Sometimes, but far less often than the pitch suggests. A synthetic control arm uses historical trial data or real-world records, often matched with AI, in place of a randomized placebo group. It can work in rare, severe diseases with predictable courses and large effects. For most common conditions, the risk of hidden bias is too high.
Ask a parent of a child with a rapidly fatal rare disease whether their child should be randomized to placebo, and you will understand why synthetic control arms exist. When a disease is devastating and predictable, and when there are only a few dozen patients in a country, a classic placebo-controlled trial can be both unethical and impossible to fill.
In April 2017, the FDA approved cerliponase alfa for CLN2 disease, a form of Batten disease that destroys children’s ability to walk and speak. The pivotal evidence was not a placebo-controlled trial. It was 22 treated children compared against 42 untreated children from a natural history cohort. The treated children lost motor function far more slowly than the historical comparison group. Nobody seriously doubted the effect.
That is the best case for external controls. The question now being asked, with increasing commercial energy, is whether AI and large real-world datasets can take that approach beyond rare diseases into mainstream drug development. Could machine learning, trained on electronic health records and past trials, build a “synthetic” placebo group so good that sponsors could skip randomizing patients to it?
The thesis here is blunt. External controls work in exactly the circumstances where you needed the least help: rare, severe diseases with predictable courses and large treatment effects. Everywhere else, AI makes external controls look more convincing without making them more reliable. Better matching on the variables you can see does nothing about the variables you cannot.
What is a synthetic control arm?
A synthetic or external control arm is a comparison group drawn from outside the trial. Instead of randomizing some participants to placebo or standard care, investigators give everyone the experimental treatment and compare their outcomes with a group of similar patients from elsewhere.
That comparison group can come from earlier clinical trials, disease registries, natural history studies, insurance claims or electronic health records. The FDA’s February 2023 draft guidance on externally controlled trials describes both historical controls, drawn from earlier time periods, and concurrent controls, drawn from other settings at the same time.
Where AI enters is the matching. Real-world patients differ from trial participants in countless ways. Machine learning methods can build comparison groups that resemble trial participants on dozens or hundreds of recorded characteristics, and can extract details from clinical notes that older statistical approaches ignored. Several companies now sell platforms that promise exactly this.
When do external controls work?
The conditions are fairly well understood.
The disease course must be predictable. If untreated patients reliably decline at a known rate, a departure from that trajectory is easy to see. CLN2 disease fits. So do some rare pediatric neuromuscular and metabolic conditions.
The effect must be large. A treatment that halts a progressive disease, or produces tumor shrinkage that rarely happens spontaneously, stands out against almost any reasonable comparison. A treatment that reduces heart attacks by 15 percent does not.
The outcome must be objective and measured consistently. Survival, or a validated motor scale measured the same way in the trial and in the natural history cohort, is far more robust than symptoms recorded differently across clinics.
And randomization must be genuinely impractical or unethical. The FDA’s guidance frames external controls as an option where randomized trials may be impractical, not as a cheaper substitute when they are merely inconvenient. It asks sponsors to prespecify the design and analysis, obtain patient-level data for the external group where possible, and run sensitivity analyses on potential bias. The European Medicines Agency has drafted a reflection paper on single-arm trials as pivotal evidence that strikes a similarly cautious tone.
Why are external controls risky almost everywhere else?
Because randomization does something no matching algorithm can. It balances, on average, the factors nobody measured.
Patients who get a new treatment in a trial are selected. They met eligibility criteria, agreed to participate, turned up for visits and were cared for by research teams who monitor closely. Real-world patients who did not get that treatment differ in ways that records only partly capture: frailty, social support, adherence, how aggressively their doctors intervene, how outcomes were recorded. Matching on age, diagnosis, lab values and prior treatments helps. It cannot match on things that were never written down.
The best evidence on how well non-randomized designs can reproduce randomized results comes from RCT-DUPLICATE, a project led by Shirley Wang and Sebastian Schneeweiss at Brigham and Women’s Hospital. In a 2023 JAMA paper, they used insurance claims data to emulate 32 randomized medication trials. Overall, the correlation between the database estimates and the trial results was 0.82, and the two agreed on statistical significance 75 percent of the time.
The detail matters more than the headline. In the 16 trials where the design could be closely emulated, agreement was high: a correlation of 0.93 and 94 percent agreement on significance. In the 16 where emulation was difficult, the correlation fell to 0.53 and significance agreement to 50 percent, a coin flip. The authors noted obstacles such as therapies started in hospital, trial run-in phases that select for responders, and patients in routine practice staying on treatment for shorter periods.
That is a study by some of the most careful practitioners in the field, using established drugs with known effects. A synthetic control arm for a new drug has no randomized answer to check against. You never find out which half you were in.
Does AI make synthetic controls more reliable?
It makes them more precise. Whether it makes them more accurate is a different question.
Machine learning can match on far more variables than traditional propensity score methods, and can pull information from free-text notes. That reduces some bias. But it can also create a false sense of security. A comparison group that matches trial participants on 200 recorded variables looks overwhelmingly convincing in a table. If the key difference is on variable 201, which nobody recorded, the table is irrelevant.
There is a further risk specific to flexible models. When there are many ways to build the comparison group, there are many ways to get the answer you want. That is why the FDA’s emphasis on prespecification matters so much. An external control whose construction was fixed before outcomes were examined is evidence. One tuned afterwards is advocacy.
