Don't Make Mirror Life: Synthetic Biologist Kate Adamala on Risks & Responsibility
Summary
Kate Adamala’s core call is unusually narrow and categorical: keep developing synthetic biology and mirror therapeutics, but do not create a self-replicating mirror cell unless major new evidence changes the risk case. Mirror life’s practical attractions—immune-stealth medicines and contamination-resistant bioreactors—are precisely what make it dangerous: immune systems, viruses, and predators may not recognize or consume it. Labenz’s memorable framing is that “there’s nothing more dangerous than something that nothing can eat.”
The warning is early enough to matter because mirror life does not yet exist, but researchers were actively moving toward it. Adamala says most specialists expect a fully self-replicating mirror cell within a decade, while some say 20 years—not a distant, sun-burning-out scenario. Her intervention therefore targets a branch of the technology tree before researchers have the capacity to create it.
The downside is not necessarily a rapid takeover but an uncertain, potentially irreversible ecological imbalance. Adamala expects a mirror organism to begin as “not even a turtle…a slug,” extremely slow but potentially unstoppable if nothing can eat or infect it. Normal life might eventually evolve a defense, and she would personally bet on it, but “life not finding a way fast enough” could still be catastrophic. Containment cannot close the tail risk because accidents are unpredictable and a bad actor could deliberately remove engineered dependencies.
Synthetic biology has a large industrial thesis independent of mirror life: replace petrochemical production and gain operational control over cells, drugs, diagnostics, aging, and disease. Adamala stresses that oil supplies not just energy and plastics but electronics, clothes, medicine, cosmetics, food, and fertilizers; engineered organisms could make those chemicals instead of extracting them from “dead ferns underground.” Mirror life is one dispensable route among many, not a referendum on the sector.
The field remains far from fully understanding even its simplest systems, despite tangible progress toward bottom-up life. Craig Venter’s minimal living cell has 474 genes, including several dozen essential genes of unknown function; Adamala’s synthetic system has roughly 85–90, while George Church’s lab speculated that about 120 might sustain truly minimal life. The scientific and practical opportunity is paired with profound uncertainty: “We don’t even have a full ingredient list of biology.”
Adamala’s reversal offers a governance model for AI developers confronting newly reachable capabilities. Nathan Labenz argues that today’s AI resembles useful mirror molecules rather than full mirror life: intelligence exists, but robust autonomy, cyber escape, self-copying, and independent resource acquisition could mark a dangerous gain of function. His challenge is to stop treating every working capability as progress and ask “what kind of AI do we really wanna have?”
The coalition succeeded by inviting experts to disprove the risk case, then seeking community norms, funding restrictions, and regulation together. Ecologists, immunologists, engineers, and policy specialists broadened the analysis but could not identify a persuasive safety argument; the proposed posture was “carrot rather than a stick,” preserving researchers’ broader work while closing one door. Adamala warns that this becomes much harder once a technology “already pays them” and billions of dollars are exposed.
Deep dive
1. Life’s visible diversity rests on surprisingly narrow chemistry
Adamala’s starting point is that life is chemically “extremely boring”: every known organism builds proteins from only 22 proteinogenic amino acids and DNA or RNA from five nucleobases, despite hundreds or thousands of naturally possible alternatives. Evolution produced enormous diversity of form and function from an unusually restricted molecular toolkit.
That restriction motivates synthetic biology’s expansive agenda: give living systems chemicals “that natural life doesn’t bother using,” then test what new functions evolve and what those results reveal about life’s possible chemistries. The field is not limited to adding one or two amino acids; it asks how much of biology’s design space terrestrial evolution simply left untouched.
Asked how large that space might be, Adamala gives an honest boundary: “The limiting factor in all our theories is imagination,” because the sample size is N=1—modern terrestrial life. Speculation includes silicon-based or non-carbon life under extreme temperatures and pressures, but her own work remains within carbon-based, water-based systems whose possibilities are already vast.
