So many ways genomics can go wrong

Consumer-tech thinking can kill people in genomics

I have worked with people from online video, consumer retail and ordinary digital-product backgrounds who looked genuinely confused when I challenged a decision: “But I’m the project manager; it’s my call.” My reply was: “I know, but I’m the geneticist, and if you implement this algorithm incorrectly, people could die needlessly. I’m sure you don’t want that.”

In most consumer technology, a bad product means a frustrated user, a cancelled subscription, a returned item, a complaint, a bad review or a product manager saying “we’ll iterate”. In clinical genomics, a bad product can mean a missed diagnosis, a wrongly reassured parent, a cancer variant not reported, an embryo selected on weak evidence, or a baby sent home without the treatment that might have saved them.

This is why genomics is not just software. Yes, there are pipelines, APIs, dashboards, cloud platforms, machine-learning models and databases. But the rows in the table are not customers in a funnel; they are patients, embryos, tumours, newborns, families and sometimes people who are already critically ill. It is a balancing act: providing life-saving care quickly while remaining cautious and investing heavily in quality and testing. If something goes wrong, establishing accountability for a genomics product designed and tested by perhaps 50 people over several years can be nearly impossible. Correcting the error may be too late.

Go to top

Clinical genomics may scale faster than quality control

A genomic test is not one simple event. It is a chain of sample handling, consent, sequencing quality, variant calling, annotation, interpretation, reporting, counselling, storage and possible reanalysis. That chain can easily fail quietly – a wrong database entry, a bad phenotype, a pipeline giving a rare unvalidated result, an ancestry mismatch, poor reporting or a clinician without enough support can all turn a technically impressive test into a bad medical decision. Professional standards exist for a reason, including the widely used ACMG/AMP guidelines for sequence variant interpretation.

The boring parts of genomics—validation, documentation, governance, audit trails, clinical oversight—are essential. Everyone I have worked with over the years has enormous respect for the difficulty of getting it right, consistently.

Go to top

Gene therapy can be miraculous, and unforgiving

Gene therapy is one of the reasons I am not anti-genomics. When it works, it can look like a miracle. Children who were going blind can gain functional vision. People with sickle-cell disease can receive therapies aimed at changing the biology that has caused them pain, hospitalisations and shortened lives. The FDA has approved gene therapies for inherited retinal disease and sickle-cell disease, including Casgevy, the first FDA-approved therapy using CRISPR/Cas9 genome editing.

But the body is not a neutral container waiting politely for our best ideas. Gene therapy has to get genetic material into the right cells, at the right dose, for long enough, without provoking the immune system into catastrophe, inserting DNA in a dangerous place, damaging the liver, triggering cancer, or causing harm that only appears years later. The immune system is doing the job that usually keeps us alive, but it can go into overdrive an become lethal.

The history of gene therapy contains extraordinary success and terrible loss. In 1999, 18-year-old Jesse Gelsinger died in a safety trial for ornithine transcarbamylase deficiency after receiving an adenoviral vector; reports describe a severe immune reaction to the vector and death four days later. He was not a terminally ill patient at the end of life. He had a milder form of the disease and joined a trial intended to help develop treatment for babies with the severe form.

Early gene therapy trials for X-linked severe combined immunodeficiency also showed the double truth of the field. Some children had dramatic immune restoration. But later follow-up reported that acute leukaemia developed in four patients, and one died, after retroviral gene therapy. The treatment had helped save children from one deadly disease while revealing another danger: a therapeutic gene can land in the genome in a way that helps drive cancer.

More recently, high-dose AAV gene therapies have raised new safety alarms. In the ASPIRO trial for X-linked myotubular myopathy, four participants died after receiving AT132, all with cholestatic liver failure at the time of death. In 2025, the FDA said it had received reports of fatal acute liver failure following Sarepta AAVrh74 gene therapies, including Elevidys for Duchenne muscular dystrophy.

These are not arguments for giving up. Many of the people entering these trials have devastating diseases and few options. Their families are not reckless for wanting hope. Scientists and clinicians are not wrong to keep trying. But every single loss feels like one too many.

