Neurogenesis: The Fight Over Whether Adult Brains Grow New Neurons

When Two Laboratories Read the Same Brain and Disagree

In March 2018, a team led by Shawn Sorrells at the University of California, San Francisco published a paper in Nature with a title that left no room for hedging. Human hippocampal neurogenesis, they wrote, drops sharply in children to undetectable levels in adults [1]. They had examined 59 human brains spanning fourteen gestational weeks to seventy-seven years of age. After roughly age thirteen, they found nothing.

Six weeks later, Maura Boldrini at Columbia University published in Cell Stem Cell. Same organ. Same question. Opposite answer. Human hippocampal neurogenesis, her team reported, persists throughout aging [2]. They counted thousands of immature neurons in people in their seventies.

Both teams were competent. Both used human autopsy tissue. Both published in journals that reject most of what they receive. And both could not be right.

That collision is the real story of neurogenesis, the process by which the brain manufactures new neurons from stem cells. It is not a settled textbook fact waiting to be summarized. It is one of the longest-running arguments in modern neuroscience, and it turns on questions most popular coverage never mentions: how long a brain sat before it was fixed in preservative, which antibody was used, whether a machine-learning algorithm counted a stem cell or a glial precursor.

Seven years after that collision, the argument has moved. A 2025 paper in Science reported finding actively dividing neural progenitor cells in adult human hippocampus [3]. A 2026 paper in Nature mapped neurogenesis across cognitively normal adults, people with Alzheimer's disease, and a rare group of eighty-year-olds with the memory of someone thirty years younger [4]. The weight of recent evidence leans toward yes. But a published dissent is still standing, and it is not a fringe one.

Six Weeks Apart, Two Answers That Cannot Both Be Right

To understand why the 2018 papers landed like a collision, it helps to know what each team was actually looking at.

Neither group could watch a neuron being born. Nobody can, in a living human. What they could do was look for molecular signatures that mark young cells. Doublecortin, usually shortened to DCX, is a protein that immature migrating neurons produce and mature neurons switch off. Ki-67 is a protein present only in cells that are actively dividing. PSA-NCAM is a sugar-coated adhesion molecule that young neurons wear on their surface. Find these markers in adult tissue and you have circumstantial evidence of recent neuron birth.

Sorrells and colleagues looked, using antibodies that bind those markers so they light up under a microscope. In children, they found plenty. In one-year-olds, young neurons were common. By age seven, far fewer. By thirteen, they were down to a handful. In adults, essentially none [1]. Their conclusion was blunt: if adult human hippocampal neurogenesis occurs at all, it is exceedingly rare.

Boldrini's team used whole hippocampi from people aged fourteen to seventy-nine who had died suddenly and had no history of neurological disease. They applied unbiased stereology, a sampling method that estimates total cell numbers from systematically spaced sections rather than from whichever field looks promising. They found intermediate progenitor cells and thousands of immature neurons at every age [2]. What did decline with age was not neuron production. It was the growth of small blood vessels feeding the region, and the size of the resting stem cell pool.

The field reacted the way fields do. Later that year, Cell Stem Cell ran a pointed exchange. Mercedes Paredes and coauthors argued the positive findings rested on markers that are not specific enough in adult humans [5]. Alexandria Tartt and colleagues from the Columbia side replied that tissue quality and processing differences, not biology, explained the negative results [6].

Gerd Kempermann, who has spent decades on this problem in Dresden, and eleven coauthors from opposing camps then did something unusual. They wrote a joint piece laying out where the evidence was strong, where it was weak, and what experiments could settle it [7]. It was closer to a ceasefire agreement than a review.

What does this mean for anyone reading about brain science? A number worth holding onto. The adult human brain contains roughly 86 billion neurons [8]. Even the most generous estimate of adult neurogenesis involves a population so small that detecting it reliably sits at the edge of what current methods can do. That is not a footnote to the controversy. It is the controversy.

The disagreement was new. The question was not.

The Man Nobody Listened To

For most of the twentieth century, the answer to this question was considered obvious, and it was no.

Santiago Ramón y Cajal, the Spanish anatomist who won a Nobel Prize for showing that the nervous system is built from discrete cells, wrote in the early twentieth century that adult nerve pathways were fixed and finished. In his view, nothing in the adult central nervous system could be regenerated once development ended. He added, to his credit, that it would be for future science to overturn this if it could. Generations of neuroscientists remembered the first part and forgot the second.

