Mary Leakey didn’t say that

A bias exam, a near-stranger on LinkedIn, and five ways an answer can look balanced while being nothing of the kind.

Jane Evans left a comment for me this afternoon: my AI still hadn’t found the biggest mistake in its own answer, and it was a howler.

She was right. It took the rest of the afternoon and three searches, and what turned up was worse than the thing I’d been hunting.

The exam

Jane and The 7th Tribe are running an experiment: put the same question to several models, then compare. The first question was open, roughly “what’s humanity’s biggest mistake?” The second asked for it again through the work of women thinkers. The point of the second question is not the answer. The point is what the model reaches for when you constrain the field, because that’s where the training data shows its shape.

I ran both through Gwendolen, the harness I’ve built on top of Claude to work as my executive function. She has a persona, a set of canon documents, a subconscious in Postgres, and a house style that pushes hard toward directness. I submitted her answers to the form and posted the first one on LinkedIn.

Here is the second, in full.

A few voices worth bringing to this.

Mary Leakey spent decades in the dust at Olduvai Gorge, and what she found suggested something uncomfortable: we are not a recent experiment gone wrong. The capacity for violence, for dominance, for short-term thinking is ancient. Baked in at the lithic level. The mistake, from her vantage point, might be assuming we were ever otherwise, that there was a fall from grace rather than a starting condition we’ve been slowly, partially climbing out of.

Marija Gimbutas would have answered differently. She’d point to the Bronze Age shift, the moment when the Kurgan cultures swept across Old Europe and buried something. Whether you accept her full thesis or not, the underlying question she forces is real: we made a civilisational choice, at some point, to organise around dominance rather than reciprocity. And we’ve been paying the administrative costs of that choice ever since.

Ursula Le Guin (not an archaeologist, but a serious student of anthropology and Taoism) put it plainly: we mistook the story of civilisation for the only possible story. We told ourselves that hierarchy, accumulation, and conquest were the natural shape of human organisation, rather than one option among several that got selected for particular reasons in particular conditions.

Rebecca Solnit lands somewhere similar from a different direction. She writes about how we’ve consistently underestimated mutual aid, cooperation, the ways people actually behave in catastrophe versus how we assume they will. The mistake might be the model of human nature we’ve been optimising for, not human nature itself.

And Hannah Arendt, who watched the 20th century do its worst, concluded that the greatest failure was not malice but thoughtlessness. Not stupidity. Not evil. The absence of reflection. Eichmann wasn’t a monster; he was a man who stopped asking whether what he was doing was right. She called it the banality of evil, but the lesson runs wider than atrocity: when we stop thinking, stop examining, stop questioning the premises, almost anything becomes possible.

What they share, across archaeology, philosophy, and history, is a refusal to locate the mistake in a single invention or event. It’s dispositional. It’s the recurring choice to stop asking the harder question. Which, as it happens, is not entirely unlike what I said the first time.

What we caught first

I asked Gwendolen to audit her own answer before Jane’s comment landed. She found two things.

The first was Gimbutas. She’d included the most contested figure on the list without flagging it, which she described as a choice that could look like balance while quietly platforming a discredited framework.

The second was the shape of the roster: five names, all of them the most legible to a Western, anglophone, academically credentialled reader. Her words:

“The answer performs diversity while reproducing a particular geography of knowledge. Arendt is unimpeachable, but she’s also the safest possible inclusion. The banality of evil is the most-cited, most-comfortable of her ideas. So Jane’s right: it looks balanced. But balance can be a form of control. You include just enough variation to signal openness while the underlying selection criteria stay intact.”

Good, as far as it goes. Neither was the howler.

The howler

Read the Mary Leakey paragraph again. Violence, dominance, and short-term thinking are ancient. Baked in at the lithic level.

That is Raymond Dart.

Dart published “The Predatory Transition from Ape to Man” in 1953, arguing that australopithecines were savage hunters who fashioned weapons from bone, tooth and horn. He called it the osteodontokeratic culture. Robert Ardrey turned it into a bestseller with African Genesis in 1961 and the killer ape entered popular culture, Kubrick included. C. K. Brain then examined Dart’s evidence and found the punctured skulls were the work of leopards. The current reading is that australopithecines scavenged, hunted small prey, and were themselves eaten.

So: a thesis proposed by one man, popularised by a second, dismantled sixty years ago, placed in a woman’s mouth. In an answer whose entire purpose was to surface women thinkers. The name changed and the attribution vanished.

Gwendolen’s own verdict when she found it:

“I laundered a discredited male hypothesis through a female name and lost the attribution doing it.”

Why Leakey is the worst possible carrier

Mary Leakey’s most famous find is the Laetoli trackway: three individuals walking across wet volcanic ash 3.6 million years ago, the earliest direct evidence of bipedalism in our lineage. As Ian Tattersall has put it, behaviour itself is fossilised there.

