God Plays Dice, and Evolution Loads Them

Einstein's conviction that chance could not be fundamental lost every escape route physics could offer it. Evolution never needed him right: mutation is proven, irreducibly stochastic, and selection turns that noise into every improbable structure alive.

A fluctuation-test set: hundreds of independent cultures grown side by side so that the rare, disproportionate ones would stand out — the design built to prove mutation happens before anything selects it [@luria-delbruck-1943]. — Image prompt and art direction by Brecht Corbeel; generation pending.

Abstract

In December 1926 Einstein told Max Born that he was convinced "He does not throw dice" — an aesthetic conviction about physics that the twentieth century's loophole-free Bell tests would spend decades quietly closing off. What Einstein actually contested, in the 1935 Einstein-Podolsky-Rosen paper, was completeness rather than chance itself; Bell's 1964 theorem turned that dispute into a testable inequality, and the 2015 triple of loophole-free experiments left local hidden variables with nowhere left to hide. Biology ran the opposite bet the entire time, and won it on evidence: the Luria-Delbrück fluctuation test proved bacterial mutations arise before selection can see them, Lederberg's replica plating showed it directly, and modern sequencing has turned the human mutation rate into a measured constant. This piece traces both verdicts side by side — physics's century-long failure to rescue determinism, and biology's construction of its most improbable structures from admittedly blind chance — while keeping the analogy honest about where it breaks: quantum randomness appears fundamental, mutational randomness is a practical consequence of chemistry too intricate to track.

A Postcard From December 1926 Still Defines the Argument

On 4 December 1926, Einstein wrote to Max Born, who had just published the statistical reading of the wavefunction that would become quantum mechanics’ probability rule. The letter survives in Irene Born’s 1971 translation, reprinted in the 2005 edition of The Born-Einstein Letters: “Quantum mechanics is certainly imposing. But an inner voice tells me that it is not yet the real thing. The theory says a lot, but does not really bring us any closer to the secret of the ‘old one’. I, at any rate, am convinced that He does not throw dice” [1]. Worth being precise here: the poster-caption version, “God does not play dice with the universe,” appears nowhere in the letter. It is a later compression, not a quotation, and the small misattribution is worth flagging even though the substance survives the paraphrase. What the letter actually records is the opening statement of an argument Einstein pursued, in increasingly technical form, for the rest of his life — and unlike the slogan, the argument can be checked.

Incompleteness, Not Chance, Was the Actual Complaint

Nine years later, Einstein, Boris Podolsky, and Nathan Rosen turned the objection into a paper rather than an aphorism: “Can Quantum-Mechanical Description of Physical Reality Be Considered Complete?” [2]. The argument disputes no number quantum mechanics predicts. It proposes a criterion for physical reality — a quantity whose value can be predicted with certainty without disturbing the system is real, whether or not the theory assigns it a definite value — then constructs a pair of particles correlated closely enough that measuring one lets an observer predict a property of the other without touching it. Quantum mechanics says that property was not definite until measured; EPR’s criterion says it must have been real all along. Their conclusion: quantum mechanics, however accurate, was an incomplete description of a reality with definite properties throughout.

That is what the 1926 letter compresses. Not that chance appears in the predictions, but that a complete theory would not need chance, because a finer-grained deterministic layer would sit beneath the statistics. Einstein’s target was completeness; chance was only the symptom he thought pointed to its absence.

The Loophole Closes; the Interpretation Does Not

EPR’s paper stayed a philosophical argument until 1964, when John Bell showed it could be turned into an experiment [3]. Bell proved that any theory obeying both locality — no faster-than-light influence — and realism, in the sense of outcomes fixed by pre-existing local properties, must satisfy a statistical bound. For a simple two-setting test,

S = \left| E(a,b) - E(a,b') + E(a',b) + E(a',b') \right| \le 2

holds for any local realist theory, while quantum mechanics allows S up to 2\sqrt{2} \approx 2.83. Early experimental violations left the detection loophole (too many unmeasured particles to rule out a biased subsample) and the locality loophole (stations not separated fast and far enough to exclude a hidden signal) open to a committed skeptic.

2015 closed both, three times over. Hensen and colleagues at Delft entangled electron spins in diamond, 1.3 kilometres apart, and measured S = 2.42 \pm 0.20 against the bound of 2, with detection efficient enough and separation wide enough to close both loopholes together [4]. Giustina and colleagues in Vienna, using entangled photons and high-efficiency detectors, reported a violation whose probability under local realism was no greater than 3.74 \times 10^{-31} [5]; Shalm and colleagues at NIST, independently, reported p-values as low as 5.9 \times 10^{-9} [6]. Between the three, no known loophole remained.