Gene therapy teaches the central lesson of genomic medicine: we should move forward, but with humility. The goal is to learn enough to make life-saving treatments possible, while understanding that manipulating the body is never simple.

Go to top

Investors may fund biology they do not understand

I have seen investors pour serious money into genomic and AI-health technologies they did not understand. Not just technologies where they misunderstood the market, but technologies where they seemed not to fully understand the uncertainties of biology, or the real risks.

This matters because money accelerates weak ideas as well as good ones. A company with a visionary founder, a polished deck, a famous investor and a few scientific words arranged in the right order can look much more solid than it is.

Theranos Inc. is a stark warning. A company promised to transform blood testing, attracted enormous investment and prestige, and collapsed when the claims could not survive contact with reality. Genomics and AI-health are vulnerable to the same failure mode because the science is complex, the language is impressive and most investors cannot easily tell the difference between a genuine breakthrough and a beautiful slide deck full of completely unrealistic but impressive sounding claims.

Biology is messy. Cells are messy, patients are complicated, datasets are biased, clinical evidence is expensive, and an elegant dashboard cannot rescue a product built on a weak scientific claim. If your background is in video-game development, astrophysics, AI or electrical engineering, you are probably a very smart person. You may still, in my experience, be massively underestimating the complexity of biological systems. Much of the apparent messiness of cell biology is functional: layered biological fail-safes make us more robust as an organism – which is a good thing. It can also effectively counteract our futile attempts at stable outcomes of gene manipulation in the lab – which can be very bad for your investment, and your customers.

Go to top

Direct-to-consumer genomics sells false certainty

I have seen direct-to-consumer genomic services marketed in ways I considered deeply misleading. The branding is usually friendly and empowering: know yourself, optimise your health, personalise your diet, understand your future, choose better, live longer.

Behind that soft language may sit weak evidence, uncertain interpretation, poor follow-up, vague consent and unclear data reuse. The consumer may think they are buying a simple personal report, while actually entering a commercial genomic ecosystem involving relatives, future children, commercial databases and medical claims they may not be able to judge, while their data is being quietly sold to someone else.

The National Human Genome Research Institute notes that many genetic tests still go to market without independent analysis to verify the seller’s claims. It seems to me wreckless to leave the general public alone to navigate glossy promises and a checkout button – most of us simply do not know enough to make an informed choice.

Go to top

AI can make fragile biology products look convincing

I have seen AI-health ideas sold with astonishing confidence despite weak foundations. In one case, I saw a startup offering AI services for healthcare win a multi-million-dollar contract with a large hospital, while the actual offering seemed to me technically fragile and badly underdeveloped.

That is exactly the kind of thing that worries me. A concept sounds visionary, hospital executives like the language, investors like the scale, and suddenly something fragile is being implemented in clinical reality, influencing life and death decisions.

AI is a somewhat dangerous force multiplier. If the data are biased, AI scales bias. If the workflow is wrong, AI automates wrong answers. If the claims are vague, AI makes them sound more accurate or confident than they are. If governance is weak, AI makes the whole system faster, larger, more opaque and harder to challenge.

Go to top

Genomic data leaks can be permanent

Genomic data are not like a password. You cannot reset it after a breach, and it can reveal information about relatives, children and future descendants who never agreed to be part of the dataset.

The 23andMe breach is the obvious warning case. Reuters reported that 23andMe agreed to settle a data-breach lawsuit involving 6.9 million customers, and the incident showed how family-matching features can spread risk far beyond the accounts directly accessed.

The UK Biobank incident makes the point even harder to ignore, they had to tell participants that de-identified participant data had been listed for sale on a Chinese consumer website owned by Alibaba. The data did not include names, addresses, exact dates of birth, and apparently the listings were removed before any sale was made – at least not on that platform. Expect to see more such incidents in the future, unless the data is stored and managed in a completely different way.

Go to top

Consent cannot carry this much weight

Many genomics systems rely on consent, but the public cannot realistically understand every future use of genomic data. People are asked to consent to storage, reanalysis, research, data sharing, commercial access, international collaboration, AI training, family implications, incidental findings and technologies that do not yet exist.