Then, in 1962, a researcher at MIT named Joseph Altman published a short paper in Science with a question for a title: are new neurons formed in the brains of adult mammals [9]. His method was tritiated thymidine, a radioactive building block of DNA. Any cell copying its DNA in preparation for division would incorporate it. Cells that never divide would not. Inject it, wait, then slice the brain and place it against photographic film. Dividing cells develop as dark spots.

Altman found them. In 1965, with Gopal Das, he published autoradiographic and histological evidence of postnatal hippocampal neurogenesis in rats [10]. Later the same year, the pair described newly generated small neurons elsewhere in the rat brain [11].

The field ignored him. Almost completely. Altman never got a tenured position at MIT and moved to Purdue. His work sat, correct and unread, for the better part of two decades.

The first serious rescue came in 1977. Michael Kaplan and James Hinds took the obvious objection seriously. Light microscopy could not prove that a labeled cell was a neuron rather than a glial cell, one of the support cells that outnumber neurons and do divide in adulthood. So they went to electron microscopy, which resolves the fine structure that distinguishes a neuron from anything else, and confirmed that labeled cells in the dentate gyrus and olfactory bulb of three-month-old rats were genuinely neurons [12].

Still the dogma held. What finally cracked it came from birds.

Fernando Nottebohm at Rockefeller University studied canaries, which learn new songs seasonally. In 1983, with Steven Goldman, he showed that new neurons were being produced, migrating, and differentiating inside a brain nucleus that controls song in adult female canaries [13]. That was interesting. What came next was decisive. In 1984, John Paton and Nottebohm recorded from these newly generated cells and showed they fired action potentials and were wired into working circuits [14].

Not just born. Employed.

The last piece arrived in 1992. Brent Reynolds and Samuel Weiss extracted cells from adult mouse brain, put them in a dish with growth factors, and watched them multiply into floating spheres that could become neurons or glia [15]. The same year, Linda Richards, Trevor Kilpatrick and Perry Bartlett generated neurons from adult mouse brain cells in culture [16]. Adult brains contained stem cells. That was no longer arguable.

Every one of those experiments was done in an animal. The human question was still open, and the obvious experiment was impossible. You cannot inject a healthy person with a radioactive DNA label and wait for them to die.

The First Human Evidence Came From Cancer Patients

The solution arrived from oncology, and it was a matter of timing.

In the 1990s, doctors treating certain cancers gave patients bromodeoxyuridine, usually called BrdU. It is a synthetic molecule that behaves like a DNA building block and gets incorporated into any cell dividing at the time. Clinicians used it to measure how fast a tumor was growing. It also, incidentally, labeled every other dividing cell in the body.

Peter Eriksson, working with Fred Gage at the Salk Institute, realized what that meant. Five patients who had received BrdU and later died had, without anyone planning it, run the exact experiment that ethics forbade. In 1998, the team published in Nature Medicine. In the dentate gyrus of every one of those five brains, including a seventy-two-year-old, they found BrdU-labeled cells that also carried neuronal markers [17].

New neurons. In adult humans. Genuinely new, because the label could only have entered during cell division after the injection.

It was a landmark, and it was also thin. Five brains. A method nobody could repeat, since BrdU stopped being used that way. And a background assumption borrowed from rodents, where the age-related decline in progenitor proliferation had already been carefully measured [18].

Attempts to fill the gap with marker studies gave mixed results. A 2010 study examining human hippocampi from newborns to centenarians found the molecular hallmarks of neurogenesis present across the lifespan, but at rates that fell steeply and early [19]. By 2011, when Guo-li Ming and Hongjun Song wrote their much-cited assessment of the field, the honest summary was that adult neurogenesis was established in rodents, plausible in humans, and quantitatively unknown [20].

Then came a method nobody in neuroscience had expected, and it came from nuclear weapons.

Above-ground nuclear testing between 1955 and 1963 pushed atmospheric carbon-14 far above natural levels. The 1963 Partial Test Ban Treaty stopped the testing, and the excess carbon-14 has been declining steadily ever since, following a curve that is documented year by year. Plants absorb it. Animals eat plants. People eat both. And when a cell divides and copies its DNA, it locks in the carbon-14 concentration of that moment and holds it for the life of the cell.

