Test the Grand Continent. The ideas and analyses you won’t find anywhere else. But you’ll see them everywhere soon. In your inbox starting at 8 euros per month
Bill Gates, the whistleblower Jacob Coxon, and tech leaders themselves, such as Dario Amodei or Sam Altman, warn that AI could “create biological weapons capable of eradicating the human species.” Yet the same voices also claim it could “cure all cancers within ten years.” Beyond the headlines, what is the real impact of AI on biotechnology? Should we fear extinction through super-biological weapons or hope for curing all cancers within ten years?
AI as pharmakon: the same tools forge poisons and remedies
Biotechnologies can produce both cures for cancer and devastating biological weapons—or chemical ones, a distinction that is often misjudged. So why not concentrate solely on the tools that yield cures, while leaving aside the most dangerous ones?
François Jacob, a companion of the Liberation, a French biologist and Nobel Prize in Medicine in 1965, noted: “Natural selection operates not as an engineer, but as a tinkerer; a tinkerer who does not yet know what he will produce, but recycles whatever comes to hand.”
This quotation offers a crucial clue: the entire living world (in its broad sense, including viruses) is intrinsically connected. Natural selection has built life by endlessly recycling sequences of DNA, protein domains, membranes and biomolecules. The protein domain of a deadly bacterial toxin could thus be found in the active site of an enzyme essential to fighting the proliferation of cancer cells. The tools used to develop cancer remedies are therefore, roughly speaking, the same ones used to develop biological and chemical weapons. Poison and remedy intertwine in a single “AI pharmakon” that carries all possibilities within it.
AlphaFold, which probably constitutes the first major “proof of concept” of AI’s full potential in biotechnology, perfectly embodies this idea. Developed by Google DeepMind, it determines the three-dimensional structure of proteins—a breakthrough rewarded with the 2024 Nobel Prize in Chemistry, and one from which humanity will no doubt reap substantial benefits. It can help design a protein capable of targeting cancer cells and aiding healing, just as it could design a deadly toxin intended to kill.
If the concept of a “pharmakon” is especially apt, it is because Socrates, more than twenty-five centuries ago, already directed to writing the criticisms that many direct at AI today. Writing, in his view, “acts like a drug,” “drives people to rely on it and to delegate to it their cognitive activity,” and “reduces or even eliminates the friction necessary to sustain human cognition.”
In Plato’s Pharmacy, Jacques Derrida shows that Plato does not merely reject writing nor reduce it to a poison for cognition. He regards it as ambivalent: at once poison and remedy, a true “pharmakon for the mind.”
AI, for its part, is a “pharmakon” for the living. It is described as a “dual-use” technology: it carries the potential to create devastating poisons as well as nearly miraculous remedies. It is neither good nor bad in itself; one must grasp its nature to maximize benefits and minimize risks.
From the digital world to the physical world: a sudden jump
The 2026 news cycle was dominated by AI performances in cybersecurity, mathematics and literature—worrisome rather than cheering. To gauge its impact on biotechnology, one must recall a fundamental difference between this field and the ones it has already disrupted.
In biotechnology, fully exploiting AI will often require moving from the digital world, where it operates, to the physical world. Even the most precise algorithms cannot today account for all phenomena in the real world. (This is also the grand project of Yann LeCun: to develop AI capable of truly understanding the physical functioning of the real world.)
This leap is abrupt, hence a logistic choke point: conducting cutting-edge biotechnical experiments, rather than mere simulations, requires a fully equipped laboratory, qualified personnel, and thus substantial resources.
Mathematics, a “conceptual” science, allows solving problems with equations alone; biology and medicine, by contrast, are experimental sciences where one must produce, test and analyze.
The catalyst paradox: AI can have maximal impact only if the puzzle pieces are already in place
The impact of AI in any field directly depends on the knowledge, data, and tools already available there. In biotechnology, genomic data illustrate this clearly.
Crucial for training a cutting-edge biotechnology AI, these data are the product of a colossal scientific undertaking: the Human Genome Project. Made possible by centuries of scientific progress, this international collaboration stretched over more than a decade and culminated, at the turn of the millennium, in sequencing nearly the entire human genome. This achievement opened the door to “genome sequencing commoditization,” which accelerated throughout the 21st century following what some call Carlson’s curve, or the “genomic Moore’s law”: at regular intervals, the cost and time required to sequence a whole genome plummet. The number of freely accessible whole genomes on the Internet surged, with new genomes regularly published.