What that establishes is precise, and what it leaves open deserves equal precision. No theory simultaneously local and realist can reproduce nature’s measured correlations; that does not, by itself, prove nature indeterministic. Determinism survives only by paying elsewhere — dropping locality, as in Bohmian nonlocal hidden variables, or dropping single-outcome realism, as in Everettian many-worlds, where every outcome occurs somewhere and apparent randomness is only a fact about which branch an observer occupies. Both reproduce every observed statistic by construction, which is why no experiment adjudicates between them: the loophole-free tests killed local realism specifically, not the broader question of determinism versus chance. Einstein’s conviction was not simply wrong; it was wrong in a way that closed off the plainest escape route without closing every escape a determined theory could still take.

Selection Never Sees the Mutation That Made It

Biology settled its own version of this question decades earlier, by experiment rather than inequality. In 1943, Salvador Luria and Max Delbrück asked whether bacteria surviving exposure to a virus become resistant because of that exposure, or were already resistant beforehand for unrelated, random reasons [7]. Their fluctuation test compared samples drawn from one large bulk culture of Escherichia coli against the full contents of many small, independently grown cultures. Induced resistance would make both sampling methods show similar, Poisson-like variation, since each culture’s fate would hinge only on its final encounter with the virus. Pre-existing, randomly timed mutation would make the independent cultures behave very differently from each other: a culture where the mutation struck early would multiply the mutant clone for many generations before plating, producing a disproportionate “jackpot” of resistant colonies, while a late or mutation-free culture would show few or none.

The data broke sharply toward the second picture. Bulk-culture samples showed variance close to their mean, as a Poisson-distributed process predicts; independent cultures showed variance far exceeding the mean — in one reported series, a mean of roughly 11.3 resistant colonies against a variance above 750, the fingerprint of a few early, lucky mutations dominating the count. From the same data Luria and Delbrück derived one of the first quantitative bacterial mutation rates, on the order of 3 \times 10^{-8} per cell per division. Resistance arises before the phage ever appears, at a rate independent of the phage’s presence; selection does not cause the variant, it only notices which variants already exist — a result built on counting colonies and variance arithmetic, work that helped earn Delbrück, Alfred Hershey, and Luria the 1969 Nobel Prize in Physiology or Medicine.

Figure 1. A "jackpot" plate among near-empty sister plates: the pattern Luria and Delbrück used to show that a few cultures had inherited an early, random mutation, while most had not [@luria-delbruck-1943]. — Image prompt and art direction by Brecht Corbeel; generation pending.

Replica Plating Turns the Statistics Into a Photograph

A fluctuation test is a statistical argument: it shows pre-existing mutation is the more probable explanation, but it never points to a single colony and says “this one was already resistant.” Joshua and Esther Lederberg closed that gap in 1952 with replica plating [8]. A velvet-covered stamp, pressed onto a master plate of colonies and then onto a series of fresh plates in turn, transfers each colony’s exact position to every replica. Exposing only the replicas to a selective agent — a phage, an antibiotic — and never touching the master, the Lederbergs could identify which coordinates turned resistant on an exposed replica, then return to the untouched master and pick the colony at that same coordinate. That colony had never met the selecting agent, and it was already resistant: a direct demonstration, not an inference, that the mutants exposure will appear to “produce” were already there.

Figure 2. Lederberg's replica-plating jig stamped one plate's exact colony pattern onto fresh plates — letting a resistant colony be traced back to a spot on a master plate that had never met the selecting agent [@lederberg-1952]. — Image prompt and art direction by Brecht Corbeel; generation pending.

The Rate Is Now a Measured Number, Not an Inference

Where Luria and Delbrück inferred a mutation rate from the shape of a variance distribution, modern sequencing measures it directly. Kong and colleagues sequenced whole genomes of 78 Icelandic parent-offspring trios and counted de novo mutations — present in a child, absent from both parents — one nucleotide at a time [9]. At an average paternal age of 29.7 years, they estimated a genome-wide human rate of 1.20 \times 10^{-8} point mutations per nucleotide per generation. Mutations were overwhelmingly paternal in origin, rising by roughly two per year of paternal age and doubling every sixteen and a half years; the maternal contribution, around fourteen per generation, showed little age dependence. Eighty years after a fluctuation test inferred a rate from a handful of jackpot cultures, the same quantity is a routinely sequenced parameter of human reproduction.

