Babies picked by algorithm: AI embryo selection sells a certainty that genetics doesn't have yet

THE THESIS. Embryo selection by genetic score is no longer science fiction. It is a product with a price tag, advertising, and backers such as Peter Thiel and Elon Musk, according to MIT Technology Review. Our reading separates two things that the marketing deliberately blends. One is modestly lowering the risk of serious disease. That technology is imperfect, but it has a scientific basis. The other is "optimizing" a child for IQ, height or longevity. There the evidence is weak and the ethical stakes are high. The near-term problem isn't a master race. It is a market that sells small probabilities as if they were control, with no law to regulate it.

THE MARKET ALREADY EXISTS. Genetic testing of IVF embryos has existed since the 1990s. It detects chromosomal abnormalities or single-gene disorders such as cystic fibrosis. The big shift came in 2019. That year Genomic Prediction launched the first clinical polygenic test, known as PGT-P. Orchid came next. It sequences the embryo's whole genome and focuses on health risks. In 2025 Nucleus Genomics and Herasight entered the market, and both now promise to score traits such as intelligence. Nucleus Embryo costs $5,999. The company says it analyzes up to 20 embryos across more than 900 conditions and more than 40 traits. MIT Technology Review estimates the full process can cost up to $50,000 and says PGT-P is offered at more than 100 US clinics. In January 2026 the magazine named embryo scoring one of its ten breakthrough technologies of the year. That choice needs careful reading. The magazine is flagging a technology that will matter, not one that has been shown to work. In the UK, selecting embryos for non-medical traits is illegal. Spanish law allows embryo diagnosis only for serious diseases, and only with authorization.

WHAT A POLYGENIC SCORE ACTUALLY MEASURES. A polygenic score is a weighted sum of thousands or millions of small DNA variants. Each variant nudges a trait or disease only slightly. The weights come from studies of hundreds of thousands of adults. The scores work reasonably well for some common, highly heritable diseases, such as type 2 diabetes or coronary disease. There they can separate high-risk groups from low-risk ones. But there are three limits the brochure leaves out. First, the scores predict population trends, not individual fates. The US National Human Genome Research Institute puts it this way: they indicate relative risk, not certainty. Second, comparing embryos from the same parents is harder than comparing strangers. Siblings share half their DNA, so the differences between them are small. Some of what a score picks up in the general population, such as family, environment and socioeconomic status, also disappears within a family. Third, most of the data comes from people of European ancestry, and accuracy falls for other groups.

THE FINE PRINT ON IQ. The most cited academic reference is a 2019 Cell study by Karavani and colleagues. It found that, with the technology of the time, picking the top-scoring embryo gained about 2.5 IQ points and about 2.5 cm of height on average, with very wide error margins. In real data from large families, the child with the highest height score was the tallest only 25% of the time. In 2024, a Japanese team presented a simulation at the European Society of Human Genetics. It covered 500 couples with 10 embryos each and tested six common scoring methods. No two methods picked the same embryo more than half the time. In other words, which embryo counted as "best" depended largely on the method. Herasight says its intelligence predictor, CogPGT, reaches a 0.45 correlation within families. Correlation measures, on a scale of 0 to 1, how closely two variables move together. According to the company, this makes gains of up to about 9 IQ points possible in the best case. Herasight validated the predictor on about 4,600 sibling pairs from UK Biobank, a large British database of genetic and medical records. That is more rigorous than what some competitors offer. But the figure comes from a model, not from children measured years later. The real gain also depends on something the algorithm can't control: how many viable embryos a couple has. Many have two or three, not ten. Our reading is that the debate isn't "it works" versus "it doesn't." A score can carry some statistical signal and still be worth little for one family's actual decision. It is the same mistake we flagged with deepfake detectors: the average figure hides where the system fails.

WHERE AI COMES IN (AND WHERE IT DOESN'T). This is where artificial intelligence enters, and precision matters. AI doesn't "design" babies or edit genes. It does two statistical jobs. The first is imputation. An embryo biopsy yields only a few cells and incomplete DNA, and models fill in the gaps using patterns learned from large databases. Herasight says its ImputePGTA algorithm can produce polygenic scores from routine chromosome screening (PGT-A), which is already used in more than 60% of US IVF cycles. The second job is prediction. In October 2025, Nucleus launched Origin, nine models covering conditions such as Alzheimer's, diabetes and several cancers. The company says Origin analyzes 7 million markers and was trained on data from 1.5 million people. Nucleus released the model weights openly. It also claims that with five embryos, couples can cut their risk of some common diseases by more than 50%. That figure comes from the company and hasn't been independently verified. Releasing the models openly is good news because it allows outside audits. But open code is not the same as validation. The key question is the one we ask of any medical AI: which population was it validated on, and who could audit the data? So far, the answer is thin. There is a Nucleus-linked observational study registered on ClinicalTrials.gov (NCT07841769) and approved in September 2026. It plans about 2,000 embryo transfers. But it will only measure whether picking embryos by genetic score lowers implantation rates compared with the classic method of judging how the embryo looks. It won't measure whether those children end up healthier or smarter. The protocol itself admits that no prospective study of these scores' usefulness exists.

STARTUP WARS AND THE GHOST OF THERANOS. The industry is also accusing itself. Herasight left Nucleus out of its technical comparison and wrote that Nucleus's scores would not add a "reliable or meaningful" analysis. In October 2025, Genomic Prediction sued Nucleus and two of its own former executives in federal court in New Jersey for alleged trade-secret theft. Nucleus denies the allegations. According to its law firm, it defeated the request for a preliminary injunction. Separately, a class action accuses Genomic Prediction of misleading patients about the accuracy of its tests, according to AboutLawsuits.com. Accusations of fake reviews against Nucleus also circulate on social media, and the company rejects them. None of these allegations has been proven in court. In July 2025, bioethicists Arthur Caplan and James Tabery published a Scientific American op-ed comparing the case to Theranos. They made clear that they see no deliberate deception of the Theranos kind at Nucleus. What they see are "striking parallels": turning a research tool into a consumer product that promises more than it can deliver. The American College of Medical Genetics and Genomics and the US reproductive medicine society agree that the science can't yet produce reliable estimates. Nucleus replies that its service helps families and never guarantees outcomes, only better odds.

IMPLICATIONS: WHAT'S MISSING IS A LAW THAT NAMES IT. The near-term problem is concrete. Families already carrying the cost and stress of fertility treatment pay thousands of dollars for a ranking that can change with the scoring method. Picking for IQ can also mean passing over the embryo most likely to implant. And the US is normalizing market-driven eugenics: no one imposes it, but the price makes it a privilege. We are not arguing for a ban driven by panic. We are arguing for regulation based on evidence. First, the law should separate medical screening from trait selection. Second, every company should have to publish how accurate it is within families and for each ancestry. Third, advertising gains that haven't been measured in real people should be banned. In the long run, we are optimistic. Genetic prediction, improved by AI and validated on data from millions of people, could genuinely help prevent diabetes, heart disease and some inherited cancers. It could be part of a medicine that extends healthy life. But that future will be built on trials and transparency, not on ads urging parents to pick their "best baby." A child's intelligence doesn't fit in a score. Today, science can't even rank it reliably.