In chemistry, a catalyst is a compound, often protein-based in the biomedical realm, that significantly accelerates a chemical reaction without being consumed. If A reacts with B to produce C and D, and X is a catalyst for this reaction, then at the end the amounts of A and B have decreased and C and D have increased, while the amount of X remains unchanged. AI catalyzes human progress.
The more advanced a field, the better we can feed models with high-quality data, and the more sophisticated AI contributions can be. AI thus made its entrance as biotechnologies and medicine were already enjoying a golden age. Two fields in particular bear witness to this:
- induced pluripotent stem cells; this discovery, awarded the 2012 Nobel Prize in Medicine, rests on a simple principle. If you take any cell from the human body—skin cell, muscle cell, kidney cell, etc.—and inject four genes (Oct4, Myc, Klf4, Sox2), you transform it into a stem cell. This is what is called an “induced pluripotent stem cell” (iPSC): it can differentiate into almost any type of human cell and greatly broadens the possibilities in regenerative medicine.
- immunotherapy; the work underlying it was recognized with the 2018 Nobel Prize in Medicine. It has been discovered that, in a large number of cancers, the human immune system is naturally capable of recognizing, targeting, and destroying cancer cells. If it doesn’t do so, it is because cancer cells hijack molecular signals to masquerade as healthy, “innocent” cells to immune cells. Immunotherapy therefore involves developing in the laboratory antibodies directed against these signals, in order to “unblock” the immune system and enable it to destroy cancer cells.
Coupled with the explosion of sequencing and genome editing capabilities enabled by CRISPR, as well as advances in microfluidics and high-resolution microscopy, these two fields ushered medicine and biotechnology into a golden age, even before AI arrived. If AI’s potential is so great in biotechnology, it is because the reagents were already in place: it can now catalyze an immense number of reactions.
The making of novel genomes at Stanford: the turning point of 2026
On August 6, 2026, an article appeared about an experiment conducted at Stanford. Slightly commented upon, apart from concerns that “AI is now producing virus genomes,” it marks a major turning point in the history of biotechnology and medicine. What happened?
An AI specialist, Evo 2, developed by teams at Stanford and the Arc Institute around Brian Hie and Patrick Hsu, designed genomes which, when tested in the lab, proved to be “viable” at extremely high rates. Why is this a revolution?
To gauge the magnitude of this breakthrough, one must pause on the notion of a “viable genome.” A genome is a sequence of millions of letters—or billions in humans. A viable genome is a sequence of characters that, when introduced into a cell or organism, allows it to develop and live normally. In other words, it is a genome “compatible with life.” What is the proportion of such genomes?
For a given genome size, that proportion is extremely small. Why? Compare a genome to a book of a similar size. Suppose one writes it by randomly selecting each character—a letter, a digit, or punctuation. A “viable” genome would be analogous to a book written this way that somehow tells a coherent story. It would require that the letters form meaningful words in a single language often enough to form sentences, and that those sentences tell a coherent tale. The task is so demanding that the success rate is vanishingly small. To obtain a single coherent book by this method, one would have to write billions of such books and have someone read them all.
Since it is not necessary for every sentence to be coherent, evaluating a book requires reading a large portion of it. Producing all these books is extremely tedious; testing their coherence is no less so. Genome editing in biotechnology works in much the same way. One could, to some extent, view it as an adaptive landscape application of Maynard Smith and Émile Borel’s monkey-typing paradox in the space of genomic sequences.
Writing a viable genome is akin to writing a book in a language one barely knows. The best approach is to start from already published books, whose stories are known to be coherent, and to modify a letter here and there. This approach ensures producing stories that make sense, but they remain extremely close to what already exists. Similarly, in biotechnology, one progressively introduces changes into existing genomes. Why is that done?
For human genomes, it is by analyzing the effects of certain modifications that we understand how genes function and discover the genetic causes of diseases such as cystic fibrosis or sickle cell disease. It is also how we have understood the roles of genes involved in cancers, and how we can diagnose them earlier and more accurately.