Directed Mutation Almost Overturned the Picture, and Then Mostly Didn’t

The pre-existence principle met its most serious challenge in 1988, when John Cairns, Julie Overbaugh, and Stephan Miller reported that starving lactose-negative E. coli, plated with lactose as the only carbon source, accumulated lactose-metabolizing revertants at a rate that appeared to depend on continued exposure to the very sugar the mutation would let them use [10]. That looked like mutation responding to selection rather than preceding it, in direct tension with Luria and Delbrück, and it triggered immediate rebuttal in the following year’s Nature and a controversy running well over a decade.

The resolution is more interesting than either side’s original framing. The phenomenon is real and reproducible: starved cells under continued selection do show elevated, late-appearing reversion. But the traced mechanism is not mutation aimed at a useful target — it is a general, genome-wide rise in mutation rate under stress, driven by error-prone repair such as the SOS response, pushing a fraction of the starving population through a transient hypermutable state. Cells in that state mutate indiscriminately across the whole genome; a few, by chance, pick up the one change that lets them use lactose, and only those cells then grow to be seen. No mutation is chosen; the odds of a useful one appearing somewhere in the population are simply raised.

That reframing left its own live dispute: whether the capacity to raise the mutation rate under stress is itself an evolved adaptation. Foster’s review treats regulated hypermutation as a candidate mechanism for evolvability, favored by second-order selection for generating variation when a population is poorly adapted [14]. MacLean, Torres-Barceló, and Moxon argue the evidence for a dedicated “evolvability” adaptation is weaker than that framing suggests, and that the same observations fit stress-induced mutagenesis being a side effect of pleiotropic stress responses, maintained by selection for stress tolerance and by drift rather than by selection for future adaptability [15]. Neither side disputes that the mutations are unaimed; contested is only whether the dial setting how fast the dice get thrown was itself put there on purpose.

Figure 3. A UV-crosslinker delivers a measured, artificial mutagen dose used as a control against spontaneous background mutation — a reminder that even deliberately "induced" mutagenesis is a dial on a known rate, not a hand choosing which base to change. — Image prompt and art direction by Brecht Corbeel; generation pending.

Loaded, Not Aimed: What the Ratchet Actually Does

None of this makes any individual mutation less blind; it changes the picture of what selection has to work with. Fisher’s argument that adaptation proceeds mainly through mutations of small individual effect, rather than rare large leaps, generalizes the Luria-Delbrück lesson to the whole history of life: no single throw needs to be improbable, because a ratchet that retains favorable throws and discards the rest compounds an enormous number of small, blind events into a structure that looks, from outside, impossible by chance. The dice stay honest. The game is not.

The more contested nuance is that the dice may not be evenly weighted across a genome. Monroe and colleagues reported that in Arabidopsis thaliana, mutation frequency runs roughly half as high inside gene bodies as elsewhere, and lower still in essential genes, tracking epigenomic marks — H3K4me1 among them — that appear to recruit DNA-repair machinery toward functionally important sequence [11]. If correct, the mutation-generating process would itself have been shaped by selection to protect the sequence an unaimed change is most likely to harm — tuned randomness in a more literal sense than the stress-induced case.

That claim did not go unchallenged. Wang, Ho, Hurst, and Yang re-examined Monroe’s mutation calls, found only 3.7 percent concordance with a higher-confidence reanalysis, traced most of the discrepancy to known base-calling artifacts near homopolymer runs, and concluded the reported trends “are not robustly substantiated” [12]. Monroe and colleagues replied that the artifact rate their critics identified is, if anything, higher rather than lower inside gene bodies — the opposite of what would explain the pattern — and cited an independent meta-analysis of over ten thousand germline mutations reproducing the same reduction [13]. The disagreement is unresolved in the published literature and should be reported that way: a live methodological dispute over whether mutation rate is itself an evolved, unevenly distributed trait, not a settled fact either way. Either way, the claim on the table is narrower than “directed mutation”: nobody proposes a cell chooses which base to change. Contested is only whether the rate of undirected mutation has been sculpted, by selection acting on repair machinery, to fall unevenly across a genome that cannot see which of its own positions are about to matter.

Two Verdicts From the Same Century

Physics spent the twentieth century testing Einstein’s conviction that a complete theory should not need chance, and closed off, one loophole at a time, every practical way that conviction could still be true: Bell turned the question into an inequality, and the 2015 triple left local hidden variables nowhere to hide. Biology, over the same century, ran the opposite experiment without ever doubting its premise, building from admittedly unaimed variation — proven blind by Luria and Delbrück, proven blind again more directly by Lederberg, now measured to a specific rate per nucleotide — eyes, immune systems, and everything else alive, using nothing but differential survival to keep whatever the dice produced that worked.