Kirsty Spalding and Jonas Frisén at the Karolinska Institutet turned this into a dating technique. Their 2005 proof of concept showed that measuring carbon-14 in genomic DNA could retrospectively establish when a human cell was born, with an accuracy of roughly plus or minus 1.5 years [21]. In 2013, they applied it to the hippocampus [22]. That paper produced the single most quoted statistic in this entire field, and also the most abused one. It gets its own section later, because the number people repeat has been stripped of everything that makes it meaningful.

A Decade of Contradiction

After 2018, the field did not converge. It escalated.

In 2019, María Llorens-Martín's group in Madrid published in Nature Medicine. Using tissue collected and fixed under tightly controlled conditions, they reported thousands of immature neurons in the dentate gyrus of neurologically healthy people into their nineties, and a sharp drop in patients with Alzheimer's disease [23]. Their argument was pointed. The reason other groups found nothing, they said, was that other groups' tissue had been handled differently.

The same year, Matthew Tobin and Orly Lazarov reported DCX-positive and PCNA-positive neuroblasts in people into their tenth decade, with higher counts in those who had scored better cognitively [24]. They also included a caveat that the positive camp does not always emphasize: the proxy markers used in humans may not demonstrate adult neurogenesis as reliably as they do in rodents.

In 2021, the Journal of Neuroscience did something almost theatrical. It published the two sides back to back. Sorrells and colleagues argued that when you include proper positive controls, meaning tissue from children that definitely contains young neurons, the adult signal disappears [25]. Llorens-Martín's group argued the opposite in the adjacent article [26].

Then the technology changed, and the argument changed with it.

Single-nucleus RNA sequencing reads which genes are switched on inside individual cell nuclei. Instead of asking whether one protein is present, it profiles thousands of genes at once and sorts cells by their molecular identity. It looked like the tool that would end the fight.

It did not. In 2022, Yi Zhou and colleagues published in Nature, using machine learning trained on validated developmental data, and identified immature dentate granule cells across the human lifespan, reduced in Alzheimer's disease [27]. That was evidence for. The same year, Daniel Franjic and colleagues published in Neuron, sequencing adult human, macaque, pig and mouse hippocampus. They found clear neurogenic signatures in mouse, pig and macaque. In human, they did not [28]. That was evidence against.

Franjic's team also raised an objection that has proved hard to dismiss: in adult humans, the transcriptional signature of so-called immature neurons overlaps substantially with that of inhibitory interneurons, a completely different and long-lived cell type. If the two look similar in the data, counting one as the other is not a hypothetical error.

By 2023, the reviews were saying so out loud. Julia Terreros-Roncal and colleagues published a detailed methods compendium arguing that technical variables, not biology, drove the disagreement [29]. Giorgia Tosoni and coauthors asked in the title of their Neuron review whether single-cell transcriptomics was reconciling the controversy or fueling it, and concluded, essentially, the latter [30].

Five years of better technology. No resolution. Which raises the question the popular coverage almost never asks.

Why Careful Laboratories Reach Opposite Answers

The honest answer is not that one side is sloppy. It is that this particular measurement is unusually fragile, and the fragility comes from four places.

The first is time. Human brain tissue does not come from a controlled experiment. It comes from people who died, whose bodies were transported, whose autopsies happened on a schedule set by hospitals and families. The post-mortem interval, meaning the gap between death and preservation, can be four hours or twenty-six. Proteins degrade during that window, and delicate markers degrade fastest.

The second is fixation. Brain tissue is preserved in formaldehyde, which cross-links proteins and locks structure in place. Leave tissue in fixative too long and those cross-links start hiding the very epitopes that antibodies need to bind. This is not speculation. In 2023, Gallardo-Caballero and colleagues ran the controlled version of the experiment in mice, where they could set post-mortem delay and fixation time deliberately. Prolonged fixation and longer post-mortem delay measurably impeded detection of neurogenesis markers [31]. In other words, a laboratory can look at genuinely neurogenic tissue and see nothing, purely because of how the tissue was handled. The protocol paper the Madrid group published in Nature Protocols in 2020 exists precisely to control these variables [32].

The third is the markers themselves. Doublecortin is the workhorse of this field and it is not a perfect witness. In rodents it reliably tags young migrating neurons. In adult humans, whether it does the same job is exactly what is in dispute, and Tobin's own 2019 paper conceded the point [24].

