Life in the UniverseExplainer
How Does Evolution Actually Work?
Small inherited differences, filtered over countless generations, add up to the variety of life. Here is how evolution works and how we know.
THE QUESTION LIBRARY LIFE IN THE UNIVERSE
If we could replay the history of life, would evolution turn out the same?
Evolution mixes chance with patterns that repeat, so here is what flasks of bacteria, eyes that evolved more than once, and a famous debate reveal about how predictable life really is.
THE SHORT ANSWER
Partly. If life’s history could be rerun, many details would very likely differ, because chance mutations and rare events steer each path. But evolution also repeats itself. Bacteria in separate flasks often evolve the same kinds of fixes, and eyes capable of true vision arose separately several times in different animal groups. So evolution mixes chance with repeatable patterns. How similar a rerun living world would be in its big picture is still debated among scientists. Separate populations test repeatability; simulations explore assumptions; plans recorded before outcomes help distinguish predictions from discoveries made afterward.
- In 1989, the scientist Stephen Jay Gould argued that replaying the “tape of life” would give a very different living world.
- The Cambridge scientist Simon Conway Morris argues the opposite: that evolution often reaches the same solutions again and again.
- In Richard Lenski’s experiment, 12 flasks of initially identical E. coli bacteria, started in 1988, often evolved in parallel.
- By the time of a 2008 report, only one of the 12 populations had evolved the ability to use citrate (by generation 31,500), and it depended on earlier changes.
- Some mutations simply happen more often than others, and a 2026 study found this can shape which fixes win, even in large populations.
On this page
About this picture
The 12 populations of Richard Lenski’s long-term evolution experiment on 25 June 2008. The cloudy flask, Ara-3 (labeled A-3), holds the population that evolved to use citrate, so it grows far more cells than the others.
WHERE THE EVIDENCE STANDS
How sure are we?
- Evolution involves both chance events and repeatable patterns.A 2018 review in Science by Blount, Lenski and Losos: natural selection is a deterministic force, but it works on variation from random mutation, and replay experiments often show parallel changes alongside one-of-a-kind outcomes.
- Separate populations facing the same challenge often evolve similar changes.In Lenski’s experiment, all 12 populations improved in fitness and grew faster on their sugar, and at least three genes changed in all 12 (Blount, Borland and Lenski, PNAS 2008).
- In one Lenski population, the ability to use citrate evolved only after earlier changes made it possible.Replays from frozen samples re-evolved the trait only from later-generation clones, pointing to a potentiating change by generation 20,000 (PNAS 2008); genome work traced the trait to a gene duplication (Nature 2012).
- A replay of all of life would give a broadly similar living world, perhaps including beings like us.Nobody can replay all of life. Blount, Lenski and Losos (Science, 2018) find evolution surprisingly repeatable among close relatives, but disparate outcomes become more likely as history runs deeper; Conway Morris argues for strong predictability and himself calls the idea heterodox.
- Mutation bias, meaning some mutations happen more often, helps decide which adaptations win.Barber and Couce (Nature Communications, 2026) found the effect need not fade in large E. coli populations and can persist or even grow, depending on how the bias lines up with selection; they note the idea is still controversial.
- A computer model alone confirms how real bacteria will evolve.Its rules are assumptions; predictions need checks against living populations.
WHAT WE DON’T KNOW YET
Nobody knows how often key steps in life’s history, such as the rise of complex cells or of minds like ours, would happen again in a replay. It is also hard to count how often convergence happens across all of nature.
WHAT WOULD CHANGE THIS ANSWER
More replay experiments in many kinds of living things, run for longer and started from more distant points in history, would sharpen this answer. Finding life elsewhere, with its own separate history, would give the first test of evolution on another world.
Partly. If we could rewind the history of life and press play again, many details would very likely come out differently. But not everything. Evolution keeps finding some of the same solutions, again and again.
The question has a famous history. In 1989, the scientist Stephen Jay Gould wrote a book called Wonderful Life. He argued that evolution is full of chance and quirky turns. If we replayed the “tape of life” from the distant past, he said, the living world would look far different from today’s. Simon Conway Morris, an Earth scientist at the University of Cambridge, argued back. He pointed to convergent evolution, when unrelated living things evolve the same kind of trait on their own. In his view, life takes many routes but reaches only a few destinations.
We cannot rewind all of life. But we can replay small pieces of it. In 1988, the biologist Richard Lenski started 12 flasks of the same bacteria, E. coli, and let them evolve side by side. By 2008 they had passed 40,000 generations. Samples are frozen along the way, so the team can go back in time and run history again.