So it is clear that consent alone cannot carry the whole ethical burden. We need strong default protections, institutional duties, audit trails, penalties, data minimisation, participant representation and limits on future use.

Otherwise, informed consent becomes a ritual that protects institutions more than people. In genomics, the most important ethical question is not only “did the person click yes?” but whether the system should exist in that form at all. There are perfectly good options, such as deleting the data after interpretation, or engineering the system so that the person always remains in active control over their own data. Taking the lazy option of “give me all your data and let me do whatever I want with it” may sound good for a company, and be a perfectly valid approach when the data is music playlists, but not when the data is someones genomic information.

The problem becomes even more serious when companies fail, merge or are sold. Once genomic data has been copied, linked, shared, merged with health records or used in AI training, even “delete my data” may be much less powerful than people imagine.

Go to top

Population genomes can become private assets

Population genomics is often presented as a public good, and sometimes it is. Large genomic datasets can help research, diagnosis and drug discovery, especially when they are governed well and built with public trust.

But Iceland shows how quickly the genomes of a population can be extracted, and lost from their control. In 1998, Iceland passed the Health Sector Database Act, allowing a private company the right to build and operate a national health database with exclusive rights to national health data, and then make it legal for that company to construct an electronic database of the country’s health records.

This was not just a few people filling in a research form. This was an entire small country’s health, genealogy and genetic potential being turned into a commercial research platform forever more. The Icelandic population was valuable precisely because it was small, relatively isolated, medically well documented and genealogically traceable.

The company; deCODE Genetics, later was aquired by the large international company Amgen for USD 415 million. And that is how easy an entire country’s population genomic data was be converted into a private company’s strategic asset, with the people of Iceland no longer in control the use of their data.

That is why large genome projects need more than cheerful language about innovation; they need serious democratic debate, data management allowing active opt-outs and opt-ins, strong long-term governance, and protections strong enough to survive acquisitions, bankruptcies and political change. Think national central bank level type of control – a genomic Fort Knox.

Go to top

Genomic extraction can target underrepresented populations

Because existing genomic datasets are biased, underrepresented populations become scientifically and commercially valuable. That creates the risk of genomic extraction: samples and data flow out of a country or community, while patents, products, profits and publications are made elsewhere.

Yes, we urgently need more African, Indian, Indigenous, island, isolated and historically understudied populations represented in genomics. Without that, medicine will be less accurate and less fair. But representation alone is not enough, if the data are extracted, analysed elsewhere, commercialised elsewhere and sold back as products the source populations cannot afford.

India provides a concrete warning. Two biomedical companies collected genetic data from more than 15,000 Indians , with and the samples from five Indian states were sent to the US. This is especially shocking to anyone who knows how tightly India restricts the export of biological material.

This is why global genomics cannot simply mean “more samples in large databases.” The answer has to include local leadership, participant-led science, fair contracts, benefit sharing, capacity building and medicines or diagnostics that people participating can actually access. Better representation in genomics is essential, but representation without power becomes extraction and exploitation.

Go to top

Genomics worsen health inequalities

Many genomic datasets still overrepresent people of European ancestry. That means the reference data used to interpret variants, estimate disease risk, build polygenic scores and discover drug targets may work better for some populations than others.

Researchers have warned that clinical use of current genomic polygenic risk scores may worsen health disparities because they often perform less accurately outside European-ancestry groups. This is not a marginal technical flaw; it is how old medical inequality is perpetuated.

What makes this even more absurd is that Africa contains the greatest human genetic diversity on Earth, and yet African genomes remain deeply understudied. The scale of what is missing is astonishing. A major H3Africa study sequenced 426 individuals from 13 African countries and 50 ethnolinguistic groups, and uncovered more than 3 million previously undescribed genetic variants. That is a flashing warning sign that global genomics has been built on a dangerously incomplete picture of human variation.

With genomic medicine trained mostly on European-ancestry datasets, the tools, treatment and medical discoveries that follow will all tilt in the same direction, and be less accurate for the people who were left out at the beginning.