The fourth is what happens when you swap microscopy for sequencing. Transcriptomic methods do not escape the specificity problem, they relocate it. Now the question is whether an algorithm has correctly labeled a cluster of cells. The 2026 Nature study noted that neuroblast signatures partially overlapped with those of mature oligodendrocytes, the cells that make myelin [4]. Overlapping signatures mean annotation choices, and annotation choices mean judgment.

Underneath all of this sits a problem the field has acknowledged but not solved: there is no standardized way to count. Xinyu Zhao and Henriette van Praag laid out the case in 2020 for applying rigorous stereological methods consistently across laboratories, precisely because inconsistent quantification makes results incomparable [33].

Here is a comparison of what each method can and cannot establish.

What does this mean for a reader trying to judge a headline? When a study reports that adult human neurogenesis exists or does not exist, the useful follow-up question is never how many neurons. It is how the tissue was handled and what counted as a young neuron. Bergmann, Spalding and Frisén made this point in their 2015 assessment of the human evidence [34], and it has aged well.

Set the counting problem aside for a moment. There is a separate question that rodent work has answered in remarkable detail, and it is arguably more interesting.

What a Newborn Neuron Actually Does

Suppose new neurons are added to the dentate gyrus, the input layer of the hippocampus. Why would that matter? A brain with 86 billion neurons gaining a few hundred more sounds like adding a sentence to a library.

The answer is that new neurons are not equivalent to old ones. For a window of a few weeks, they behave very differently, and the difference is what makes them useful.

Christoph Schmidt-Hieber, Peter Jonas and Josef Bischofberger showed in 2004 that newly generated granule cells in the adult hippocampus have enhanced synaptic plasticity, meaning their connections strengthen far more readily than those of mature neighbors [35]. Three years later, Shaoyu Ge and colleagues mapped the boundaries of that window, showing a defined critical period during which the young cells are unusually easy to modify [36]. This is the same molecular machinery that underlies long-term potentiation, running at a higher gain.

The clearest demonstration of what that means functionally came in 2012. Antonia Marín-Burgin and Alejandro Schinder, working in Buenos Aires, recorded from mature and immature granule cells while stimulating their inputs. Weak input recruited only a few mature cells but activated a substantial proportion of the immature ones. The young cells were poorly selective, responding to almost anything, and that selectivity sharpened over time [37].

Think of it as the difference between a specialist and a new hire. The specialist responds only to their exact area. The new hire responds to everything, badly, and gradually narrows.

Earlier work had already established that adult-born neurons integrate functionally into hippocampal circuits rather than sitting inert [38]. The question was what job the circuit gave them.

The leading answer is pattern separation, the ability to store two similar experiences as distinct memories rather than blurring them together. Where you parked today versus where you parked yesterday. Two colleagues with similar faces. Two exam questions that differ by one clause.

The evidence for this is unusually clean because it runs in both directions. In 2009, Clelland and colleagues reduced neurogenesis in mice and found specific deficits in distinguishing similar spatial locations, while other memory functions stayed intact [39]. Then in 2011, Amar Sahay working with René Hen went the other way. By genetically blocking the programmed death of new neurons, they increased the surviving population and found that increasing adult hippocampal neurogenesis was sufficient to improve pattern separation [40]. Hen has disclosed consulting relationships in that publication, which is worth knowing when weighing any single laboratory's interpretation.

Toshiaki Nakashiba and colleagues then divided the labor explicitly, showing that young granule cells handle pattern separation while old ones support pattern completion, the complementary ability to retrieve a whole memory from a fragment [41]. Sahay, Donald Wilson and Hen argued the same computational logic applies in the olfactory bulb [42]. Michael Yassa and Craig Stark reviewed the human behavioral and imaging evidence for pattern separation in the hippocampus, though human work cannot isolate the contribution of new neurons specifically [43].

One more finding closes the loop. Nohjin Kee and Paul Frankland showed that adult-generated granule cells are preferentially recruited into spatial memory networks, meaning the brain does not merely tolerate the new cells, it selects them [44]. Wei Deng, James Aimone and Fred Gage synthesized the picture in 2010 [45].

What does this mean in practice? If pattern separation depends partly on a small pool of hyper-plastic young cells, then the memory operations most sensitive to neurogenesis are the ones where similar things must be kept apart. Which is, unhelpfully for students, most of what makes studying difficult.