For non-human genomes, the same approach has enabled us to understand the workings of many pathogens—viruses, bacteria, fungi, and parasites—and how they evade the immune system or antibiotics, leading to rapid advances in diagnostics and treatment.
Writing new viable genomes is therefore not an act of hubris, but a means of acquiring knowledge likely to save many lives.
CRISPR—the equivalent of a word-processing tool—revolutionized genome editing by allowing scientists to write viable genomes by typing the text of their choice. But they still did not know what to write: the minuscule probability of randomly obtaining a coherent text and the burden of testing remained.
Until this Stanford experiment. AI proposed a hundred novel genomes: the proportion of viable genomes, which should have been minuscule, turned out to be very high. It succeeded where human intellect failed, lacking the capacity to grasp the scale of genomes, their complexity, and the multiplicity of parameters governing life. This means that human scientists will now be able to rely on AI to produce viable genomes. While such a breakthrough can legitimately provoke worry, it should also inspire hope: the golden age of biotechnology and the progress already achieved, notably against cancers, existed before AI. It did not make headlines, but it will leave a lasting mark.
Biological weapons, pandemics and apocalypse: why extinction will not occur
The answer to apocalyptic talk about AI capable of manufacturing super-weapons and to the tech leaders who claim “10% odds of eradicating humanity” lies in three letters: HLA.
HLA, short for Human Leukocyte Antigen: this is a group of genes located on chromosome 6. These genes govern the regulation of the immune system. They are among the most polymorphic genes in the human species, meaning they vary the most from one individual to another. There are currently more than 35,000 “different versions” within the human species, a substantial diversity.
These genes determine the “target repertoire” that the host’s immune system can recognize and attack. They were first identified in the context of organ transplantation: two people with very similar—almost never identical—HLA profiles are far more likely to be compatible.
Yet their role extends beyond transplantation. Some types are associated with autoimmune diseases, while others correlate with markedly greater resistance to a given pathogen. This is why the extinction of humanity by a pathogen seems extremely unlikely. The diversity of HLA genes is such that in any human population there will always be a sizable portion of individuals whose immune system can mount a highly effective response against any pathogen. In science fiction films, these are the “immune” ones. And what if the virus mutates?
The virus can accumulate mutations that allow it to become invisible to the most common HLA-targeting systems and to replicate unimpeded in a large portion of the population. But there will remain minority HLA profiles for which these same mutations will have the opposite effect, making them particularly visible.
This brings us back to the idea of pharmakon and tinkering: in biology, everything is linked, and what protects against one threat can expose to another. Take the mutation responsible for sickle-cell disease, a disease of red blood cells. Homozygotes—carrying the mutation on both paternal and maternal chromosomes—develop the disease, while heterozygotes with one copy are relatively “protected” against malaria. Evolution has kept this mutation in malaria-endemic regions: depending on whether it is present once or twice, it protects against one disease or triggers another. This mutation perfectly illustrates the pharmakon concept: it carries the potential for poison and for cure.
In the same way, any mutation induced by a “super-bioweapon” would likely affect some HLA genes while increasing vulnerability to others. It’s an inextricable problem. Imagining a pathogen that escapes all HLA profiles on Earth would be like imagining a politician winning 100% of the vote in a truly democratic, secret, and sincere planetary election. Every stance, every silence would win some voters and cost others. He might approach a majority by aligning with the dominant positions, but would simultaneously alienate those whose views are minority.
All of this was understood and demonstrated by Peter Doherty and Rolf Zinkernagel, Nobel Prize in Medicine 1996. Their work was continued by Austin Hughes and Masatoshi Nei in a 1988 Nature article, and by the mathematical modeling of HLA allele dynamics proposed by Takahata and Nei in Genetics in 1990. The extraordinary progress in sequencing in the 21st century has allowed detailed analysis of HLA diversity and dynamics in human populations, confirming these hypotheses once again.
And if, to bypass this limit, an AI decided to use not a single weapon but a cocktail of biological weapons, with the idea that individuals resistant to one pathogen would be vulnerable to another? The calculations such an approach would require seem beyond reach today, but are unlikely to remain so in the medium to long term, given AI’s progress in mathematics and genetics in 2026.