We also need to respect fully the difference between genetic ancestry and where a person lives now. Genetic ancestry can matter for health and for the interpretation of a genomic diagnosis, but it is not a reason for anyone to receive worse care than others in the same healthcare system. This is already an issue for example with organ donations – that it is difficult to find a matching donor with enough similar genetic background, if you have a minority ethnic background. Genomics, genomic medicine and AI in healthcare will almost certainly exacerbate health inequalities.

Go to top

Embryo selection can turn parental fear into a luxury market

This risk is not theoretical for me. I once interviewed for a company that wanted to offer polygenic risk scores for embryo selection, and I declined the job. I did not decline because embryo testing is always wrong. Screening embryos for serious monogenic disease can for some families be profoundly important. I declined because the commercial direction of polygenic embryo selection worries me deeply.

The problem begins when embryo selection moves from avoiding severe childhood disease into probabilistic ranking for complex traits: future disease risk, height, intelligence, psychiatric risk, longevity or general “genetic quality”. A 2024 review of polygenic embryo screening warned that possible societal harms include discarded embryos, increased demand for “designer babies,” overemphasis on genetic determinants of disease, unequal access and lower utility in people of non-European ancestries. For everyone who watches reality shows with parents with body dysmorphia taking their children to pageants, or encouraging them to do cosmetic plastic surgery – this is what the genomic future could also be like, if we allow it!

Go to top

Genomic surveillance may expand quietly

Genomic data are useful far beyond healthcare. It can be used in policing, immigration, insurance, employment, education, fertility, sport, ancestry databases and national security.

The danger is function creep. A person may give a sample for healthcare, but a relative’s ancestry test can make them easier to identify. A forensic database built for serious crime may expand into routine identification. Insurance companies or employers may not need a whole genome to discriminate; probabilistic risk signals may be enough.

Europe’s data-protection framework treats genetic data as sensitive personal data, and the United States has legal protections against genetic discrimination in health insurance and employment. Those protections matter, but they do not resist every future use of genomic inference by police, insurers, employers, platforms or states. Once again – your genome is the most sensitive data you have, by far. If I knew yours, I could easily get away with killing you, I could blackmail you, hedge bets on you and your family, or give you a allergic reaction (but I won’t because I’m a nice person).

Go to top

Engineered biology may leak into the biosphere

As genome editing and synthetic biology become cheaper and more powerful, biological pollution becomes a real risk. Humans are starting to create novel genes, synthetic pathways, engineered microbes, gene drives and powerful enzymes. Once that material enters soil, crops, insects, microbes, waterways or wild ecosystems, there may be no reliable way to get it back.

The problem is not just one escaped organism. It is persistence, spread and recombination. DNA can move. Microbes can swap genetic material. Organisms reproduce. Ecosystems are connected in ways we rarely understand until something goes wrong.

We have seen this pattern before with other human inventions. Plastic was useful, cheap and controllable — until it was everywhere, including inside our own bodies. Engineered genes, viruses or microbes could create a similar problem, except the material may also copy itself, mutate and evolve.

This is not an argument against environmental biotechnology. Some engineered organisms may help clean pollution, protect crops or restore damaged ecosystems. But once modified life is released, the usual logic of product safety breaks down. You cannot simply recall a gene from a river, a field, a mosquito population or a microbial community.

Go to top

AI-designed biology may outrun safety testing

Biotechnology increasingly uses living cells as factories. Engineered microbes and cell lines already produce medicines, enzymes, food ingredients, chemicals and materials. AI now makes this more powerful by helping researchers design proteins, pathways, binders, enzymes and biological systems that nature did not conveniently evolve for us.

That is thrilling if you want cleaner chemistry, cheaper medicines, better materials or less destructive manufacturing. It is also risky. A model can optimise for “binds strongly,” “cuts efficiently,” “makes more product” or “survives industrial conditions” without understanding every context that molecule, gene or organism may later enter.

A clever enzyme in a sealed production vat is one thing. The same enzyme in wastewater, soil, gut microbes, a contaminated bioreactor, a damaged supply chain or another species is a different question. Biology is not a stainless-steel pipe. The factory is alive.

