Muneera Bano and Didar Zowghi of CSIRO published a piece last month arguing that AI bias is not an error in the algorithm but a chain of human decisions [1]. They are right, and the argument gets stronger if you push it further than they do.
Their case is that bias enters through choices, not code. People decide what data to collect, how to label and categorise it, whose experiences are treated as important, what safeguards ship with the product, and how a company responds when someone reports being harmed. The evidence is their own 2025 study putting GPT-4 and Microsoft Copilot through recruitment scenarios. Both systems favoured male profiles, particularly for senior roles, and the images they generated skewed younger, slimmer and lighter-skinned.
The section I want to take seriously is the one titled "Why technical fixes are not enough". It makes three moves. Technical work matters but treats symptoms rather than causes. Data does not enter an AI system as a neutral record of reality. And testing each protected attribute separately can pass while people at the intersection of several are still harmed, a point established a decade ago when Buolamwini and Gebru showed commercial gender classifiers failing darker-skinned women at rates their per-attribute numbers concealed [2].
All of that is correct. But there is one sentence elsewhere in the article that decides the whole conclusion, and it passes without examination: "AI models learn from those associations and formalise, automate, and repeat them at a larger scale."
Associations. That noun is carrying the argument.
The toolbox is a property of the architecture
Look at what the article offers as the technical options: bias identification, re-balancing datasets, and adjusting outputs. Reweight the input, probe the behaviour, post-process the output.
That is not a survey of what engineering can do about bias. It is the complete and exhaustive list of what you can do to a large statistical model, and the reason the list is so short is structural. In a transformer, the association you object to is not located anywhere in particular. It is distributed across billions of parameters, and it is observable only as a statistical tendency in the outputs across many prompts. You cannot open it up and look at it. You can only measure its behaviour from the outside and then lean on what goes in or what comes out.
So the authors find the toolbox shallow, and they are quite right that it is. Then they generalise from that to a claim about technical work as such, and that step does not follow. The shallowness is a property of correlational architectures. It is not a property of engineering. The conclusion is architectural, and the architecture is never named.
I want to be fair about why. These are serious responsible-AI researchers who certainly know that other kinds of AI exist. Their study was of GPT-4 and Copilot, so the scope of the evidence is entirely defensible. What is not defensible is drawing an AI-wide conclusion from it without flagging that the limit belongs to the system they tested.
What changes when bias is locatable
Some architectures do not learn associations at all. I should declare an interest here, because I build one of them. In those systems the same bias stops being a diffuse statistical tendency and becomes an object you can point at.
The worked example is computational mechanics. In the tradition of Shalizi and Crutchfield, a category is a causal state: a set of histories that induce the same distribution over what happens next, and provably the smallest such set that predicts the process as well as it can be predicted [3]. The states are discrete, countable, and reconstructed from the data rather than fitted as weights. Four consequences follow, and none of them is cosmetic.
Bias becomes an inspectable object
If contexts about senior leadership and contexts about men collapse into a single causal state, that merge is a specific, named thing, with an explicit record of the evidence that justified it. You can open it, read why it happened, and argue about it. The equivalent procedure in a large model is five hundred prompts and a significance test, which tells you that something is wrong without ever telling you where it lives.
It becomes a falsifiable claim about the corpus
A causal state asserts something checkable: in this data, these histories predict the same futures. That relocates the argument. Instead of trading anecdotes about biased image outputs, you have a specific exhibit on the table and a concrete question about whether that predictive equivalence is one your institution is willing to act on.
Splitting is not the same as rebalancing
Re-balancing a dataset is a correction applied by hand with no truth condition attached to it. Nothing tells you whether you have gone far enough, or too far. Splitting a state on predictive insufficiency is a different kind of operation: the merge survives only if the data genuinely makes those histories interchangeable, and dissolves if it does not. One is a thumb on the scale. The other is a claim you can test.
Abstention answers the intersectionality point directly
This is the one that matters most, and it lands squarely on the article's strongest argument. Failing people at the intersection of several identities is, mechanically, a thin-evidence problem. A statistical model fills thin regions by interpolating from dense ones, because that is exactly how it generalises, and that interpolation is the harm. The system has no way to distinguish "I have seen many people like this" from "I have seen almost nobody like this", so it answers both with equal fluency.
A system that reads its own confidence at that thin intersection and declines to answer fails in a completely different way. It fails visibly, as a refusal that someone can escalate, rather than invisibly, as a confident wrong answer that nobody catches. That is not a mitigation bolted on afterwards. It falls out of how the machine is built.
Where the argument has to stop
None of this is an escape from the authors' thesis, and I would rather state the limits myself than have someone else state them for me.