One logistical problem would persist: “inoculating” a single outbreak is already hugely challenging. SARS-CoV-1 and MERS, two close relatives of SARS-CoV-2, failed to spark a pandemic. Likewise, influenza—despite its high variability, the reservoir of healthy carriers in aquatic birds and pigs, and the regular emergence of worrying strains—produced only one pandemic in the 21st century (2009-2010), and that one was not highly lethal. Triggering a pandemic is therefore far harder than it seems. But could one trigger several pandemics at once, with multiple pathogens?
That seems extremely difficult, and this is a “real-world problem” that no simulation or equation can fully solve. Several pathogens would compete for the same hosts, and their mutation dynamics could interfere with one another. Moreover, it would require an AI to design a set of biological agents covering all of the human species’ HLA gene diversity—more than 35,000 alleles—then produce them and trigger a pandemic such that every individual, wherever they are, encounters an agent able to bypass their defenses. The probability is almost nil.
Thus, the “HLA shield” makes the chance of humanity’s extinction by one or more pathogens nearly zero. But it is not “absolute,” and a devastating pandemic that bypasses innate immunity remains a real danger. How can we best guard against it?
Wastewater and molecular biology: the health semaphore
The best protection against a pandemic would be to maintain substantial investment in a state-of-the-art wastewater analysis system. Everyone who is infected sheds, even in tiny amounts, copies of the pathogen and its genome into the wastewater. Thanks to advances in sequencing and molecular biology, large investments could allow, in the short to medium term, to distinguish harmless viruses and bacteria from dangerous pathogens.
Such a system does not yet exist in a form capable of detecting and isolating a new pathogen present at low levels, and no surveillance system is ever infallible. But the golden age of producing and analyzing genomic data, opened by the Human Genome Project and amplified by AI, and perhaps tomorrow by quantum computing, makes the prospect of “health semaphores” that can identify discrete pathogen genomes in wastewater before circulation becomes visible seem achievable in a relatively near horizon.
These health semaphores would require close international scientific cooperation, with dedicated services around the world that publish their analyses and, in cases of detected potentially dangerous genomes, would investigate to determine their area of circulation.
Molecular tests functioning even in the absence of symptoms would be rapidly deployed to mass-screen and isolate contagious individuals. Many deaths would not be avoided, but transmission could be halted long before humanity is reached.
The bio boomerang paradox: why biological weapons are seldom used
It seems extremely unlikely that AI would cause humanity’s extinction. But what if a state used AI to develop a super-biological weapon capable of ravaging its enemies?
Such an undertaking would require substantial resources. Synthesizing a biological weapon calls for state-of-the-art laboratories, perfectly equipped benches, hoods, incubators, microscopes, all the reagents needed for cell cultures, and highly qualified personnel.
A peculiarly cynical state might bypass a high-security P4 lab and the safety measures necessary to protect its personnel by turning researchers into patient zero. But such personnel are difficult to replace, and such a replacement would handicap a development process that is already long. Moreover, even if a robot endowed with AI could perform bench manipulations precise enough to render staff—and perhaps even the P4—superfluous, it would be enormously expensive, and a fully equipped laboratory with all tools and reagents would remain necessary.
So, a state would need a certain level of resources to develop a biological weapon. Not on par with some drones that can be cobbled together and fitted with grenades or explosives, but still substantial.
Furthermore, a biological weapon is highly unpredictable, hence the notion of a “bio boomerang.” Collateral damage—even domestic—would be difficult to avoid if used. In the Middle Ages, the plague is said to have been partly imported into Europe by Mongol horsesmen who, besieging the Genoese outpost of Caffa in Crimea, hurled infected corpses over the walls. That outbreak delivered a brutal check to the Mongol expansion and marked the end of their golden age. Today the world is far more interconnected, so the military value of a biological weapon would be much more limited: the user could suffer as much harm as their enemies.
Moreover, biological weapons, like nuclear weapons, are a highly sensitive issue for the international community—precisely because the global risks of use are understood. Whoever develops or employs biological weapons risks drawing the attention of intelligence services worldwide and mobilization by the armed forces of the world’s major powers.
The powers with the resources to produce, test, and tune biological weapons have a great incentive to see the international community focus on them. Why develop such weapons when there are many other, powerful instruments that do not provoke as much upheaval? And while obtaining nuclear weapons can alter one’s status on the international stage, no country proclaims itself a “biological power.”