And there is a cost, which almost nobody mentions.

The Neurons That Erase

Adding new cells to a circuit that already stores information is not a free operation. Something has to give.

In 2014, Katherine Akers, Paul Frankland and Sheena Josselyn at the Hospital for Sick Children in Toronto tested the obvious implication. If new neurons rewire the dentate gyrus, they should disturb memories already stored there. They trained adult mice, then raised neurogenesis afterward, either through running or genetically. The mice forgot [46].

Read that again. Increasing neuron production after a memory formed caused that memory to fade. Frankland, Stefan Köhler and Josselyn had laid out the theoretical case the previous year [47]. This is a different mechanism from the enzymatic active forgetting processes that operate at the synapse, but it points the same direction: forgetting is something brains do on purpose.

The second half of the 2014 paper is where it becomes genuinely strange.

Human beings do not remember their infancy. Almost nobody has reliable episodic memories from before age three, a phenomenon called infantile amnesia. Josselyn and Frankland had proposed in 2012 that the cause might be neurogenesis itself, since neuron production is at its lifetime peak in infancy [49]. Infant brains would be erasing memories as fast as they formed them.

Testing that required a comparison. Mice are altricial, born helpless with most hippocampal neurons still to be generated. Guinea pigs and degus are precocial, born with the dentate gyrus largely complete and neurogenesis already low. If the hypothesis was right, precocial infants should not show infantile amnesia.

They did not. Infant guinea pigs and degus retained memories that infant mice lost. Then the team raised neurogenesis in those precocial infants artificially, and infantile amnesia appeared [46]. Lucas Mongiat and Schinder, writing a companion piece in the same issue of Science, called it a price to pay for adult neurogenesis [48].

Christoph Anacker and Hen later framed the broader function as cognitive flexibility, connecting the memory and mood literatures [50]. Clearing outdated representations is not a bug in that framing. It is the point.

What does this mean for anyone trying to learn? Mostly it means caution about a tempting inference. There is no direct evidence that adult neurogenesis specifically mediates the spacing effect, retrieval practice, or any other study technique. The defensible link is narrower: hippocampus-dependent learning improves the survival of newly born neurons in rodents, and new neurons remodel the circuit in ways that both sharpen new distinctions and destabilize old ones. Anyone who tells you that a particular study schedule grows neurons in your brain is extrapolating well past the data.

There is another extrapolation that happens even more often, and it involves anatomy.

The Rodent Map Does Not Fit the Human Brain

Open almost any popular explanation of neurogenesis and you will find the same diagram. Two neurogenic zones. The subgranular zone of the hippocampus, producing granule cells for memory. And the subventricular zone lining the fluid-filled ventricles, sending young neurons along a highway called the rostral migratory stream to the olfactory bulb, where they support smell.

That diagram is accurate. For a mouse.

For an adult human it is wrong on both halves, and the error is not subtle.

The trouble surfaced in 2004, when Nader Sanai and Arturo Alvarez-Buylla examined the human subventricular zone and found something unexpected. There was a ribbon of astrocytes with stem cell properties. But the chain migration seen so clearly in rodents, where young neurons travel nose to tail in long files, was absent [51]. Maurice Curtis and colleagues reported in 2007 that a human rostral migratory stream did exist, running through a ventricular extension [52], and the finding was contested immediately.

The resolution came in 2011, and it was developmental. Sanai and colleagues showed that a corridor of migrating young neurons genuinely exists in the human subventricular zone during infancy, and then declines. By around eighteen months it is nearly gone [53]. The same paper described a transient migratory pathway in infants heading toward prefrontal cortex, a route with no rodent counterpart at all.

Then carbon-14 settled the destination question. In 2012, Olaf Bergmann and colleagues dated neurons in the human olfactory bulb. If new neurons were arriving throughout adulthood, the average neuron would be much younger than the person. It was not. Human olfactory bulb neurons are essentially as old as the individual [54].

So the rodent route runs, in adult humans, almost nowhere.

And then the reversal. In 2014, Aurélie Ernst working with Frisén applied the same dating method to the human striatum, a deep structure involved in movement and habit. New interneurons are added there continuously throughout adult life. In patients with Huntington's disease, that addition was depleted [55]. Rodents do not do this. The human brain has a neurogenic route that the standard model does not contain.