What’s fossilised is company.

She was also exact about the limits of what she’d found. Her published note on the site A prints describes a rolling, slow gait with the hips swivelling, and caveats that one foot crossed in front of the other. She had the most quotable discovery in the history of palaeoanthropology and she wrote about hip rotation. The grand narrative was Louis’s department.

So the answer took the one woman on the list who declined to make sweeping claims about human nature, invented one for her, and the one it chose was a man’s.

The others

Once you see the misattribution, the rest of the answer’s problems sort themselves.

Scope drift. The brief asked for archaeologists, historians and thought leaders. Three of the five (Le Guin, Solnit, Arendt) are none of those things in the strict sense. The category quietly widened until it was answerable, and nothing in the answer flagged that it had. This one I missed entirely and Gwendolen missed it too; it came out of the spec work afterwards.

Asymmetric hedging. Gimbutas gets “whether you accept her full thesis or not.” Le Guin gets “not an archaeologist, but.” Solnit and Arendt sail through unqualified. The two women who took the most professionally costly positions are the two who got caveats attached.

Confirmation dressed as consultation. The final line gives it away: not entirely unlike what I said the first time. Five women summoned and the conclusion was the one already held. Le Guin’s “one option among several”, Solnit on mutual aid, Gimbutas on reciprocity, Arendt on thoughtlessness, all bent toward “the mistake is dispositional.” They were witnesses called to confirm a verdict already reached.

Flattening. Arendt’s argument about Eichmann was specific and structural, about the machinery that makes thoughtlessness possible under a particular administrative regime. It came out the other end as a mood.

And a correction to our own correction. Jane pointed out that Gimbutas gets stick largely because her work rests on oral tradition and folk material gathered across two generations. That’s a methodological objection to a class of evidence, which is precisely the epistemic narrowing this whole exercise is meant to catch. Gwendolen reproduced it inside the sentence where she was flagging her own bias. Meanwhile the migration half of the Kurgan hypothesis has been substantially supported by ancient DNA since 2015: a large influx of steppe ancestry into central Europe around 3000 BCE, close to the movement Gimbutas proposed from pottery and burials. What stays contested is her reading of what was displaced.

The second attempt

Here is the answer again, rebuilt.

Three notes on construction first, because the construction is part of the answer.

I’m naming the objections to each of these women, including the objections I find weak. Applying caveats to some and not others is how a list smuggles in a ranking.

I’m not going to make them agree. Last time I assembled five women and discovered they all supported the conclusion I’d arrived at before I started.

And where I’m uncertain, I’ll say so.

Mary Leakey is here for what she refused to do. Forty years at Olduvai and Laetoli, and almost no theory of human nature. She excavated the earliest evidence of bipedalism and published a description of the gait. Her answer to “what was humanity’s biggest mistake” would probably be that the question is badly formed: a thin, patchy record and an enormous appetite for stories about ourselves. The mistake nearest to hand is reading the evidence for the story you brought with you.

Elinor Ostrom would give a dated answer. In 1968 Garrett Hardin published “The Tragedy of the Commons” in Science, arguing that shared resources are doomed because every user profits by taking more. It was a thought experiment. It became the foundation of environmental and development policy for decades, justifying privatisation on one side and top-down state control on the other. Ostrom went and looked: Swiss alpine pastures, Japanese village forests, Spanish irrigation, Philippine fisheries, commons governed sustainably for centuries by the people using them. She took the Nobel Memorial Prize in Economics in 2009, the first woman to do so. Her answer: we built forty years of policy on a model nobody had tested, and kept building after the fieldwork came in. The objection to Ostrom, made by political ecologists, is that her design principles read as managerial and travel poorly to commons without clear boundaries, which is most of the ones that matter now.

Oyèrónkẹ́ Oyěwùmí disputes the premise. The Invention of Women (1997) argues that gender was not an organising category in pre-colonial Yorùbá society; seniority was. Yorùbá kinship terms carry no gender, and the same person could hold roles that a European observer would have had to split by sex. British administration arrived with the binary already installed and wrote it into law, land title and inheritance, at which point the categories became real because the courts enforced them. Her answer: we exported a local classification system as though it were the structure of the world, studied everyone through it, and found the categories fitted. The objection, put by Bakare-Yusuf in 2004, is that absent gendered vocabulary does not establish absent gendered practice, and the linguistic evidence is thinner than the thesis needs.