A different architecture over a biased corpus still finds the bias. If the historical record makes gender predictive of seniority, the induction will discover precisely that, and it will be right to, because it is a real regularity in the data it was shown. A causal state is a minimal sufficient statistic of an observed history. It describes the process in front of it. It has no opinion about whether that process is just.
Nor is structure the same as intervention. In Pearl's terms this is the bottom rung of the ladder [4]. Reconstructing the states of an observed process tells you nothing about what a fairer world would look like, and using the word "causal" for those states does not quietly buy you counterfactual reasoning.
And interpretability is not accountability. Being able to point at the state that merged leadership with maleness does not decide whether to intervene, who pays if you do not, or what an acceptable residual risk is. That is entirely the human chain the authors are describing, and their governance argument survives all of this untouched.
There is a fourth point that cuts against my own architecture rather than for it. Someone chooses the corpus, and someone chooses what counts as the signal in the first place. That second choice determines what structure can be found at all, which arguably makes it a more consequential human decision than anything in the transformer pipeline. It is a chain-of-human-decisions point, and the authors would have made it had they been looking at systems like mine.
What is not being measured, and why
Here is the deeper problem. The article could not easily have been written any other way, because there is almost nothing else operating at a scale worth writing about.
On the Stanford AI Index figures for 2025, generative AI grew by more than 200 per cent and captured nearly half of all private AI funding worldwide [5]. One architecture is absorbing roughly half the money going into the entire field.
When that happens for long enough, the limits of an architecture stop looking like the limits of an architecture. They start looking like the limits of AI. This article is a careful, well-evidenced instance of exactly that slippage, written by people who know better, which is what makes it worth writing about rather than merely disagreeing with.
So consider what is genuinely unknown. We have a decade of fairness literature about correlational models and almost none about reconstructed-state ones. Not because the question was asked and answered, but because those systems do not exist at a size anyone has been funded to audit. Nobody has measured whether locatable bias is actually easier to govern in practice, or whether it merely moves the argument somewhere more technical. Nobody has tested whether abstention at thin evidence reduces intersectional harm or just converts it into denial of service, which would be a real cost and might be the larger one. Nobody has established what an audit standard for such a system should even require.
Those are answerable empirical questions. They are unanswered because of where the money went.
There is a second reason to care that has nothing to do with bias. A monoculture shares its failure modes. If every hospital, court, insurer and employer runs the same architecture, they inherit the same blind spots at the same moment, and there is no independent system anywhere in the ecosystem positioned to disagree. Diversity of architecture is a resilience argument before it is ever a fairness argument.
What to ask for
The authors' recommendations are right and should be adopted: interdisciplinary teams, genuine participation from affected communities, testing whether benefits and harms distribute fairly across groups, and governance that actually monitors systems once they are live.
But notice what those require in practice. In a correlational system, "test whether benefits and harms distribute fairly" is an attestation exercise. You sample outputs, run statistics, write it up, and the regulator takes your process on trust, because there is nothing else to examine. In a system whose states are enumerable, the same requirement could in principle be a mechanical check against an inspectable structure. Whether it actually works that way is an open question, and that is the point: it is worth money to find out.
Three things follow. Fund the comparison rather than the advocacy: same task, same biased corpus, two architectures, and measure whether locatable bias is genuinely easier to detect, contest and correct, with a published negative result if it is not. Write procurement rules that can reward inspectability, so that a system whose internals an auditor can examine directly does not score identically to one where the only available evidence is a document describing good intentions. And treat architectural diversity as sovereign capability, because Australia will not out-spend the frontier labs on scale, but it can fund the part of the field that the scaling race has no commercial reason to explore, which happens to be the part where the governance questions in this article are tractable.
The authors close by saying that human decisions about which risks are acceptable remain central. That is exactly right, and it is the strongest argument for their own funding case rather than against it.
A society that has built only one kind of AI has already made that decision once, by default, on everybody's behalf. Making it properly requires having something to decide between.
References
- Bano, M., and Zowghi, D. (2026). AI bias isn't just an error in the algorithm. It's a chain of human decisions. The Conversation, 20 August 2026. Republished by CSIRO.
- Buolamwini, J., and Gebru, T. (2018). Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Proceedings of Machine Learning Research, 81, 77-91.
- Shalizi, C. R., and Crutchfield, J. P. (2001). Computational Mechanics: Pattern and Prediction, Structure and Simplicity. Journal of Statistical Physics, 104, 817-879.
- Pearl, J., and Mackenzie, D. (2018). The Book of Why: The New Science of Cause and Effect. Basic Books.
- Stanford Institute for Human-Centered AI (2026). The 2026 AI Index Report, Economy chapter. Figures for 2025 private investment.