What does this mean? Two things. First, any claim that adult humans grow new neurons for smell should be treated skeptically. Second, and more broadly, the transfer of rodent findings to humans in this field has a documented failure rate, and the failures run in both directions. Some things rodents do, humans do not. Some things humans do, rodents do not.

Which brings us to the claims people actually search for.

What Actually Raises Neurogenesis, and in Which Species

This is where popular coverage does the most damage, and the damage is a single recurring move: taking a mouse result and writing it as a human fact.

Start with what is genuinely well established, in rodents.

Environmental enrichment came first. In 1997, Kempermann, H. Georg Kuhn and Gage housed mice in cages with tunnels, toys and running wheels and found more hippocampal neurons than in standard cages [56]. Two years later, Henriette van Praag isolated the active ingredient. Running alone increased cell proliferation and neurogenesis in the adult mouse dentate gyrus [57], and a companion paper showed running also improved learning and synaptic plasticity [58]. Elizabeth Gould and colleagues showed that hippocampus-dependent learning itself improves the survival of newly born neurons [59].

Stress runs the other way. Gould's group demonstrated suppression of granule cell precursor proliferation by psychosocial stress in tree shrews [60] and then in adult marmoset monkeys [61]. That is primate data, which is closer to human than mouse, but it is still not human.

The antidepressant story is the most instructive, because it shows a strong claim being narrowed over time. In 2003, Luca Santarelli and colleagues reported that blocking hippocampal neurogenesis by targeted irradiation abolished the behavioral effects of antidepressants in mice [62]. The headline version became: antidepressants work by growing neurons. Jason Snyder and Heather Cameron later showed that adult neurogenesis buffers stress responses and depressive behavior, identifying a specific subset of dentate neurons involved in controlling the stress hormone axis [63]. Real findings. But the strict requirement claim did not survive replication across strains, species and behavioral paradigms, and antidepressants demonstrably act through several mechanisms that have nothing to do with neurogenesis. Snyder's own 2019 review argued for recalibrating how much the field claims [64].

Now the human side, and this is the part that gets misreported.

In 2007, Ana Pereira and Scott Small measured cerebral blood volume in the dentate gyrus of exercising humans using MRI, having first shown in mice that blood volume there tracks neurogenesis [65]. They measured blood volume. Not neurons. In 2011, Kirk Erickson and colleagues ran a randomized controlled trial in 120 older adults and found that aerobic exercise training increased hippocampal volume by about 2 percent while the stretching control group declined by roughly 1.4 percent [66]. They measured volume by MRI. Not neurons.

Hippocampal volume can change for many reasons. Vascular changes, glial changes, dendritic growth, changes in extracellular fluid. Neurogenesis at the scale anyone proposes for humans could not plausibly account for a 2 percent volume change on its own. These are important, well-conducted human studies. They are not measurements of neurogenesis, and citing them as such is a factual error that has spread widely.

The table below separates what was measured from what is often claimed.

None of this means exercise is pointless or that stress is harmless. Aerobic exercise has strong evidence behind it for cardiovascular health, mood and cognition, and the hippocampal volume result is real. It simply is not a neuron count. James Aimone and colleagues gave the field a careful synthesis of what regulates neurogenesis and what that regulation does [67], and Tomohisa Toda with Gage reviewed the disease implications [68]. Both are worth more than any single headline.

There is one number, though, that has escaped its own paper entirely.

Where the Number 700 Actually Comes From

Search almost anything about brain health and a figure appears: the adult human brain makes 700 new neurons every day. It is printed as though it were a measured headcount, like a census.

It is not. It is a modeled estimate from one paper, and it has four qualifiers that are almost always deleted.

The paper is Spalding and colleagues in Cell in 2013 [22]. The method is the carbon-14 dating described earlier. Neuronal nuclei were separated from human hippocampal tissue by flow cytometry, their DNA extracted, and the carbon-14 concentration measured and matched against the known atmospheric curve.

Here is the sentence as the authors wrote it, with everything intact. In adult humans, 700 new neurons are added in each hippocampus per day, corresponding to an annual turnover of 1.75 percent of the neurons within the renewing fraction, with a modest decline during aging.

Four qualifiers. Per hippocampus, meaning per side, not per brain. An annual turnover of 1.75 percent. Within a renewing fraction, which the paper estimated at roughly one third of dentate granule cells, the other two thirds being as old as the person. And derived from a mathematical model fitted to carbon-14 measurements, not from counting neurons one by one.