The risk is not that AI-designed biology is automatically dangerous. The risk is speed. Design may become cheap, fast and automated, while testing remains slow, expensive and awkward. That gap is where mistakes happen.

AI-designed biology needs boring, suspicious safety work: containment, evolution, escape, by-products, host effects, vulnerable populations, ecological exposure and what happens when the clever molecule meets the stupidly complicated real world.

Go to top

Biosecurity is not science fiction

Gene editing, DNA synthesis, synthetic biology and AI-assisted design are turning biology into something more searchable, editable and buildable. That is thrilling in medicine. It is also why biosecurity can no longer be treated as a niche concern for specialists in windowless rooms.

Putin’s reported toilet protocol sounds like autocrat slapstick: a travelling team whose job includes collecting the president’s poo. Multiple outlets have reported that Vladimir Putin’s security staff collect his faeces and urine on foreign trips, allegedly to stop foreign intelligence services from analysing clues about his health.

A persons poo might reveal illness, medication, infection, diet, and human DNA. A genetic marker that makes one drug safe for most people can make the same drug dangerous for someone else. Precision medicine uses that knowledge to treat people better. Precision targeting asks the nastier question: what could an enemy do with the same biological information?

Most of us do not need hazmat suits or poo collectors to go to a party. But if powerful people like Putin — who clearly knows a lot more about this topic than we do — guard his biological traces like state secrets, perhaps the rest of us should stop for caution?

Go to top

Scaling changes the risk

Some genomics and reproductive technologies look very different at the scale of one person, one clinic, or one product than they do after mass rollout. A single successful case can be celebrated as a breakthrough. Ten years later, the same technology may be an industry, and the real consequences may only become visible when it has touched thousands, or millions of people.

IVF is the obvious example. The world’s first IVF baby, Louise Brown, was born in 1978, a milestone recognised by the Nobel Prize as the breakthrough that opened modern reproductive medicine. Since then, assisted reproduction has moved from extraordinary to increasingly ordinary; the European Society of Human Reproduction and Embryology estimates that more than 10 million babies have been born worldwide since the first IVF baby, and the UK fertility regulator reported that IVF accounted for 3.1% of UK births in 2023, roughly one child in every classroom.

IVF has brought millions of wanted children into the world, and for many families it is a profound medical and personal good. But it shows how quickly a technology can move from miracle to infrastructure, and infrastructure always deserves a different kind of scrutiny. So what are the long-term, population-level effects when IVF accounts for 3% of births? I don’t know if anyone really knows?

Many other genomic technologies may create similar scaling problems, but with higher stakes. What happens when 1%, 5% or 10% of children are selected using embryo polygenic scores? What happens when 5% or 10% of people carry a genetic modification, or when a longevity product claims to extend healthy life by 20% and targets hundreds of millions of customers? What happens when a gene-editing company does not aim for a rare disease population, but for 10% of the world? I don’t know if there is any international oversight body which can regulate on such matters, and prevent distortion of the entire human population genetics.

So every genomics company should be asked a question that is almost never in the pitch deck: if you succeed completely, what happens then? If your product reaches 10% of the world’s population, 10% of births, 10% of crops, 10% of embryos or 10% of hospitals, what will actually happen to humanity?

Go to top

Evolution cannot be controlled

A biological product can look safe at approval and behave differently after years of evolution, spread and selection. That is because living systems keep on evolving long after the paperwork is finished.

Glyphosate-resistant creeping bentgrass is a clean warning. It was planted in Oregon under regulated conditions, but the engineered glyphosate-resistance trait moved through pollen-mediated gene flow. Researchers later found the transgene established in feral creeping bentgrass and compatible relatives outside the original control area. Approval had a boundary, but the pollen did not.

This is the risk: a trait designed for one field, crop, or pest can change once millions of organisms and hectares get involved. The safety question cannot only be: “Is this safe at launch?”. It also has to be: How can this evolve? Where can it move? What can it cross with? What resistance will it select for? Who is monitoring longterm effects? You can never fully predict or protect against evolution.