The results show both sides. All 12 populations got better at living in their flasks, often through similar changes. But only one population evolved the ability to eat a food called citrate, after more than 30,000 generations, as a 2008 paper reported. Replays showed that it needed earlier changes first. That is chance and history at work.
Nature shows the same mix. Eyes capable of true vision arose separately several times in different animal groups. Sharp, high-resolution eyes come in different designs, such as camera eyes and compound eyes.
So the plain answer is that evolution mixes chance with patterns that repeat. That much is Established. Our overall answer, “partly,” is Supported, still debated. Whether a full replay would give a broadly similar world, including anything like us, is an Open question.
THE LONG ANSWER
Free long answer 1 of 3 this month.
What did Gould mean by replaying the tape?
Stephen Jay Gould made the idea famous in his 1989 book Wonderful Life: The Burgess Shale and the Nature of History. A 2021 article in BioScience sums up his claim: if we rewound the tape of life and replayed it from the Cambrian, a time more than 500 million years ago, we would be unlikely to end up with anything like humans. A 2008 paper by Zachary Blount, Christina Borland and Richard Lenski in the journal PNAS explains his reasoning. Evolution depends on long chains of earlier events, and small changes in them can have huge effects later. So, Gould argued, replaying the tape of life from some point in the distant past would give a living world far different from the one we see today. Scientists call this contingency: when an outcome depends on the particular history that came before.
The same paper gives the other side. Simon Conway Morris, of the University of Cambridge’s Department of Earth Sciences, argued that natural selection pushes living things toward a fairly small number of good designs. His summary: the routes are many, but the destinations are limited. In a 2010 essay for the Royal Society, he went further. He wrote that evolution is remarkably predictable, and that some major steps in life’s history were very probable, if not inevitable. He also called this a heterodox idea, meaning it goes against the usual view.
Neither side says evolution is pure chance or pure destiny. The argument is about the balance.
How do scientists replay evolution in a lab?
Lenski’s long-term evolution experiment started in 1988 with 12 populations of E. coli, all founded from the same starting strain. They live in a liquid with a little glucose, a sugar, as food. Each day, a small sample of each is moved into a fresh flask. The bacteria grow about 6.6 generations a day, so the experiment moves through thousands of generations each year. By 2008 it had passed 40,000 generations. Every so often, samples are frozen. Those frozen samples are a living fossil record that can be thawed and run again.
The 12 flasks show a lot of parallel change. According to the 2008 paper, all 12 populations improved in fitness, quickly at first and then more slowly. All evolved faster growth on glucose and bigger cells. At least three genes changed in all 12 populations. That is the repeatable side of evolution, and it answers one of the most common questions: bacteria in separate flasks often evolve the very same fix because they face the very same challenge.
What did the citrate story show?
The flasks also contain citrate, a compound E. coli normally cannot use as food when oxygen is present. For more than 30,000 generations, no population learned to eat it, even though each one tried out billions of mutations. Then, by generation 31,500, one population, called Ara-3, evolved the trait. The team noticed it after about 33,127 generations, when that flask turned much cloudier with extra cells.
Was this just one rare lucky mutation? The team tested it by replaying history. They thawed clones from different points in Ara-3’s past and started 72 new replay populations, which evolved for about 3,700 generations. The citrate trait evolved again in four of them, and all four were founded from later-generation clones. The researchers concluded that some earlier change, which arose by generation 20,000, made the trait much easier to reach. They call this a potentiating mutation.
A 2012 paper in Nature by Blount and colleagues read the genomes. The trait arose when a stretch of DNA was duplicated, which switched on a citrate transporter gene that had been silent. The genome work also found that several earlier potentiating mutations arose during the population’s history. So the citrate story is both: a new trait that became reachable only in a population whose history had set the stage.
Why does a mutation test need separate populations?
One population gives one history. Separate populations can take different paths in the same environment, because mutations occur in different orders. Comparing populations reveals how often an outcome appears and how much results vary. Many cells from one population do not provide many separate evolutionary histories.
A 2010 experiment started six populations per food regime. Testing evolved bacteria against their ancestor under matched conditions distinguished inherited improvement from temporary adjustment to the surroundings.
What can a computer simulation tell us?
It shows what follows from chosen rules for reproduction, survival and population size. Those assumptions matter: the authors of SLiM 3, an evolution simulator, describe how simplifying life cycles can limit which situations a model represents.
The 2026 mutation-bias study used a simple model to explore possibilities, then tested predictions in real bacterial populations. Running a simulation is useful; checking its predictions against living things supplies the biological test. This does not mean a model captures all of evolution.