We made a related argument in real-world evidence is winning the easy cases. Real-world data succeeds when the question is narrow, the effect large and the confounding manageable. Synthetic control arms follow the same rule, and AI does not repeal it.
A worked example: two submissions
Consider two hypothetical programs, with illustrative details.
Program A is a gene therapy for a rare pediatric neuromuscular disease. Untreated children lose the ability to sit unaided by a predictable age, documented in a well-run natural history registry using the same motor scale as the trial. The sponsor treats 20 children. Most reach motor milestones that essentially no untreated child in the registry reached. The external comparison is prespecified, uses patient-level registry data, and the effect is so large that plausible bias could not explain it. Regulators are likely to take this seriously, and randomizing children to placebo would be hard to justify.
Program B is a new drug for type 2 diabetes aiming to reduce cardiovascular events. The sponsor proposes to treat 3,000 patients and compare them with 30,000 AI-matched patients from health records, avoiding the cost of a placebo arm. The matching is excellent on paper. But the expected effect is modest, perhaps a 10 to 20 percent relative reduction, and cardiovascular outcomes depend on adherence, lifestyle, access to care and clinician behavior that records capture poorly. Trial participants are known to be healthier than similar patients in routine care. A bias of the same size as the claimed benefit is entirely plausible.
Program A is the case external controls were built for. Program B is the case they are being sold for. The difference is not the quality of the algorithm. It is the ratio between the effect size and the plausible bias.
The strongest case for synthetic control arms
The counterargument has force, and it is partly ethical.
Placebo arms have costs. Patients randomized to them forgo a treatment that might help. Trials that require placebo arms are slower to fill because many patients will not accept a chance of getting nothing. In rare diseases, the population may be too small to run two arms at all, and insisting on randomization can mean no trial happens.
There is also a middle ground. Hybrid designs randomize fewer patients to control, for example two to one in favor of treatment, and supplement the smaller control group with external data. That keeps randomization as an anchor while reducing the number of patients on placebo. Used this way, external data augments a trial rather than replacing its foundation.
Those points are right, and in rare diseases they should prevail. But the ethical argument cuts both ways. Approving a drug on biased evidence exposes thousands of future patients to a treatment that may not work, and delays the search for one that does. As we argued in fast approval isn’t the danger, slow withdrawal is, weak evidence at approval is costly precisely because it is so hard to reverse later.
What this means for sponsors, regulators and patients
For drug developers, the practical rule is to match the design to the effect size. If the expected effect is large and the disease course predictable, invest in high-quality natural history data and prespecified external comparisons, and talk to regulators early. If the effect is modest, a synthetic control arm is likely to cost more in regulatory questions than it saves in trial costs. Hybrid designs are often the realistic compromise.
For regulators, the challenge is consistency. External control analyses built with flexible AI methods need the same prespecification, transparency and sensitivity analysis as any other, plus access to the underlying data and code. Clear public examples of accepted and rejected designs would help sponsors avoid wasting years.
For data providers and AI vendors, the credible path is validation against randomized results, as RCT-DUPLICATE did, and honest reporting of where emulation fails. A vendor that claims its synthetic arms can replace randomization across therapeutic areas is claiming something the best published evidence does not support.
For patients and advocates, especially in rare diseases, external controls can mean faster access and fewer children on placebo. The price is uncertainty about how well a treatment really works, which lands on patients and clinicians after approval. That trade-off is worth making in the right cases, and worth questioning in the rest.
Frequently asked questions
Has the FDA approved drugs using synthetic control arms?
Yes, mainly in rare diseases and some cancers. Cerliponase alfa for CLN2 disease was approved in 2017 based on 22 treated children compared with 42 untreated children from a natural history cohort. The FDA’s 2023 draft guidance sets out how sponsors should design externally controlled trials, emphasizing prespecification, patient-level data and sensitivity analyses. Such approvals remain the exception rather than the norm.
What is the difference between a synthetic control arm and a placebo group?
A placebo group is randomized within the trial, so on average it resembles the treated group in both measured and unmeasured ways. A synthetic control arm is assembled from outside data, such as earlier trials or health records, and matched to trial participants on recorded characteristics. It cannot balance factors that were never measured, which is its central weakness.
Can real-world data reproduce randomized trial results?
Sometimes. The RCT-DUPLICATE project, published in JAMA in 2023, emulated 32 trials with insurance claims data. When trial designs could be closely emulated, agreement was high. When they could not, agreement on statistical significance fell to about 50 percent. Real-world analyses work best for narrow questions with measurable confounders and well-defined treatments.
Are synthetic control arms more ethical than placebo?
In rare, severe diseases with no effective treatment, they can be, because they spare patients from being randomized to nothing. In common diseases with modest expected effects, the ethical balance shifts, since biased evidence can lead to approval of ineffective drugs that many future patients will take. Hybrid designs, with smaller randomized control groups supplemented by external data, offer a middle path.
The takeaway
Synthetic control arms are a real tool for a narrow set of problems. AI makes them easier to build and harder to argue with, which is not the same as making them right. The test is simple: could a plausible hidden bias be as large as the effect you are claiming? If yes, randomize.
Where would you draw the line for external controls: rare diseases only, or is there a case in common conditions too? Tell us in the comments.
