by Ashutosh Jogalekar
In 1966, Michael Polanyi began his book The Tacit Dimension with a sentence that every experimental scientist understands: “we can know more than we can tell.” Polanyi had been a chemist before he became a philosopher. He knew the difference between the written account of an experiment and the experience of doing it. His point was not that science is vague or mystical, it was that scientific judgment rests on inherited practice, trained perception, and habits learned from other people. The University of Chicago Press summary of The Tacit Dimension describes tacit knowledge – tradition, inherited practices, implied values, and prejudgments – as a crucial part of scientific knowledge.
The point is easy to see in chemistry. A synthetic procedure may ask the chemist to warm a solution with “gentle swirling,” to add a reagent slowly enough to maintain a “gently exothermic reaction,” or to continue stirring until the mixture turns “pale yellow.” These are ordinary phrases. They are also compact pieces of apprenticeship. A young chemist wants to know how gentle is gentle, how pale is pale, and how much warmth counts as a controlled exotherm. The experienced chemist has seen enough flasks to supply the answer. Even Organic Syntheses, perhaps the most careful journal of synthetic procedures, acknowledges the difficulty. Its instructions say that its procedures are written in greater detail than usual journal procedures, checked for reproducibility in an editor’s laboratory, and still may cause trouble when checkers try to reproduce the submitters’ results.
There are foundational examples in the old seminal papers. Alexander Fleming’s 1929 paper on penicillin begins with an accident that would have annoyed most bacteriologists. He had set aside plates of staphylococci on a laboratory bench. During repeated examinations they were exposed to air and became contaminated. Around one contaminating mold colony, he noticed that the staphylococcal colonies “became transparent and were obviously undergoing lysis.” A spoiled plate became a discovery because Fleming knew what he was seeing. The later paper contains numbers, dilutions, temperatures, and culture conditions. The first act was visual judgment. It was the recognition that a nuisance had become a phenomenon to be investigated.
A second example comes from a very different part of science. Robert Millikan’s 1913 paper on the elementary electric charge is remembered as a monument of precision. It gave the electron charge a number. Yet the method depended on a person watching oil drops through an optical system and knowing when the observation could be trusted. Millikan worried that an oil drop might drift during a long observation; if the drift occurred along the line of sight, it “could in no way be noticed by the observer.” His improved apparatus kept the drop “continually in sharp focus,” so that a motion of half a millimeter blurred the image. The number depended on voltage, viscosity, gravity, and Stokes’s law. It also depended on an observer who knew what a trustworthy drop looked like.
Synthetic biology now gives this old problem a sharper edge. Addgene’s protocol for amplifying CRISPR libraries moves easily between exact numbers and the language of craft. It tells the user to add a specified amount of DNA, “flick gently to mix,” avoid repeated pipetting, spread cells without ripping the agar, and later scrape bacteria while avoiding bubbles and gouging. A cell-culture guide gives the same kind of instruction in another setting. Healthy cells should be “round and plump,” culture medium should be “pinky orange,” and cells may be split at roughly 70–80 percent confluence. These are signals. They tell a trained person whether living material is behaving well.
This is where large language models become interesting. Their importance for tacit knowledge is not simply that they retrieve information. Retrieval is the old trick. The new trick is comparison at scale. Imagine a model reading a hundred papers in which a protocol says “stir until turbid.” In one set of papers, turbidity may precede high yield. In another, it may precede poor purity. Across the larger literature the model may correlate that phrase with solvent, concentration, temperature, reaction time, particle size, filtration method, and later words such as “milky,” “opalescent,” “fine suspension,” or “persistent emulsion.” It may also correlate turbidity with a dozen other tacit indicators: color, viscosity, foaming, pellet appearance, colony size, media color, or cell density.
This is a plausible path by which AI can illuminate the novice researcher. The model can make a vague phrase less vague by placing it among similar phrases and outcomes. It can say, in effect, that in this part of the literature the word “turbid” usually travels with one kind of result, while in a neighboring part it travels with another. Recent chemistry work already points in this direction: researchers have fine-tuned LLMs to extract structured reaction information from prose descriptions of organic synthesis procedures, with the goal of supporting tasks such as reaction prediction and condition recommendation. I have made the broader point in a WMD context: LLMs can mine tacit knowledge by comparing vague laboratory phrases such as “gently shake” or “straw yellow” across many experiments.
For ordinary science, this is a wonderful prospect. A student without a famous adviser can learn faster. A small laboratory can avoid old mistakes. A researcher far from the traditional centers of science can gain access to practical wisdom that once moved mainly through elite networks. The history of science is full of such moments. Printing widened access to texts. Journals widened access to results. Photographs, videos, preprints, open-source software, and public databases widened access again. AI may widen access to the part of science that has always been hardest to write down.
The unease begins when this same thought is applied to weapons of mass destruction. Dangerous technologies have never depended on information alone. Nuclear, chemical, and biological weapons require materials, instruments, space, measurements, organization, secrecy, and people who can keep a chain of operations from falling apart. Biology is especially difficult because living systems misbehave. They mutate, contaminate, die, drift, and respond to conditions in ways that defeat confident plans. This is why intent and expertise have often lived in different worlds. Those with the strongest desire to do harm usually lack the craft. Those with the craft usually work inside institutions with colleagues, rules, reputations, and supply chains, wanting to do good rather than harm.