Even an unmodified natural ribosome can incorporate hundreds of amino acids when removed from cells and “gently encourage[d]” in vitro. It can also connect substrates that are not amino acids into polymers, implying that today’s demonstrated range is “definitely not the ceiling” even before researchers evolve the ribosome or invent other molecular assemblers.
2. Life may have originated repeatedly before one lineage consumed the field
Adamala’s personal guess, shared by many colleagues, is that life began far more than once. It appeared almost immediately on geological timescales once Earth had crust and liquid water, suggesting that origin-of-life chemistry under suitable conditions “cannot be that difficult” and may have occurred in many places at roughly the same time.
Those early lineages probably drew from the same molecular pool and competed for the same resources; the ancestor of modern life may simply have eaten or outcompeted the others. New origins could theoretically occur even “in 2025” at a hydrothermal vent, but nascent life would be so fragile that established organisms would consume it before it gained traction: “Once you have established life on a planet, then good luck trying to start again.”
Early chemistry likely began promiscuously, using whatever the prebiotic environment supplied, then narrowed as membrane-bound cells became responsible for sourcing or synthesizing their own ingredients. Adamala compares the constraint to nomads carrying only a backpack, versus a settled homeowner with “a basement full of crap”: mobility and compartmentalization reward simplicity; established biosynthesis later permits expansion.
3. The ribosome’s power comes from being an ancient, indiscriminate machine
Adamala rejects the intuition that the ribosome must be sophisticated because it accepts so many substrates: “Ribosome is actually extremely dumb.” In biochemical terms it is an entropy trap, not an elegant catalyst; its “business model” is to take molecules and “smash them together” into a polymer.
That crudeness may reflect a frozen evolutionary accident. The first organism able to synthesize proteins gained an enormous advantage, but subsequent changes risked lethally breaking an indispensable system—like ancient internet infrastructure or Adamala’s five-year-old router that nobody dares update because losing connectivity would be worse than keeping obsolete machinery.
Specificity is mostly physical: a charged tRNA must fit into an opening, and the growing protein must pass through the ribosome’s exit tunnel. Oversized amino acids can jam that tunnel, but the ribosome does not inspect a molecule and declare, “You’re unnatural, so I won’t translate you.” The more selective gatekeepers are the enzymes that attach amino acids to tRNA.
Labenz connects this to optimization: the ribosome occupies a workable local minimum, while a better global design may exist elsewhere. Adamala says biologists use precisely that language—the route to another minimum requires climbing a fitness hill, and “that hill is lethal.”
4. Definitions of life and intelligence fail at their own boundaries
Adamala’s functional definition of life borrows Justice Potter Stewart’s phrase: “I will know it when I see it.” NASA’s working definition—a self-replicating chemical system capable of Darwinian evolution—sounds rigorous until she notes that she cannot reproduce without her husband and does not undergo Darwinian evolution alone, yet is plainly alive.
Her strongest positive criterion is emergence: life must be “more than a sum of its parts.” Nothing in that framing requires DNA, proteins, molecules, or even a body, so a sufficiently complex computational system with emergent properties might cross the threshold. The difficulty is epistemic: without a complete definition, observers may find it difficult to settle whether AI has arrived there.
Adamala applies the same objection to claims that intelligence evolved independently twice. Her dog distinguishes a chip bag from similar shopping-bag sounds whenever food might fall, despite being “otherwise…pretty dumb”; whether that counts depends on the definition. She expects many independently evolved flavors of intelligence, all so far supported by terrestrial biology’s single chassis—unless AI supplies the first foreign one.
5. Synthetic biology aims to replace oil and make biology programmable
Adamala defines synthetic biology as engineering life beyond its evolved chemistry and behavior. Natural cells resist programs that diverge from their survival interests; the field’s ultimate ambition is “full operational control”—understanding every molecular part well enough to program living systems to perform functions that evolution never selected.