What does preregistered mean?
Researchers record their question, planned measurements and analysis before learning the outcomes, usually in a registry. This separates an advance prediction from an explanation built after seeing results. Plans can change; researchers should report what changed and why. Preregistration does not guarantee a sound experiment or a correct idea.
Existing data can still be useful, but researchers must say what they already knew. A pattern discovered while examining data is exploratory; a later test can fix its plan before examining fresh outcomes.
Why did eyes evolve again and again?
Outside the lab, scientists look for convergent evolution. Eyes are a classic case. In a 2013 review, Dan-E. Nilsson of the Lund Vision Group at Lund University in Sweden concluded that the move from simply sensing light toward true vision happened several times on its own in different animal groups. Eyes built for sharp, high-resolution vision evolved only in vertebrates (animals with backbones), cephalopods (such as octopuses and squid) and arthropods (such as insects and crabs). Some are camera-type eyes, like ours. Others are compound eyes, like an insect’s, made of many small units. These are different solutions to the same problem.
There is a twist. A gene called Pax-6 helps control how eyes develop in both vertebrates and many invertebrates. Some scientists took that as a sign that eyes evolved only once. Nilsson argues that his picture of several separate origins, built from simple light-sensing cells that early animals already shared, fits with that shared gene too. Why do eyes keep evolving? In Nilsson’s account, animals kept evolving more demanding behaviors guided by sight, and those behaviors pushed eyes to perform better.
What is the strongest objection?
The strongest objection to “evolution would repeat” is that we have never replayed the big picture. A 2018 review in Science by Blount, Lenski and Losos weighed the evidence. They found that evolution tends to be surprisingly repeatable among closely related lineages. But different outcomes become more likely the deeper history runs. They also note that it is hard to measure how often convergence happens in nature.
A newer piece of the puzzle is mutation bias: some mutations simply happen more often than others. A 2026 study in Nature Communications by J. N. Barber and A. Couce used two strains of E. coli with high, lopsided mutation rates that evolved resistance to the antibiotic cefotaxime. The usual view is that mutation bias matters less in large populations. They found it can fade in some cases, but it can also persist and even grow stronger in large populations, depending on how well it lines up with what selection favors. They describe a view in which evolution favors not just the fittest, but “the likeliest among the fittest.” They also write that the idea is still controversial.
What do people often get wrong?
“Evolution is random.” Mutations arise by chance, but natural selection is not random. The 2018 review calls selection a deterministic force, meaning it pushes steadily in a direction.
“Repeated traits mean everything was bound to happen.” Repeats show that some solutions are likely. They do not show that every step, or every species, would come back.
“Lenski’s bacteria prove Gould right,” or “prove him wrong.” They support parts of both views: lots of parallel change, and one key trait that depended on history.
How can you check this yourself?
- Read the 2008 paper. The Blount, Borland and Lenski paper is free on PubMed Central, and its opening pages explain Gould and Conway Morris in plain terms.
- Look at the flasks. The photo on this page shows the 12 populations in 2008. One flask is cloudier than the rest.
- Ask “how related?” When you hear that evolution repeated itself, ask how closely related the living things were. Close relatives repeat more often.
- Read more. Our story How Does Evolution Actually Work? covers the basics, and our answer on whether life is common in the universe asks what this means for life elsewhere.
The answer stays the same after all this detail. Evolution mixes chance with repeatable patterns. How alike a replayed world would be is still debated.
How is the author's "natural dominance" idea different from "survival of the fittest"?
Author’s hypothesis. Ricardo’s natural-dominance proposal compares how often competing mutation routes occur and when they take hold. A route here means a possible genetic change. The software description for version 2.6.5, dated 5 September 2026, calls it a proposed framework. It considers the supply of new mutations, their biological accessibility, whether they establish, and the setting.
Its question is which routes we could predict, and how often. For the standard meaning of “survival of the fittest,” and how fitness, luck and timing relate, see our fitness and evolution answer. TMD still needs to show whether its models predict better than appropriate alternatives.
Does the natural dominance idea challenge Darwin, or build on his work?
Author’s hypothesis. The relationship stated in the Research Notebook, reviewed 1 October 2026, is narrower: TMD has not established a new law of evolution or replaced natural selection. It proposes tests of mutation-route models. That leaves a research question to investigate; it does not establish that Darwin’s account has been overturned. The cited passages do not settle the author’s precise historical comparison with Darwin.
Why does the author call his evolution idea a hypothesis, not a proven theory?
