Biotech policy writer Ruxandra Teslo says AI has not yet delivered a world-changing new drug because the core bottleneck is not intelligence. It is the system built around drug development. She argues that a new drug takes an average of 7 years to move from discovery to market, costs more than $1 billion, and spends much of that time inside regulatory process rather than science.
In a recent article, Teslo framed the issue through two lenses. One comes from the San Francisco tech scene, where she says belief in AI as a universal answer can become close to doctrine. The other comes from medicine, where stronger tools have not removed the industry’s hardest constraints.
More spending, fewer approved drugs
Teslo described one encounter with an employee at an AI lab after she said she planned to keep writing about clinical trial policy. She said he looked at her as if she were a doomed hamster, convinced that AGI would soon out-argue professional advocates and make the remaining problems irrelevant.
Her evidence goes the other way. She points to Eroom’s law, the idea that over the past several decades, the number of new drugs approved per $1 billion in R&D spending has kept falling. Scientific tools have improved, but efficiency has not risen with them. That cuts against the intuition that better tools should automatically speed progress.
Scientific success does not guarantee scalable deployment
Teslo cites Adaptimmune as one example. The biotech company won U.S. Food and Drug Administration, or FDA, accelerated approvals for two cell therapies and treated patients with rare cancers. Even so, development costs remained so high that the company struggled on and ended up facing Nasdaq delisting pressure.
She makes a similar point with the case of “baby KJ.” A medical team used a customized gene-editing therapy to save an infant with a rare metabolic disease. The science, she argues, was sound. The problem was that regulation-driven manufacturing costs made the approach difficult to repeat for the next child.
AI may help find better endpoints, but validation still moves slowly
One area Teslo sees as especially promising is the use of AI to identify better surrogate endpoints and biomarkers. Instead of waiting years to learn whether a drug works, researchers could use earlier and continuously measurable signals of treatment effect. She estimates that the right surrogate endpoint could make clinical trials as much as 10 times faster while cutting cost at the same time.
But that effort, in her telling, is blocked by governance and data access rather than model performance. Companies trying to build such biomarkers often hit the same wall: they cannot get the data they need. She says one company has already spent a year waiting for the National Institutes of Health, or NIH, to release an imaging dataset. If the endpoint itself also needs FDA validation, the timeline stretches much further.
Her example is bone mineral density, or BMD, in osteoporosis trials. Teslo says the qualification process for BMD as a surrogate endpoint took 12 years, even though the supporting data were already in place and the analysis, in her words, was essentially a regression model.
She also challenges a common assumption about regulation. Many people treat it as a hurdle that matters only at the final approval stage, believing that if a drug works well enough, regulation ceases to be the real issue. Teslo argues that this misses where the delay actually happens. The bigger drag comes earlier, in whether human data can be collected at all and whether those data can actually be used once collected.
Patents protect molecules, not biological insight
Teslo identifies another structural constraint in biotech: target crowding. Drugmakers tend to cluster around the same set of already validated biological targets because those targets carry the lowest risk. As a result, the industry explores a much narrower biological space than its capabilities would otherwise allow.
She breaks drug development risk into two parts. One is target risk: whether a protein is truly linked to a disease and whether hitting it will be safe and effective. The other is molecular design risk: once the target is known, how to build a compound that is more precise and safer. The second problem is a chemistry problem and is more manageable. Given a target, a strong team can usually produce a workable molecule.
The patent system, however, mainly protects that second piece. Composition-of-matter patents protect a specific molecule, not the scientific insight that a target is worth pursuing. That means the first team to validate a novel target takes the largest risk but cannot secure the same level of reward. Once a first-in-class drug posts strong clinical data, the target has effectively been validated in public. Competitors can then design different molecules and still obtain strong patent protection.
The result, Teslo argues, is wave after wave of drugs built around the same biology, with weaker incentives to look for the next target.
Capital keeps backing molecular design
She says this distorted incentive structure also explains why funding keeps flowing to the same kind of AI biotech companies.
Teslo points to the two largest AI biotech fundraising rounds this year. Chai Discovery closed a $400 million Series C in July at a $3.8 billion valuation, nearly tripling in 7 months. Its models are already being used in drug development by Eli Lilly, Pfizer, and Novartis.
Isomorphic Labs, part of Google DeepMind, closed a $2.1 billion Series B in May, with Novartis also among its partners. In Teslo’s framing, both companies are focused on the same core task: making molecule design faster and more precise once a target has already been selected. That is the lower-risk half of the problem, and the half with the strongest patent protection. It is not the part where the industry is most stuck.
The bottleneck is institutional, not computational
Teslo’s conclusion is direct. Biotech does not simply need another few generations of better AI. It needs someone willing to bear the cost of reforming 7-year clinical trials, 12-year endpoint validation, and a patent regime that protects molecules but not targets.
Without changes there, she argues, advances in AI alone are unlikely to transform the pace and output of new drug development.