But the data also shows that AI-discovered drugs are not exempt from the hard realities of drug development. The next two to three years will be critical: if rentosertib and GB-0895 succeed in Phase 3 and reach patients, the narrative will shift from "AI can discover drugs" to "AI can discover drugs that actually work." If they fail, the field will need to confront deeper questions about what machine learning can and cannot predict about human biology.