Does AI change the odds?
The question this registry exists to answer, and what the data can say so far.
The question
Most drugs that enter human trials never reach patients. Across the industry, about 1 in 13 candidates that start Phase 1 are eventually approved, and Phase 2 is the steepest step: fewer than 3 in 10 get through it.1
Developers that use AI say they choose better targets and design better molecules. If that is true, it should show up in the clinic as better odds at each phase, faster progress, or both. Testing the claim needs a denominator. It cannot be every molecule AI has touched: AI input is now routine across drug discovery, and much of it is never disclosed. What can be built is a cohort: the molecules dosed in people whose developers credit AI with choosing the target or producing the molecule, counted to a written definition, with what happened to each, including the programs that quietly stopped. Most published counts do not release their list. This registry does, so the comparison below can be checked and recomputed by anyone.
What the data can say so far
For each step, the bar under its name shows what has happened to the 75 strict-cohort molecules that reached it. The chart beside it compares the share that moved on with what published studies report for drug development as a whole.
- Moved to the next phase
- Stopped
- Still running
- This registry: share that moved on, with 95% CI
- Industry: median of 5 studies
- Range of the 5 studies
The chart needs the registry data to load. The table below has the same figures.
Show as a table
| Step | Reached | Moved on | Stopped | Still running | This registry: moved on, of known outcomes (95% CI) | Industry: median (range of 5 studies) |
|---|
| Study | Data | Phase 1 to 2 | Phase 2 to 3 | Phase 3 to filing |
|---|
Why it is too early to call
- Small numbers. The intervals are wide because few molecules have an outcome yet. At Phase 2, the rate rests on a handful of molecules, and one more outcome either way would move it by several points; at Phase 3, on none.
- Most are recent. Most of the strict cohort first entered humans in 2022 or later. Phase 2 and Phase 3 trials take years, so most outcomes are still to come.
- The cohort may flatter AI. It holds only molecules whose developers credit AI in public, and developers credit their methods more readily when things go well. Each claim is recorded when it is made, so a molecule that later fails stays in.
- Early outcomes may not be typical. The first molecules to resolve are the earliest entrants and the fastest decisions. They may not represent the cohort that follows.
- Stopping is not always failure. A program can stop for money, strategy or a merger. The industry figures count those suspensions the same way,1 so the comparison is like for like on that point. Developers rarely say why a program stopped, so the registry records that it stopped, not why.
- Published benchmarks disagree. Five widely cited studies put Phase 2 success between 29% and 58%. Four of them, which count each drug's phase transitions, cluster around 30%. The fifth tracks every trial path and reports higher rates.6 The chart shows the full range rather than one number.
- Other counts reach other answers. An analysis of AI-native company pipelines to the end of 2023 found Phase 1 success of 80–90% and Phase 2 success of about 40%.2 It counted a different set of molecules under different rules; see Methodology § 6.
What would settle it
Outcomes, mainly from Phase 2. Several strict-cohort molecules are in Phase 2 or Phase 3 and still running. The registry checks their trial records every week and records what happens to them, and this page recomputes from the data each time.
What these numbers suggest so far is argued in the companion essay, Turning the Crank Faster (Backed By Reality, 2 November 2026).
How this is computed
- Only the strict cohort is counted. Borderline entries are not.
- The registry measures what public records show: target novelty, phase outcomes and timing. It cannot measure cost, which no public record reports.
- A molecule reached a phase if its furthest phase is that phase or later. It moved on if its furthest phase is later. A molecule approved without a registered Phase 3 trial counts as moving on from Phase 2 but not as reaching Phase 3. As in the industry studies, a combined trial counts as the later phase (Wong et al. state the convention6): a molecule whose first trial in people was a combined Phase 1/2 trial enters at Phase 2, so it is not counted in the Phase 1 step, and a Phase 2/3 trial counts as Phase 3. A molecule that entered at Phase 1 moved on once it reached Phase 2, including through a later Phase 1/2 trial. The home page chart still shows Phase 1/2 as its own column.
- A molecule that reached a phase but went no further is stopped if its status is stopped, and still running otherwise (including status unknown).
- The rate is moved on ÷ (moved on + stopped). The interval is the Wilson score interval. Until at least 10 molecules have an outcome at a step, the rate is shown in grey and not compared. Ten is the floor for showing a comparison, not an adequate sample: with so few outcomes the interval is wide, and it is the interval, not the rate, that should be read against the industry's figures.
- Industry figures are all-indication phase-transition success rates from five studies covering 1995–2020.1,3–6 The line is their median; the band is their range. The Phase 3 figures measure progress to filing, except Wong et al., which measures progress to approval.
Sources
- BIO, Informa Pharma Intelligence and QLS Advisors. Clinical Development Success Rates and Contributing Factors 2011–2020. February 2021. 12,728 phase transitions; likelihood of approval from Phase 1, 7.9%.
- Jayatunga MKP et al. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today, 2024.
- Hay M et al. Clinical development success rates for investigational drugs. Nature Biotechnology 32:40–51, 2014. Data 2003–2011.
- DiMasi JA, Grabowski HG, Hansen RW. Innovation in the pharmaceutical industry: new estimates of R&D costs. Journal of Health Economics 47:20–33, 2016. Compounds first tested in humans 1995–2007.
- Thomas DW et al. Clinical Development Success Rates 2006–2015. BIO, Biomedtracker and Amplion, 2016.
- Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics 20:273–286, 2019. Data 2000–2015.