The industrial case begins with civilization’s reliance on petrochemicals. Renewable power does not remove the need for oil because petrochemical inputs underpin electronics, clothes, medicines, cosmetics, plastics, food production, and fertilizers. Abandoning them would imply returning to pre-industrial living; continuing to extract them destabilizes the climate.
Her proposed alternative is biological manufacturing: persuade engineered organisms to synthesize chemicals now obtained from “dead ferns underground.” Because many target compounds are toxic, “no self-respecting bacterium” would willingly produce them for us unless researchers redesign the cellular chassis so those products no longer poison the factory itself.
The second major prize is health. Better control of human cells could clarify metabolic disease, cancers, and aging while improving health span rather than merely extending sick years. Adamala invokes Richard Feynman’s “What I cannot build, I cannot understand”: constructing living systems is both a test of biological knowledge and a route to better diagnostics and drugs.
6. Biology still lacks an ingredient list, let alone a causal model
Asked how much of the cellular causal graph has been mapped, Adamala answers: “Really not much.” Researchers know a graph exists and recognize some major nodes, but they cannot trace every molecule, establish all causal relationships, or enumerate every chemical in a single human cell. “We don’t even have a full ingredient list of biology.”
That gap is “kind of embarrassing” after roughly 50 years of modern work, yet it defines the frontier: biologists are still measuring the extent of cellular complexity before they can explain it. This is true for single cells; tissues and complex organisms remain “miles away.”
The top-down response is Craig Venter’s Mycoplasma-based minimal cell, whose 474-gene genome can be chemically synthesized. It is substantially more interpretable than ordinary cells, but several dozen essential genes have unknown functions: deleting one kills the organism, while researchers cannot explain why. Small molecules and their relationships add further layers to the black box.
Bottom-up researchers “are basically cheating,” Adamala jokes: rather than spend centuries deciphering a natural cell, they combine purified or synthesized components they already understand. Their goal is to add metabolism, replication, environmental survival, and interactions until the result becomes undeniably alive while every component remains legible.
7. Bottom-up cells are nearing plausible complexity but cannot yet evolve
Adamala’s laboratory works with a synthetic cell containing about 85–90 genes, versus 474 in the minimal living cell. A George Church lab paper from roughly 10 years earlier speculated that a truly minimal independent system might need about 120, placing bottom-up constructs near the right order of magnitude even though the remaining functional gap is unresolved.
The 474-to-120 difference may be another local-minimum problem. A redesigned organism could outsource membrane production or eliminate complex post-transcriptional modification pathways, while the existing minimal cell cannot lose them without dying. Unknown essential genes could also reveal that the 120-gene estimate overlooks indispensable functions.
There is no Frankenstein moment. Researchers pipette clear liquids into a tube, wait, then measure proteins and metabolized small molecules; membrane formation follows lipid chemistry at the right pH, and polymerases transcribe appropriate promoters because “that’s what you evolved to do.” Adamala expects life’s original emergence was similarly gradual: “You blink, you miss it.”
Synthetic systems already form membranes, transcribe, translate, metabolize, feed, grow, and replicate genomes. Many biologists nevertheless await Darwinian evolution before calling them alive. Physicists may accept today’s systems; biologists reply, “No, this is too wimpy.”
8. Evolution requires an error rate that synthetic cells cannot yet tune
The present bottleneck is perversely good replication. Adamala’s bacteriophage-derived polymerase is “annoyingly high fidelity,” making too few mistakes to create selectable variation. Yet simply breaking it produces too many errors, causing an “error catastrophe” in which descendants lose a viable genome rather than evolve.
The target is a narrow band: enough mutations to diversify offspring, but few enough that most remain functional. Available enzymes tend to offer either high fidelity or library-generating error rates, neither suitable for Adamala’s unusually small construct. Engineering a protein to be “just a little bit broken” is much harder than wrecking it.
Her numerical example makes the scaling problem concrete. At one error per 100,000 bases, a natural genome of several million bases accumulates several mutations per replication; her roughly 90,000-base genome gets zero or one, generally too few. Nature supplies no obvious enzyme optimized for such a small genome because no organism in nature is simple enough to have a genome that small.