Strip those qualifiers and the number changes meaning entirely. It stops being an estimate about a subpopulation and becomes a promise about your whole brain. Even the Karolinska Institutet's own communications have framed it as roughly 1,400 neurons across both hemispheres, which is arithmetically fine and shows how easily the framing slides.

The paper also deserves credit for a finding that undercuts the tidy version. The decline with age was modest, which sits awkwardly beside marker-based studies reporting steep declines and beside what is otherwise known about how aging changes memory. Two methods, two different shapes of curve, both published. That tension has never been resolved.

There are also real criticisms of the approach. Carbon-14 dating measures population turnover, not visible cell division, so it can tell you a population contains younger cells than the individual without showing you a single dividing progenitor. It depends on modeling assumptions, particularly the split between renewing and non-renewing subpopulations. And the same laboratory's olfactory bulb result [54] is a warning against assuming that progenitor activity anywhere guarantees integrated neurons downstream.

For orientation, the staging framework everyone uses, from quiescent stem cell through amplifying progenitor to neuroblast to surviving integrated neuron, was formalized by Kempermann and colleagues in 2004 [69], and Chunmei Zhao with Gage set out the underlying mechanisms in detail four years later [74]. Most cells born do not complete the journey. They die before integrating, in a selection process with echoes of the synaptic pruning that reshapes the adolescent brain. Survival, not birth, is the bottleneck.

Which is exactly the point the skeptics have been making for a decade.

The Case That It Barely Happens at All

Any honest account of neurogenesis has to give the skeptics a real hearing, because their argument is not stubbornness. It is arithmetic.

Sara Cipriani and colleagues examined hippocampal radial glial subtypes, the stem-cell-like cells that sit at the top of the neurogenic chain, in human fetuses, healthy adults and Alzheimer's patients. They found the neurogenic potential of these cells declining steeply after early life [70]. If the stem cells are gone, downstream production has nowhere to come from.

Then there is the counting problem, stated most sharply by a group at Yale. In 2022, Jon Arellano, Alvaro Duque and Pasko Rakic published a formal Comment in Science responding to a paper by Terreros-Roncal and colleagues on neurogenesis in neurodegenerative disease [71]. Their objection [72] was methodological and it generalizes: cells identified as adult-born neurons in human tissue may be misidentified, and the numbers reported are difficult to reconcile with what is actually visible in the tissue.

That reconciliation problem is worth spelling out, because it is the strongest argument the skeptics have.

Take the carbon-14 numbers at face value. Over a childhood, they imply the generation of something on the order of ten million granule cells. Sustained adult production at the reported rate implies a similar magnitude again. A cell population producing millions of neurons should leave behind an abundant, obvious population of dividing progenitors. Dividing cells are not subtle. They are the easiest thing in a tissue to stain.

And yet for twenty-five years, laboratory after laboratory has struggled to find them. Not found them in small numbers. Struggled to find them at all. Sorrells and colleagues, using positive controls from children as a benchmark, reported essentially none in adults [25]. Franjic and colleagues, sequencing rather than staining, found no clear neurogenic trajectory in adult human tissue while finding one in three other mammals [28].

There is a version of this story where the skeptics are simply right. Where adult human neurogenesis is a residue of development, largely finished by adolescence, and where the positive findings reflect markers that are not as specific in humans as they are in mice.

That version has not been ruled out. It has become less likely. Those are different things, and the difference matters when reading what happened next.

2025 and 2026: Progenitors, SuperAgers, and a Dissent That Holds

On 3 July 2025, Science published the paper the field had been waiting twenty-five years for.

Ionut Dumitru, Marta Paterlini and colleagues in Frisén's laboratory at the Karolinska Institutet went looking for the missing progenitors directly. They sequenced hippocampal tissue spanning infancy through adulthood, combined it with Ki-67 antibody staining to catch cells in the act of dividing, and applied machine-learning classifiers to sort the results. In children, every stage of the neurogenic sequence was present. In adults, they identified proliferating neural progenitor cells, and located them by transcriptomic signature to the dentate gyrus [3].

They also found something that fits neither camp cleanly. The variation between individuals was enormous. Some adult brains contained many progenitors. Others contained almost none. If that holds, both sides of the 2018 collision may have been reporting accurately about different people. Frisén discloses consulting work for a genomics company, which does not bear on the data but belongs in any full account.

