In data science, algorithmic design, and institutional governance, the pursuit of perfection is increasingly treated as a necessary virtue. Modern systems are routinely optimised to erase all anomalies, minimise all errors, and smooth out the complex and untidy irregularities of human behaviour.
But what if the truth is more counter-intuitive; that a completely optimised, hyper efficient system that eliminates all error is inherently brittle and increasingly incapable of survival? This is the contention put forward by Professor Simon Reay Atkinson, an expert in complex adaptive systems from the University of Sydney Business School, in the first of Link Digital’s new Honest Series of conversations.
The theory behind Atkinson’s talk is explored in greater detail in a paper, co-authored by Atkinson and Link Digital’s Executive Director Steven De Costa, The lie is to the perfect as perfection is to life: If to err is human, to forgive is divine – what future when AI plays divine?
The following article will briefly recap Atkinson’s webinar presentation.
The need for systematic variance
In the context of Atkinson’s talk “the lie”, deceit, or error does not imply any malicious falsehood. Rather, it represents the importance of systemic variance. “We’re not encouraging people to lie,” said Atkinson. “But what I do think we’re looking at is what makes the oyster. It’s the grit in the oyster that makes the oyster. Without grit in the machine, we essentially end up in perfection and perfectionism.”
“And the work that Steven and I have been examining has led us to a conviction that there is a great threat posed by AI in that its virtue is that it removes variance and error. That very variance and error is what enables humanity to change, adapt and, if you like, to live. You remove that or you end imperfection, and you essentially remove life itself.”
For data professionals and structural designers, there are two core themes in Atkinson’s argument that raise questions about how they should collect, interpret, and implement data in an increasingly automated world.
The first centres on the necessity of systematic variance, or what Atkinson called “jitter” and “hunting”. He maintains that in complex adaptive systems, variance is not a structural failure to be purged; it is the exact mechanism that enables adaptation and resilience. When an algorithm or model achieves mathematical perfection, it leaves no room for adjustment. It freezes the institutional architecture into a state of structural rigor mortis, rendering it unable to handle unexpected real-world shifts.
He illustrated this point with reference to Australia’s now infamous Robo-debt scandal. This involved an unlawful automated welfare debt recovery scheme implemented by a previous Australian government from 2015 to 2019. It used income leveraging to incorrectly accuse hundreds of thousands of welfare recipients of owing money to the government, resulting in severe financial hardship, trauma, and suicides.
Atkinson set out how the system utilised a rigid, highly optimised averaging algorithm to detect income discrepancies. Because the model assumed its mathematical abstractions were a perfect mirror of reality, it treated any real-world variance as deliberate deceit. It lacked an analogue “hunting” loop – the capacity to hesitate around a point, estimate, and allow for systemic friction. In the case of the Robo-debt scandal, the system reduced human complexities to an unyielding mean, resulting in automated absurdity and human tragedy.
“If you have a highly optimised perfect system, there is no room for life. And, therefore, within that, the deceit – the variance – actually becomes what gives us life; that enables adaptation and change.”
For data professionals, this theme is an urgent warning against over-fitting models and over-automating decisions. When data systems operate purely within the strict, flattened confines of digital tokens, they replace nuanced conceptualisation with sterile
representationalism. This rigidity can also breed algorithmic hallucinations when a system encounters different realities.
Atkinson contends that data architects must intentionally build grit back into their systems, ensuring that models retain analogue-like estimation spaces where variance is audited and integrated rather than blindly flattened.
Why institutional ethics are not the same as human morality
The second theme in Atkinson’s presentation is the important distinction between institutional ethics and human morality, and how this divergence necessitates “cognitive variety” over simple “managed diversity”.
Atkinson argued that ethics often manifests as scripted logic – a playbook of rules, constraints, and algorithmic boundaries. Morality, conversely, is an existential human attribute rooted in lived experience, empathy, and localised judgement.
He demonstrated this point with reference to his field experience during the brutal conflict between Bosnia and Herzegovina from 1992 and 1995. In particular, the incident that occurred when a Dutch United Nations battalion, strictly adhering to their explicit rules of engagement, stood back while thousands of men and boys were marched away to their deaths in what became known as the Srebrenica massacre.
“The [Dutch] colonel obeyed the rules of engagement. The rules of engagement gave him the ethical authority to stand back. The result was that 5,000 men and boys were marched away and massacred.” According to Atkinson, the commander in question possessed complete ethical authority under their scripted logic, yet the outcome was an absolute moral failure that not only led to large-scale deaths but devastated the battalion’s long-term morale and psychological stability.
In other words, a system can be entirely ethical – and compliant with its internal rules – while being fundamentally immoral. Data-driven organisations, whether they be the modern public service or corporations run the risk of dangerous institutional blindness when they mistake adherence to regulatory data metrics or compliance checklists for genuine systemic health.
Instead, they must cultivate true cognitive variety, the ability to bring together empirical problem solvers, engineers, humanists, and artists who operate outside identical conceptual models. Without cognitive variety to challenge the data inputs, organisations become “zombie institutions”—perfectly compliant, yet entirely detached from reality. “If we have no agility in organisation, then we’ve got fragility, zombieism and unhappiness.”
The path forward for modern data practices
What is the primary takeaway of this for data scientists? Ultimately, it shifts their role from that of an agent of total optimisation to a curator of balanced complexity.
Data models should never be decoupled from the messy, existential pulse of human intent. By designing frameworks that respect systemic jitter and by building diverse cognitive networks to audit automated logic, data professionals can ensure their systems remain resilient, ethical, and fundamentally aligned with human life.
You can watch the full recording of Atkinson’s talk on our YouTube channel here.
You can find the research paper mentioned earlier, The lie is to the perfect as perfection is to life here, which directly expands on the core theme discussed in the webinar.