Go to top

International oversight is missing

Genomics can now change humans, organisms and ecosystems in ways national laws are not built to contain.

The warning is already here. He Jiankui used CRISPR to edit human embryos, leading to the birth of the first gene-edited babies in 2018. He was later sentenced to three years in prison, but prison did not undo the edit. Those children are alive with deliberately modified genomes. If they have children, those edits may enter future generations. One researcher crossed a line, and the human gene pool changed.

Fertility shows the same governance failure at a different scale. In 2023, a Dutch court ordered a sperm donor to stop donating after it was found he had fathered at least 550 donor children on several different continents, despite telling parents he would father no more than about 25. The court recognised the consequences for identity, psychosocial harm and accidental incest or inbreeding. The ruling could stop future donations. It could not undo the fact that 550 children were born, or the consequences for all of their descendants.

The World Health Organization has published an 87-page governance framework for human genome editing. But there are no enforceable global systems or oversight for a range of very powerful genomic technologies. For example, reproductive genomics, embryo selection, germline editing, synthetic embryos, engineered microbes, gene drives and synthetic biology currently move much faster than law and international coordination.

Who can stop a clinic, company, state, donor, billionaire or researcher before changes to the human gene pool or biosphere becomes permanent? Who tracks consequences across borders and generations? Who supports the children, descendants, ecosystems and species affected? How can one authority allow changes that affect all of humankind?

Without enforceable international oversight, the fastest, richest or least scrupulous actors can move first, while the rest of humanity and planet Earth is forever changed.

Go to top

Ecosystem effects are unpredictable

A transgenic organism, edited organism or newly introduced species does not arrive in an empty container. It arrives in an ecosystem already full of species eating, competing, avoiding, infecting, parasitising, pollinating and depending on one another. The new organism can and will change the relationships around it.

Some other species may be pushed out. Some may adapt. Some may evolve to exploit the newcomer. Some will change their behaviour or abundance. The effect may be small, or it may ripple through the system in ways nobody predicted at launch.

Bt crops is an example; they were designed to kill target insect pests by producing proteins from a bacteriaBacillus thuringiensis. The intended biology worked. But once deployed across real farms, the crops became part of a larger ecological system: pest populations, non-target insects, pesticide use and resistance evolution. A major review of monitoring data from 5 continents found that pest evolution can undermine Bt crops when resistance management fails.

The warning is not that Bt crops are uniquely bad. The warning is that living interventions do not remain single interventions. They become ecological events. The safety question cannot only be: “Does it do what we designed it to do?”. It also has to be: What else eats it, avoids it, hosts it or depends on it? Which species are pushed out, and which are favoured? What new interactions does it create? What happens after 10, 20 or 50 generations? What happens when it meets other engineered organisms? Who is responsible if the ecological harm appears years later?

Go to top

Good intentions are not enough

Purdue Pharma is a medical-commercial warning story. Their unsafe medicine caused an opioid addiction crisis, and showed how an approved medical product can become societally dangerous when aggressive marketing and distorted incentives push use far beyond what caution should allow. We need to learn from that before poorly thought through genomics and AI become healthcare infrastructure.

Most — if not all — people building genomics are not villains. They are excited, ambitious and often sincere. My colleagues want to cure disease, build companies, publish papers, help families, and modernise healthcare.

But good intentions do not replace clinical evidence, engineering discipline, regulation, humility or public debate. A founder can believe they are democratising healthcare while selling nonsense. An investor can believe they are accelerating innovation while funding dangerous products. A hospital can believe it is being visionary while buying a broken system.

The genomic revolution needs ambition. It also needs people willing to ask difficult questions before the product launch, before the embryo test, before the database transfer, before the hospital contract and before the engineered organism leaves the controlled environment. This requires risk management on a different level from almost any other sector, and a firm rejection of “enshittification”. We cannot compromise the safety of people, future generations or ecosystems for simple cost-cutting exercises or pressure to launch a product by a particular date. Many healthcare systems are already doing a really good job of protecting patients, and I wish similar care would percolate throughout the rapidly growing genomics industry.

Go to top