Biosecurity scholars have long stressed this point. James Revill and Catherine Jefferson argue that tacit knowledge is often marginalized in discussions of biological weapons programs, weakening our sense of how hard such programs are. They note that successful biological weapons work would require know-how across acquisition, handling, culturing, storage, scale-up, and delivery; they also describe tacit knowledge as something commonly acquired through “learning by example” and “learning by doing.” Their paper gives a vivid laboratory example: the art of “douncing” cells, where force, speed, number of strokes, and feel matter in ways that are hard to learn from a paper. Even automated lab systems, they add, require tacit knowledge to work consistently.
AI therefore raises a real dual-use problem. It may not turn a novice into an expert. It can still move the novice some distance down the road. It can explain why a protocol failed. It can identify which omitted condition matters. It can gather fragments that used to take weeks to find. In good hands this saves time and money. In bad hands it can reduce one of the barriers that has protected us.
I think that my use of mosaic theory is a useful lens here. In my Fast Company essay, I recall the case of John Aristotle Phillips, the Princeton undergraduate who in the 1970s designed a crude atomic bomb using public sources, while Freeman Dyson advised him without supplying classified information. The unsettling part, in Dyson’s telling, was the speed with which Phillips collected the material. My lesson is that public fragments can become sensitive when arranged in the right order, and that LLMs are very good at arranging fragments. In this setting, tacit knowledge is part of the mosaic. It may be hidden in adjectives, lab notes, captions, troubleshooting comments, protocols, and failed experiments. A model that compares enough of these traces can sometimes reconstruct the practical meaning.
This is why the recent open letter signed by Demis Hassabis, Sam Altman, Dario Amodei and others deserves a central place in the argument. The letter calls for mandatory screening and record keeping for synthetic nucleic acid orders and for equipment used to make them. It begins by acknowledging the good: ordering synthetic DNA online has accelerated vaccine development, powered basic research, and given small teams access to capabilities that once belonged mainly to large institutions. It then states the concern plainly: AI systems may meaningfully erode knowledge barriers that have historically kept bad actors from obtaining biological weapons.
Natasha Bajema and Mara Zarka’s VCDNP report supplies the calmer half of the argument. Their paper examines AI and WMD over the next decade and concludes that the AI-WMD nexus “does not yet constitute a revolution,” although it deserves careful attention. The report emphasizes capability, intent, WMD development pathways, and the places where barriers persist. Bajema’s own summary of the work is especially useful. AI is impressive in digital space, she writes, and less impressive in the laboratory and the real world. WMD work still needs controlled materials, specialized facilities, tacit knowledge, scarce data, and enough expertise to judge whether AI outputs are reliable. In her opinion, AI is currently a force multiplier for those who already know what they are doing, rather than an equalizer for novices.
That is reassuring for exactly the right reason. The lab still supplies friction: Cells must live. Cultures must stay clean. Instruments must be calibrated. Materials must be obtained. Bad advice must be recognized. The physical world remains an old and stubborn regulator of ambition. A language model can describe an experiment, but it cannot yet make the experiment work.
Bajema also names the threshold worth watching: models that can accurately simulate complex biological or physical systems, and AI-driven robots able to perform delicate laboratory operations autonomously, including pipetting, mixing, and culturing. This is where the problem changes. Today’s LLM may help decode the tacit dimension from text. A future laboratory system could learn the tacit dimension by running experiments, recording failures, and adjusting conditions. For medicine, materials, agriculture, and climate science, that would be a powerful development. For WMD safety, it would move the main concern from knowledge access to operational access.
The right policy follows from this distinction. Broad scientific explanation should remain widely available. The chokepoints should sit around operationalization: procurement, troubleshooting, scale-up, evasion, automated execution, and the supply chains where matter begins to move. AI systems should be evaluated over extended conversations, because the risk often lies in the direction of travel rather than in one sentence. DNA synthesis screening, customer verification, material controls, biosafety training, expert red-teaming, and record keeping all belong in the same picture. They are dull tools, but dull tools often keep civilization alive.
The larger point should not be lost in the risk analysis. Tacit knowledge is one of the great goods of science. Making more of it visible can help young scientists, poor laboratories, neglected diseases, and fields where expertise is scarce. The opportunity cost of excessive caution may be very large. If AI models are trained to refuse too much practical scientific guidance, the losses will not appear as one dramatic event. They will appear as slower diagnoses, failed experiments, weaker public-health responses, delayed drugs, and small laboratories left without help. The World Health Organization already describes AI as relevant to diagnosis and clinical care, drug development, disease surveillance, outbreak response, and health-system management. These are not marginal uses. They touch ordinary human suffering on a very large scale. A rare biosecurity event may kill many people and must be guarded against. But a policy that withholds too much tacit knowledge from millions of legitimate users may cost more lives, and may leave many more lives unimproved. This is the irony of the Precautionary Principle in this case. As Cass Sunstein has argued, strong versions of the principle can become paralyzing because every choice, including inaction, carries risks. Here, inaction has a cost measured in missed cures, slower science, and preventable disease.
AI may help spread judgment beyond the communities that once carried it. The task is to preserve that gift while keeping destructive capability difficult to assemble. For now, the physical laboratory remains a brake. In the future, as AI gains better hands, the safeguards will have to move closer to the hands.