Then came the dissent, and it arrived fast.

Within days, an eLetter appeared on the Science record for that paper, authored from the Department of Neuroscience at Yale University. It does not dispute the data. It disputes the reading. Its argument runs on the same arithmetic laid out above: if the carbon-14 model is right about how many granule cells adults generate, progenitors should be abundant, and a population that can only be resolved with a machine-learning algorithm is by definition not abundant. The letter concludes that the same evidence is consistent with a steep postnatal decline in progenitors, aligning with Cipriani, Sorrells and others, and that neurogenesis may be negligible in adult humans if it exists at all. A second eLetter on the same record raises related concerns about what sequencing can establish.

Others voiced the same worry in the press. Sorrells suggested the classifier may have been trained in a way that flags progenitors destined to become glia, which unambiguously do regenerate in adulthood, rather than neurons.

Eight months later, the picture got more detailed and more interesting.

In February 2026, Ahmed Disouky, Orly Lazarov and a large collaboration published in Nature. They applied multiomic single-cell sequencing, reading gene expression and chromatin accessibility in the same nuclei, to 355,997 nuclei from human hippocampi across five groups: young adults with intact memory, healthy older adults, adults with preclinical Alzheimer's pathology, adults with Alzheimer's dementia, and a rare cohort called SuperAgers, people over eighty whose episodic memory matches that of people in their fifties [4].

They found neural stem cells, neuroblasts and immature granule neurons in every group. In Alzheimer's disease, neuroblasts and immature neurons were sharply reduced, even though stem cell numbers went up, which suggests a blockage in the pipeline rather than an empty reservoir. In SuperAgers, neurogenesis ran at roughly twice the level of other healthy older adults, with elevated expression of brain-derived neurotrophic factor in the young cells. The epigenetic profiles of the Alzheimer's and SuperAger groups were close to mirror images of each other.

The authors were careful about the limits. Sample sizes were small, individual variation was large, and the SuperAger effect rested on a handful of brains.

Related work is filling in the regulation. In early 2026, a team at the University of Cincinnati led by Yu Luo showed that blocking a specific signaling pathway in microglia, the brain's resident immune cells, stimulated hippocampal neurogenesis and reduced anxiety-like behavior [73]. That work was done entirely in mice, and it belongs in the rodent column.

So where does the argument stand in August 2026? Two independent laboratories using different technologies have reported adult human neurogenic cells, and their signatures agree with each other. That is meaningful. A published objection from a serious group holds that the same data support the opposite conclusion, and no experiment has yet ruled it out. Both of those statements are true at once. Anyone claiming otherwise is choosing a side rather than reporting one.

Conclusion

Sixty years after Joseph Altman was ignored, the question he asked is still open. That is unusual, and it says something about how hard the human brain is to study.

What has changed is the quality of the disagreement. In 1962 the argument was whether adult mammals make new neurons at all. That is settled: rodents do, birds do, primates do. In 2018 the argument was whether humans do. That has narrowed but not closed. Today the argument is about magnitude, about whether the numbers implied by one method can be reconciled with what another method can see, and about whether a small and highly variable process is meaningful for how a person thinks and remembers.

Three things are worth carrying away.

The first is that the interesting biology is not the birth of neurons but what they do in their first few weeks. Hyper-excitable, poorly selective, easy to modify, preferentially recruited into new memories, and capable of destabilizing older ones. A small population of unusual cells can matter out of proportion to its size.

The second is that most of what circulates as fact about boosting neurogenesis in humans has never been measured in a human. Running wheels and enriched cages produced real, replicable findings in mice. MRI scans of exercising adults produced real findings about hippocampal volume and blood flow. Those are not the same measurement, and the gap between them is where a great deal of confident writing has gone wrong.

The third is about the 700. One paper, one model, per hippocampus, within a renewing fraction of about a third of granule cells, at 1.75 percent turnover a year. That is a careful, hedged, honest estimate. It became a slogan by having its hedges removed.

The brain does not owe us a clean answer, and on this question it has declined to give one for sixty years. What it has given instead is a case study in how science actually works when the measurement is at the edge of what technology can do. Not a verdict. An argument, still running, conducted by people who keep publishing their disagreements where everyone can read them.