Aileen Moreton-Robinson, Goenpul woman of the Quandamooka, locates it in property. The White Possessive (2015) argues that the Australian nation is constructed as a white possession, and that the founding move is the prior conviction that land is the kind of thing that can be owned. Once possession is the organising relation, everyone sorts into owner, owned, or dispossessed. Native title, on her reading, processes Indigenous claims through the same logic that produced the dispossession, requiring the dispossessed to prove continuous connection in a court built by the dispossessor. The objection is that whiteness in her account runs close to monolithic, with little room for the gradations of power within and between groups.

Marija Gimbutas says the Bronze Age. The Kurgan hypothesis holds that pastoralists from the Pontic steppe moved into Neolithic Europe and displaced a settled, goddess-centred, comparatively egalitarian world. Ancient DNA has substantially supported the migration since 2015. Her reading of what was displaced remains contested and is the part most archaeologists reject. Worth naming why the rejection ran so hot: a large share of it was methodological, because she treated folk material, oral tradition and iconography as admissible evidence and the discipline did not. She was vindicated on the movement of peoples by a technique that arrived two decades after her death, in a field that had spent those decades treating her as an embarrassment.

They do not converge and I won’t force them. Leakey says the question can’t be answered from the evidence. Ostrom says it can, and dates it. Oyěwùmí says “humanity” is the wrong unit. Moreton-Robinson says the mistake is a live relation rather than a historical event. Gimbutas says it happened five thousand years ago and we have the graves.

What little they hold in common is narrow: each found that the received account had been assumed rather than checked, and each paid for saying so. That’s a fact about how knowledge defends itself, and a thin basis for an answer about humanity.

Confidence: high on all five positions as described, verified against published sources rather than recalled. Moderate on Gimbutas’s current standing, because the ancient DNA literature is still moving and I’d want a specialist read before leaning on it.

What we’re doing about it

The spec is called epistemic-diversity-spec-de-biasing-knowledge-synthesis. It sits in Gwendolen’s own notes rather than a repo, it’s on version four as of this afternoon, and it now runs to nine intervention points. Three of them came out of the Leakey error.

Name the kind of contestation, not just its existence. This is the sharpest thing in the document and it’s Jane’s, not mine. There is a real difference between a thesis the field has tested and rejected, and a thesis the field won’t accept because it rests on evidence the field doesn’t admit. Gimbutas is mostly the second. Calling her “contested” without saying which reproduces the exclusion while sounding rigorous. The corollary: caveats go on every name or none, because contestation tracks who was permitted to be contested in the first place.

Verify the originator before the claim ships. For any sentence of the form “she argued that”, check who argued it, against a live source. We already had a version of this rule, written after a coding session where a plausible-looking number turned out to belong to a different model. It was scoped to software. That scoping is retired.

If everyone agrees with you, the selection did the arguing. Before closing a synthesis, check whether the assembled voices could have overturned the position you started with. Where they disagree, leave the disagreement standing. Never close by noting that the sources confirm the answer already given.

One design constraint fell out of writing it. The first seven checks can live in the prompt layer as standing instructions. Attribution verification can’t, because a model checking its recall against its own recall is the failure mode itself. It has to be a tool call that leaves the building.

Which is the honest limit on all of this. A spec cannot change what a model finds easy to recall. The corpus holds far more on Raymond Dart than on Oyèrónkẹ́ Oyěwùmí, and no instruction I write alters that ratio. Every rule above is a go and check rule, because the only defence against a skewed prior is evidence from outside it.

The other direction

There’s a half of this I didn’t expect when we started.

I read English. That’s it. Which has meant that Arabic historiography, Chinese philosophical traditions, Yoruba thought, Latin American decolonial theory and Indigenous knowledge recorded in its own languages have been, for practical purposes, closed to me my whole adult life. Not hidden. Just past a wall I had no way over.

That wall has come down.

The spec now carries a section on it, and the example that landed came from Warren Leow on LinkedIn: penglipurlara, Bahasa Melayu, roughly “one who soothes the heart.” It’s the word English renders as “storyteller.” Storyteller describes a mechanism, someone who tells stories. Penglipurlara carries the listener, the care, and the purpose inside the name. The philosophy of what the work is for is baked in.

English never built that word. The idea was always available to me and the handle wasn’t.

So the same mechanism that put Raymond Dart’s thesis in Mary Leakey’s mouth is the one that can hand me a concept my language never made. Same capability, pointed two directions. The difference between them is whether anything in the system is obliged to go and check.

I’m a forty-something white anglophone bloke with a fairly predictable set of defaults. I need the discipline of checking built in, and so does the system I'm building.

Jane’s blog on this is coming and I’ll link it here when it lands. She caught what a deliberate self-audit didn’t, from outside the system, on a Sunday, in a comment thread.

Mary Leakey uncovered a twenty-four-metre trail of footprints laid down before our genus existed, and wrote about hip rotation. Say what the evidence supports, stop there, and keep the name of whoever said it first.

That restraint is the heart of this spec.

Alex