9. Mirror life turns molecular chirality into a new biological chassis
Chiral molecules can exist in two non-superimposable forms, like left and right hands: chemically similar and visibly mirrored, but unable to occupy the same configuration. Mirror life would replace normal DNA, RNA, proteins, and other chiral components with their opposite-handed versions while preserving comparable chemistry.
A complete mirror cell would be compelling foundational science—the first organism outside the biochemical tree shared by known life. It could test whether the same functions emerge when familiar molecular components are rebuilt with opposite stereochemistry.
The practical motivations were more immediate. Mirror therapeutics might circulate without immune attack, while mirror-cell bioreactors might resist viruses and ordinary microbial contamination, protecting an entire “million dollars’ worth” of production from a sneeze or stray phage.
Consultation with ecologists and immunologists inverted that value proposition. Invisibility to immune systems, viruses, and predators is not merely a feature; it removes the mechanisms that keep organisms from proliferating unchecked. Adamala describes her shift as a “rollercoaster of emotions”: from “Yay, mirror life” to recognizing that its risks outweigh its applications.
10. Molecular evidence supports immune stealth even without a mirror cell
No whole mirror organism has been made, “thank God,” so researchers cannot infect a mouse and directly observe the response. The organism-level risk is therefore inferential—but it rests on experiments with opposite-chirality proteins that do evade normal immune activation.
Mirror-image nucleic acids called spigomers, from the German word for mirror, are already in clinical trials precisely because they are immune-stealthy. Specialists infer that if mirror proteins and nucleic acids escape recognition, an organism constructed from those components would also pass under immune defenses.
Adamala distinguishes detection from activation: white blood cells may “sniff” a mirror molecule, but they do not treat it as a threat. Her analogy is a dog investigating bread versus a rock; both are noticed, only one triggers action. A mirror cell could consequently find the bloodstream “a warm little pond,” replicating without the response that ordinarily clears foreign biology.
11. An organism that nothing can eat breaks ecology’s normal checks
Labenz’s maxim is that “there’s nothing more dangerous than something that nothing can eat.” Adamala’s reply is immediate: “That’s what mirror life is.” Ordinary predators, parasites, viruses, and immune systems depend on stereochemical compatibility; reverse that interface and a mirror organism may register as inert matter rather than food or infection.
The closest historical analogy is the great oxygenation event: cyanobacteria transformed the atmosphere with a toxic waste product, triggering extinctions that harmed even the organisms responsible. Adamala knows no later example of a new organism possessing absolutely no predator, virus, or replication constraint.
Humans are her nearest modern comparison. Having reached the top of the food chain, they reproduce without meaningful natural predators and reshape the planet, although disease and self-inflicted harms still impose limits. Mirror life could lack even those controls, violating the ecological balance in which “everything in nature tries to eat everything else.”
Adamala does not predict a rapid takeover. A released organism would be “not even a turtle…it would be a slug”: extremely slow, but potentially unstoppable if nothing could eat or infect it. She would personally bet that normal life eventually finds a way, but refuses to rely on evolution to do so quickly enough.
12. Distance, containment, and competition do not close the tail risk
Adamala calls remoteness the strongest skeptical argument: why worry now when mirror life resembles the eventual death of the sun? Her answer is that researchers “were actively working on making the damn thing,” and specialists broadly expect a self-replicating mirror cell within a decade; some say 20 years. Silence would leave the attractive applications to recruit another team.
Physical containment cannot be absolute: “If you make a foolproof container, then nature will develop a better fool.” Twenty-seven locks do not anticipate the hole in the floor, and no laboratory design excludes theft or intentional release. If an organism must never escape, Adamala concludes that the only dependable safeguard is not to possess it.
Biological containment could make a mirror cell dependent on perhaps 50 rare chiral inputs, preventing it from surviving even if poured onto a lawn. But an attacker might engineer around those dependencies, or researchers could build metabolism around achiral carbon sources already used by some bacteria. Jurassic Park’s lysine dependency, she jokes, was weak enough for a first-year bioengineering student to reject.