Finally, those who have “nothing left to lose” and would be willing to die to destroy humanity would not have the means to acquire biological weapons in secret or evade intelligence. A terrorist enterprise involving a biological weapon would be instantly perceived as a universal threat, especially after the Covid-19 pandemic, and denunciations would be numerous, quickly captured by human intelligence.
“Bio dividers”: the hybrid warfare biological weapon
Now imagine another kind of biological weapon whose aim is not to devastate a country, but to divide it, within a framework of hybrid warfare. An AI might design a highly contagious pathogen that causes visible symptoms—blemishes on the face, for instance—but slowly kills its host.
One could conceive a negative feedback mechanism: as the pathogen accumulates, it produces a signaling molecule that binds a regulatory protein and blocks its own replication, allowing its concentration to decline. The infection then becomes a long-term affliction.
What would be the point of such a pathogen, which has little chance of sparking an outbreak since its symptoms make contagious cases easy to identify? Infected individuals would remain contagious for a long time—around a decade—raising the question: what should be done with them?
In a logic of hybrid warfare, a state could use this pathogen to strike a minority community within the targeted country. Caregivers and scientists would then have no choice but to isolate symptomatic members of that community; yet months or years of isolation would breed tensions. The relatives of the sick might not tolerate the situation, while the targeted community would feel ostracized. Any individual seeking to break quarantine would likely be denounced, and a climate of paranoia would take hold.
The development of a treatment would feed the same tensions. The treatment would be costly and ostensibly not very profitable, sparking debates: should billions be spent to treat so few people?
Thus, one speaks of a “bio divider”: a weapon below conventional war but capable of provoking or intensifying tensions within a state. Genome analysis would reveal the artificial nature of the pathogen but would not identify the state responsible—much like many cyberattacks today that are identified as such but impossible to attribute.
The layered biosafety doctrine of defense
The extinction of humanity through super-biological weapons designed with AI is thus highly improbable, and a massive investment in sequencing and wastewater analysis could lead to a top-tier health semaphore. Yet pandemics capable of significantly altering life on Earth remain possible.
Wastewater analysis thus forms the “first layer” of a biosafety doctrine built on layering defense mechanisms. Each layer covers certain risks and leaves gaps that the others fill.
Early detection through wastewater helps buy time, but is only valuable if a system is in place to exploit that advantage. Upstream, screening of DNA synthesis orders provides another layer. Indeed, very few laboratories or hospitals synthesize the DNA they use themselves; this task is centralized and entrusted to a handful of specialized service providers that filter orders to ensure no sequence could be used illicitly. Downstream, defense relies on personal protective equipment, indoor air quality, rapid diagnostics, and vaccine platforms, among others.
Chemical weapons and the risk of “cognitive uplift”
Chemical weapons, when they are not forgotten, are often confused with biological weapons. Yet these two types of armaments share as much common ground as differences. Like biological weapons, chemical ones can kill a large number of people invisibly and with little means of defense, even in small quantities.
However, two major differences should be stressed. First, chemical weapons do not self-replicate and therefore cannot trigger an outbreak or an epidemic from a single localized use. Second, they are not “living” and therefore do not require the same logistics as cellular cultures: temperature control and cold-chain considerations. Chemical weapons are lethal, terrifying, far more predictable, and less complex to develop and handle.
For these reasons, unlike biological weapons, they hold genuine military interest. Their use is, in fact, more common: developed during World War I, and later employed by Saddam Hussein against Kurds or by Bashar al-Assad against Syrians. Once ethics and humanity are set aside, they allow rapid, mass killing in a horrific manner.
Because they are not living, there is no inherent potential to trigger a pandemic. However, AI could provide malicious actors with the “recipe” to synthesize one of the many chemical weapons already known, using everyday materials as feedstock.
Unfortunately, it is likely that AI could sequence reactions to synthesize chemical weapons—or even illicit drugs—from precursors that are not tightly controlled. In other words, AI could lower the resources required to produce and use these weapons.