Author’s hypothesis. A title containing “Theory” is not a biological test. The version 0.8 description, dated 12 October 2025, used “pre‑theory” and made the full name conditional on passing checks specified in advance. Version 2.6.5 describes a proposed framework and leaves improved prediction to independent comparisons planned before new outcomes are examined.
The methods report, version 0.5.0, dated 30 September 2026 and posted 1 October, claims no new laboratory experiment or matched biological validation. The cited passages support a proposal awaiting such testing, not a claim that its biological predictions have passed those tests.
What experimental result would count against the natural dominance idea?
Author’s hypothesis. The Research Notebook proposes a broad comparison: an outside lab fixes its test in advance, then TMD’s picks do no better than a standard model of mutation bias plus selection. That would count against the proposed extra predictive benefit. It is the Notebook’s suggested criterion, not a completed experiment.
The version 0.5.0 report describes a narrower condition. After its recovery correction, if the ranges from different conditions share no common value, that challenges the “proposed common contrast” under its stated assumptions. Here, the contrast is a comparison between routes. If the ranges overlap, the question remains unsettled.
This conditional result concerns a shared comparison under those assumptions. It does not establish rejection of every TMD claim or identify a biological mechanism. The cited passages do not supply a full laboratory rejection protocol for the whole framework. Follow the Test Ledger for the Notebook’s dated test status.
THREE THINGS TO REMEMBER
- Evolution mixes chance with patterns that repeat.
- Close relatives facing the same challenge often evolve the same fixes.
- How predictable life’s big picture is remains debated.
WHERE THE AUTHOR’S RESEARCH TOUCHES THIS
Author’s hypothesis. The Research Notebook’s Evolution & mutation page describes a research program by Ricardo Maldonado, author of GOD PLAYS DICE, called the Theory of Mutation by Natural Dominance (TMD). It starts from the pattern on this page: when many populations of the same microbe face the same challenge, such as an antibiotic, they often evolve similar fixes. It asks whether the “winning” mutations follow patterns we could predict, and how much the setting matters.
Its newest software version, 2.6.5, is dated 5 September 2026 and was posted on Zenodo on 30 September 2026. It is for analysis only. Its tools include a check that a lab runs first, checks on results sent in from outside, and a log of every step that others can replay. Version 2.6.4 compares the proposed patterns with competing explanations based on known mechanisms. Passing software checks does not show the pattern happens in nature.
On the same day, Ricardo posted a separate report. It says its aim is “to make predictions that can fail fairly on new data.” It corrects how earlier work used some counts, and it takes a new look at data that other scientists published in 2021. It calls that look “exploratory, not preregistered.” Because it reuses published data, it is not a test fixed in advance. Neither version 2.6.5 nor the report describes a new lab experiment, and TMD does not establish a new law of evolution or replace natural selection.
The Research Notebook’s status, reviewed 30 September 2026, is: software and workflow release; no new lab test reported. Its strongest challenge is the mainstream work on mutation bias described above: to count, TMD would have to pick the winners better than those models, on new data. You can follow what would count against it on the Test Ledger.
Keep asking
Nobody knows yet.
New here?
God Plays Dice answers big questions in plain words, with sources and one of six honest labels.
Words worth knowing (6)
- Contingency
- When an outcome depends on the particular history that came before it, so small early events can change what happens later.
- Convergent evolution
- When unrelated living things evolve similar traits on their own, such as eyes in several animal groups.
- Parallel evolution
- When closely related populations, such as Lenski’s 12 flasks, evolve the same changes separately.
- Potentiating mutation
- An earlier change that does not create a new trait by itself but makes it much easier for a later change to produce it.
- Mutation bias
- The fact that some kinds of mutations happen more often than others.
- Generation
- One round of reproduction. For Lenski’s bacteria, about 6.6 generations pass each day.
Sources & further reading
(13) · Checked Oct 1, 2026
- Nothing in Evolution Makes Sense Except in the Light of Biology (Reiskind, Moody, Bolnick, Hanifin and Farrior) ↗BioScience 71(4), 370–382 (2021); open access via PubMed Central PMC8038875Stephen Jay Gould argued in Wonderful Life: The Burgess Shale and the Nature of History (1989) that evolution would not repeat itself; if we rewound the tape of life and replayed it from the Cambrian, we would be unlikely to end up with anything like humans.