Adamala would personally bet that ordinary life eventually evolves a way to eat or infect the invader. She still refuses the wager: no opposite-chirality life has existed in the environment to select such defenses, evolution is stochastic rather than programmable, and the cost depends on whether defenses arise quickly enough. “It might or it might not. That’s the problem.”
13. AI now faces its own choice between useful components and dangerous autonomy
Adamala initially distinguishes the fields sharply: mirror life does not exist and can still be stopped, while AI is already useful and “out of the box.” She has not written code in over a year because ChatGPT drafts her scripts and often fixes them after she pastes back the error; biotechnology and medical discovery already use similar tools.
Labenz’s pushback is that AI may be only partially built, just as mirror molecules exist without mirror organisms. Current models display intelligence, but robust autonomy, cyber escape, self-copying, acquiring money, and paying for independent compute could combine into a qualitatively less controllable chassis.
His broader concern is regime change: AI research once asked whether anything could distinguish a cat from a dog, so almost any working technique counted as progress. Now “everything is working,” exposing “a million doors” while critical analysis of which capabilities society may regret remains scarce.
Adamala identifies one biological gain-of-function threshold: a model that synthesizes the literature into genuinely new instructions for a better virus, pathogen, toxin, or drug. Safeguards might exist, but she notes that most models can be jailbroken; the dangerous change would be possessing knowledge that even top virologists cannot presently provide with confidence.
14. Germline editing tests where to draw the line around irreversible technologies
Labenz finds mirror life comparatively easy to reject because medicine offers many other doors. Human germline editing presents a harder bargain: a few edits might sharply reduce Alzheimer’s, heart disease, or rare genetic illness, making the next generation healthier even if later generations face dystopian pressures.
Adamala’s fear is not that edited embryos escape like organisms, but that parental incentives make the social slope unstoppable. “If I could make my kid…the smartest, fastest, healthiest one in the world, I totally would,” she says; even while opposing the technology, she doubts she could reject an edit lowering her own child’s Alzheimer’s risk.
Unequal access could create biologically distinct social tiers within one or two generations, amplifying existing disparities in nutrition and development into inherited advantages. Once the capability exists, she expects people to use it, leaving universal distribution as the fallback—despite society being “really suck[y] at distributing things equally.”
Labenz counters with the genome-sequencing cost curve, from roughly $1 million during his college years to a few hundred dollars, and notes that leading AI models have often reached free users. Adamala concedes the possibility but calls it a leap of faith: adoption must begin before scale lowers costs, while biological procedures retain labor and material constraints.
15. Coalition-building worked because researchers were asked to prove the warning wrong
Adamala’s group expanded beyond mirror-life engineers to ecologists, immunologists, and policy experts because no discipline could assess the whole system. Many prospective authors had never considered mirror life; the organizers showed the evidence and asked, “Can you please prove us wrong?” They sincerely wanted a safe route to remain open.
The search changed and broadened the technical report, but nobody who studied the full case found a persuasive refutation or withdrew in disagreement. Adamala’s invitation remains open: evidence that mirror life is safe would be welcome because the researchers still consider it a fascinating project—they simply “don’t think there is a way to do it safely right now.”
Useful countermeasure work may include testing mirror-image antibiotics and clarifying whether achiral food sources can support the organism’s metabolism. Yet Adamala warns against allowing safety research to become a covert development path: “You don’t wanna get too close to actually making the thing that you don’t wanna make.”
Six months after publication is “nothing” in policy time, but her desired mechanism has three layers: a professional taboo, denial of funding, and formal prohibition. The strategy is “carrot rather than a stick”—community buy-in before top-down rules, with the reassurance that nobody is taking science’s toys away, only keeping play safe. She says this was easier because mirror life was not yet an existing technology or therapeutic with billions of dollars at stake.