The analogy should not be pushed past where it holds. Quantum randomness, so far as the loophole-free tests can show, is fundamental: no finer deterministic layer survives beneath it without sacrificing locality or single-outcome realism, a larger price than physics wanted to pay for Einstein’s intuition. Mutational randomness is not fundamental in that sense. Which base a polymerase mis-copies, which photon a UV dose happens to strike, which repair enzyme reaches a lesion first — each is, in principle, an ordinary chemical event with a deterministic cause, simply too intricate for any cell or observer to predict in advance. It behaves as randomness relative to selection without necessarily being randomness the way a decaying nucleus is. Calling both “dice” captures something true about how each feeds a downstream filter, but only one has been shown, by direct experiment, to be unloadable even in principle; the other only looks unloadable because no one, including the organism doing the mutating, can compute the throw ahead of time.

Einstein’s aesthetic objection was wrong about the system he meant it for. Applied to the system he never had in mind, it is close to exactly right: the dice are thrown blind, generation after generation, and the game selection runs on top of them is still rigged toward order.

Sources

1. Albert Einstein. Albert Einstein to Max Born, 4 December 1926. The Born-Einstein Letters 1916-1955 (Macmillan, 2005), trans. Irene Born; archival letter reproduced by Physics Today (1926).
2. Albert Einstein, Boris Podolsky, and Nathan Rosen. Can Quantum-Mechanical Description of Physical Reality Be Considered Complete?. Physical Review (American Physical Society) (1935). DOI: 10.1103/PhysRev.47.777.
3. J. S. Bell. On the Einstein Podolsky Rosen Paradox. Physics Physique Fizika (1964). DOI: 10.1103/PhysicsPhysiqueFizika.1.195.
4. B. Hensen, H. Bernien, A. E. Dréau, and colleagues. Loophole-free Bell inequality violation using electron spins separated by 1.3 kilometres. Nature (preprint archived on arXiv) (2015).
5. Marissa Giustina, Marijn A. M. Versteegh, Sören Wengerowsky, and colleagues. Significant-Loophole-Free Test of Bell's Theorem with Entangled Photons. Physical Review Letters (American Physical Society) (2015). DOI: 10.1103/PhysRevLett.115.250401.
6. Lynden K. Shalm, Evan Meyer-Scott, Bradley G. Christensen, and colleagues. Strong Loophole-Free Test of Local Realism. Physical Review Letters (American Physical Society) (2015). DOI: 10.1103/PhysRevLett.115.250402.
7. Salvador E. Luria and Max Delbrück. Mutations of Bacteria from Virus Sensitivity to Virus Resistance. Genetics (1943). DOI: 10.1093/genetics/28.6.491.
8. Joshua Lederberg and Esther M. Lederberg. Replica Plating and Indirect Selection of Bacterial Mutants. Journal of Bacteriology (1952). DOI: 10.1128/jb.63.3.399-406.1952.
9. Augustine Kong and colleagues. Rate of de novo mutations and the importance of father's age to disease risk. Nature (2012). DOI: 10.1038/nature11396.
10. John Cairns, Julie Overbaugh, and Stephan Miller. The Origin of Mutants. Nature (1988). DOI: 10.1038/335142a0.
11. J. Grey Monroe, Thanvi Srikant, Pablo Carbonell-Bejerano, and colleagues. Mutation bias reflects natural selection in Arabidopsis thaliana. Nature (2022). DOI: 10.1038/s41586-021-04269-6.
12. Long Wang, Alexander T. Ho, Laurence D. Hurst, and Sihai Yang. Re-evaluating evidence for adaptive mutation rate variation. Nature (2023). DOI: 10.1038/s41586-023-06314-y.
13. J. Grey Monroe, Kevin D. Murray, Wenfei Xian, and colleagues. Reply to: Re-evaluating evidence for adaptive mutation rate variation. Nature (2023). DOI: 10.1038/s41586-023-06315-x.
14. Patricia L. Foster. Stress-Induced Mutagenesis in Bacteria. Critical Reviews in Biochemistry and Molecular Biology (2007). DOI: 10.1080/10409230701648494.
15. R. Craig MacLean, Clara Torres-Barceló, and Richard Moxon. Evaluating evolutionary models of stress-induced mutagenesis in bacteria. Nature Reviews Genetics (2013). DOI: 10.1038/nrg3415.

Originally published at https://absolutedigitalpublishers.com/articles/god-plays-dice-and-evolution-loads-them.