- Historical contingency and the evolution of a key innovation in an experimental population of Escherichia coli (Blount, Borland and Lenski) ↗PNAS 105, 7899–7906 (2008), doi:10.1073/pnas.0803151105Gould, Wonderful Life (Norton, 1989) and the tape of life; Conway Morris: “the evolutionary routes are many, but the destinations are limited”; LTEE started in 1988 with 12 populations; about 6.64 generations per day; parallel fitness gains, faster growth, larger cells, at least three genes changed in all 12; no citrate use for over 30,000 generations despite billions of mutations; Cit+ by 31,500 generations, noticed after about 33,127; 72 replay populations for about 3,700 generations, Cit+ re-evolved in four, all from later clones; potentiating mutation by 20,000 generations.
- Genomic analysis of a key innovation in an experimental Escherichia coli population (Blount, Barrick, Davidson and Lenski) ↗Nature 489, 513–518 (2012)Abstract (read via Europe PMC, PMC3461117): the Cit+ trait arose by a tandem duplication that captured an aerobically expressed promoter for a previously silent citrate transporter; multiple potentiating mutations arose during the population’s history.
- Contingency and determinism in evolution: Replaying life’s tape (Blount, Lenski and Losos) ↗Science 362, eaam5979 (2018)Abstract (read via Europe PMC): selection is deterministic but acts on random mutation and drift; replay experiments often show parallel changes plus idiosyncratic outcomes; convergence is hard to quantify; evolution surprisingly repeatable among close relatives, disparate outcomes more likely as history grows deeper.
- Evolution: like any other science it is predictable (Simon Conway Morris) ↗Philosophical Transactions of the Royal Society B 365, 133–145 (2010)Open-access full text: Department of Earth Sciences, University of Cambridge; argues evolution is remarkably predictable and calls the idea heterodox; key transitions very probable, if not inevitable; cites Lenski’s experiment as the best approximation of re-running the tape.
- Eye evolution and its functional basis (Dan-E. Nilsson) ↗Visual Neuroscience 30, 5–20 (2013)Open-access full text (Lund Vision Group, Lund University): directional photoreception proceeded several times independently to true vision; high-resolution (class IV) eyes only in vertebrates, cephalopods and arthropods, of camera and compound type; eye evolution driven by increasingly demanding visually guided behavior; shared Pax-6 expression has been taken as evidence eyes evolved once, and the multiple-origin scenario is compatible with it.
- Mutation-biased adaptation is consequential even in large bacterial populations (J. N. Barber and A. Couce) ↗Nature Communications 17, 7357 (published 10 June 2026)Open access, published 10 June 2026: the idea that mutational tendencies influence adaptation is “still controversial”; two E. coli mutator lineages evolving cefotaxime resistance; mutation-biased adaptation “may not wane in large populations” and can persist or intensify depending on its alignment with selection; describes a view of “the likeliest among the fittest”.
- Experimental evolution with E. coli in diverse resource environments. I. Fluctuating environments promote divergence of replicate populations ↗Cooper and Lenski, BMC Evolutionary Biology 10:11 (2010)Methods; Discussion.
- SLiM 3: Forward Genetic Simulations Beyond the Wright–Fisher Model ↗Haller and Messer, Molecular Biology and Evolution 36:632–637 (2019)Introduction: model assumptions.
- The preregistration revolution ↗Nosek and colleagues, PNAS 115:2600–2606 (2018)Preregistration in Practice, Challenges 1, 3, 5 and 6.
- Theory of Mutation by Natural Dominance (TMD) — v2.6.5: portable campaign execution and stricter result acceptance (DOI 10.5281/zenodo.22398093) ↗Ricardo Maldonado; author’s research record on ZenodoDated 5 September 2026; posted 30 September. The saved description passages call this a proposed framework for comparing mutation routes and leave improved prediction to independent comparisons planned in advance. Cached 30 September 2026; complete descriptions and attached files were not assessed here.
- Mutation by Natural Dominance (TMD): a testable framework for evolution and the origin of life — software & data (v0.8) (DOI 10.5281/zenodo.17335715) ↗Ricardo Maldonado; author’s research record on ZenodoDated 12 October 2025. The saved description passage calls this a framework before theory status and makes the full name conditional on passing checks specified in advance. Cached 30 September 2026; this passage does not establish that those checks passed. Complete descriptions and attached files were not assessed here.
- Mutation by Natural Dominance: finite-sample bounds for recovery-corrected route comparisons (report, version 0.5.0; DOI 10.5281/zenodo.23075312) ↗Ricardo Maldonado; author’s research record on ZenodoDated 30 September 2026; posted 1 October. The saved description passages give a conditional common-route-comparison test under stated assumptions; overlap remains unsettled. The report claims no new laboratory experiment or matched biological validation. Cached 1 October 2026; complete descriptions and attached files were not assessed here.
See a mistake? Report a problem. Corrections are made openly.