Currently, chemical weapon precursors, as with many drugs, are tightly monitored. Anyone ordering these molecules is immediately flagged to intelligence and police services. For example, pseudoephedrine—a molecule found in cold medicines but also used to produce methamphetamine—receives close surveillance. (In February 2025, when Donald Trump signed the executive order 14,193 declaring a national border-wide “emergency” between the US and Canada, and then amended it on July 31, 2025, he mentions the flow of drugs and chemical precursors from Canada into the US. Without naming it directly, he targets pseudoephedrine imported from Canada to the US.)
Yet AI could compute an “indirect chemical pathway” enabling the synthesis of chemical weapons—or drugs—without relying on controlled precursors. It could almost certainly deduce chemical routes diverted in ways that they involve compounds too common to be surveilled. Imagine (the example is purely hypothetical) that it finds a way to produce a chemical weapon from household items as commonplace as consumer alcohol or acetone. Monitoring such compounds would be impossible, and in any case ineffective.
Thus, with chemical weapons we confront a double phenomenon: circumventing controlled precursors and “cognitive uplift,” which brings within reach the practical know-how that normally requires a high level of expertise.
No extinction, opportunities but also threats
AI is thus a pharmakon. Regarding the poisons it contains, it will probably not lead to super-biological weapons capable of wiping out the species, but it could facilitate the use of chemical weapons. We have not explored here the other paths by which it could precipitate humanity’s extinction.
As for the remedies it could bring, we have underlined that biotechnology already enjoyed a golden age before AI arrived and that, as a catalyst, it accelerates reactions made possible by the “reactants” already in place. The Stanford turning point of August 2026 testifies to this—an experiment that appears to provide proof of concept that humans will soon be able to leverage CRISPR to write and rewrite genomes at will. The prospects are, thus, dizzying.
What other frontiers of “beneficial” biotechnology could AI move in the near future?
Two major biotechnology questions where AI could substantially “help humanity”
AI could render humanity a considerable service by resolving two major open questions in biotechnology. If the intention is to “save the world,” tech giants could begin with these.
The differentiation of stem cells
In the mid-2000s, the Japanese scientist Shinya Yamanaka discovered a way to turn any human cell into a stem cell—allowing generation, for any patient, of stem cells fully compatible with their own body. But how does one “cultivate” these stem cells so they differentiate into specific tissues?
The goal of regenerative medicine is precisely to enable the reconstruction of different human tissues from stem cells, so as to replace what has been damaged or is malfunctioning. This requires stem cells that can become any cell type, but that must then differentiate into specific tissues. Because of the extreme complexity of this differentiation step, it remains one of the major bottlenecks to developing stem cell-based therapies.
Vertex, the American pharmaceutical company, recently succeeded in curing some patients with type 1 diabetes by replacing their insulin-secreting pancreatic cells with differentiated allogeneic stem cells. The company has formed a partnership with the French start-up TreeFrog Therapeutics to produce this treatment at scale.
Yet stem cells could also help treat many other pathologies and conditions. The “recipes” required for stem cell differentiation are, in many cases, three-dimensional in nature: to obtain certain tissues, stem cells must be grown in highly complex and specific three-dimensional environments.
AI could be mobilized to facilitate obtaining all these “recipes” needed for differentiation, thus enabling a leap forward in regenerative medicine.
Rubisco: the natural carbon trap
The enzyme known by its full name, ribulose-1,5-bisphosphate carboxylase/oxygenase—Rubisco for short—is the planet’s most abundant protein, and it plays a crucial role in photosynthesis. It helps harness the chemical energy produced by the capture of solar photons by the photosystems and chlorophyll to fix atmospheric CO2 (carboxylation).
Evolution selected Rubisco in a time when the environment—temperature, CO2 and O2 levels, and atmospheric composition—differed from today. In addition to performing the beneficial carboxylation reaction—making possible the production of carbon-based biomass essential to life on Earth—Rubisco also catalyzes a parasitic oxygenation reaction. This is unnecessary and reduces the enzyme’s efficiency.
Therefore, Rubisco could, in theory, be “optimized” to fix more carbon. One could imagine vast fields planted with species bearing these optimized Rubisco enzymes, with root structures designed to trap carbon for the long term.
Such carbon traps would be highly effective and, to a degree, “natural.” Given Rubisco’s abundance on Earth, even a moderate gain in efficiency could, in the medium term, have a substantial impact on the amount of CO2 in the atmosphere.
And if AI, which has also contributed to climate warming, could assist us in mitigating its effects?