Beyond the Pandemic of the Mind: What It Actually Looks Like to Use AI Well

As an undergraduate, I spent two years studying symbolic logic. My mother asked me, reasonably, what I thought I was going to do with it.

I did not have a good answer at the time. If she asked me the same question today, I would tell her that AI was coming, still some decades off, and that the two years I spent formalizing arguments into notation turned out to be some of the most useful preparation I could have had for understanding it.

Symbolic logic teaches you, in a way nothing else quite does, what a formal system can and cannot do with reference to truth. A valid argument, in the logician’s sense, is one where the conclusion follows necessarily from the premises. Validity has nothing to do with whether the premises are actually true. You can build a perfectly valid argument from false premises and arrive at a perfectly confident, perfectly wrong conclusion, and the logic itself will offer no warning that anything has gone astray.

Truth has to be checked against the world, separately and continuously; the formal system will never do that checking for you, no matter how airtight its internal structure appears.

That distinction, drilled into me decades before anyone was talking about large language models, is close to the single most useful thing I know about how these systems reason. An AI system is, at bottom, an extraordinarily sophisticated pattern-completion engine operating over the structure of language and argument. It can produce chains of reasoning that are fluent, internally coherent, and confidently stated, in exactly the way a valid syllogism is internally coherent, without any of that coherence guaranteeing that the premises it started from, or the facts it inserted along the way, are actually true.

Fluency is not validity, and validity is not truth. Symbolic logic is the discipline that teaches you to hold those three things apart instead of collapsing them into one impression of “this sounds right.”

It is, I think, exactly the discipline that AI use demands and that its ease of use quietly erodes.

Which brings me to a report that has just crossed my desk, written by a friend and colleague, that names this erosion directly and treats it as a matter of national consequence.

The Diagnosis

Air Vice-Marshal John Blackburn AO (RAAF Retd), Chair of the Institute for Integrated Economic Research – Australia, has just published a report that deserves a wide readership well beyond Australia’s policy community. A Pandemic of the Mind? The Potential Impact of AI on National Security and Resilience argues that AI’s most consequential effects may turn out to be cognitive rather than technological: a slow erosion of the analytical and moral judgment that democratic societies, and the institutions that defend them, depend on.

Blackburn’s central metaphor is deliberately provocative. Cognitive change, he argues, spreads through a population the way a pathogen spreads through a body: transmitted through information networks, largely invisible until well advanced, and shaped more by the design of the systems carrying it than by any individual’s choices. He traces a reinforcing cycle, increased use, cognitive atrophy, reduced confidence, greater reliance, that will be familiar to anyone who has watched a colleague stop checking an AI-generated draft because checking it has started to feel like the harder path.

He extends the same logic into the moral domain: when we stop thinking critically, he writes, we also stop questioning ethically. And he places both inside a second cycle, a foreign dependency trap, in which reliance on offshore frontier models becomes a sovereignty risk as consequential as reliance on offshore fuel or medicine.

I don’t dispute the diagnosis. I have spent the past several years immersed in exactly the kind of high-stakes, human-machine environments where Blackburn’s concern is most acute — the kill web architectures now replacing linear kill chains across air, missile defense, and maritime autonomous systems.

The report’s warning about automation bias, about decision-makers who accept machine output because interrogating it has become unfamiliar, is not abstract in that world. It is the difference between a system that extends human judgment and one that quietly displaces it.

Where I want to add something to Blackburn’s argument is on the other side of the ledger: not just the discipline required to avoid cognitive offloading, but what it actually looks like, in practice, to extract AI’s genuine value without falling into the trap he describes.

Diagnosing the disease is necessary. It is not the same as prescribing how a practitioner should actually work.

The Distinction Is Not Use Versus Non-Use

The temptation, when a report like this lands, is to read it as an argument for caution bordering on abstinence. That would be a misreading, and Blackburn himself does not make it. His own account of working with agentic systems across a twelve-month research project is instructive precisely because it shows sustained, heavy AI use producing better judgment rather than worse.

The distinction that matters is not how much AI a person uses. It is whether the human remains the party who directs the analytical process or becomes the party who merely receives its output.

In the kill web architectures I write about, this same distinction has a name: the difference between a system that disaggregates sensing and shooting so that a human retains command authority over the engagement decision, and one that collapses the loop so tightly that the human’s role becomes ratifying a decision already effectively made. The equivalent choice in knowledge work is just as concrete.

It is the difference between asking an AI system to produce your conclusion and asking it to stress-test a conclusion you have already formed independently.

Three Things That Distinguish Augmentation From Offloading

Set against Blackburn’s cycle, a few practical distinctions are worth naming plainly, because they are what separates genuine augmentation from the atrophy he describes.

Sequence matters more than frequency. The question is not how often a person consults an AI system, but where in the reasoning process the consultation happens. Forming an independent view first, and only then bringing AI in to challenge, extend, or pressure-test it, preserves the muscle that atrophies when AI is consulted first and adopted by default. This is a sequencing discipline, not a usage limit and it is compatible with very heavy AI use, provided the sequence holds.

Disagreement is the signal to watch, not agreement. A system that only ever confirms what its user already believes is not augmenting judgment; it is automating confirmation bias at scale, which is precisely the failure mode Blackburn’s cited research identifies as sycophantic reinforcement.

The useful test of whether a tool is functioning as a genuine collaborator is whether it has meaningfully changed your mind, or surfaced a consideration you would not have reached alone, within the last several uses. If it never has, the tool is not being used as a collaborator regardless of how the relationship is described.

Cross-verification is not a luxury, it is the method. Running an analytical question through more than one system, and against primary sources, is not inefficiency. It is the closest analogue available to peer review in a domain where a single model’s failure mode is invisible from inside a single conversation.

This adds friction. That friction is the point.

It is the same reason a well-designed kill web retains redundant sensor paths rather than trusting a single node: single points of failure are dangerous precisely when they are fluent and confident, which is exactly how a hallucinating model presents itself.

The Period Ahead

Blackburn frames the choice Australia faces as institutional rather than technological, a matter of design decisions, not inevitabilities. I’d extend that one step further. Over the next several years, the organizations and individuals who benefit most from AI will not be the ones who adopt it fastest or the ones who resist it longest.

They will be the ones who treat the human’s role as unapologetically load-bearing: the party who sets the question, holds the standard of evidence, and remains accountable for the conclusion, while treating the AI system the way I have come to think of it in my own work, as an exceptionally capable but fundamentally limited analytical partner, closer to a gifted intern than an oracle, whose output is a starting point for scrutiny rather than a substitute for it.

That is a harder discipline to sustain than either uncritical adoption or reflexive caution, because it offers no shortcut in either direction.

It is also, I think, the only discipline that actually delivers on the promise Blackburn is pointing toward: AI that strengthens the judgment of the people using it, rather than quietly hollowing it out while looking, at every individual step, like progress.

This article draws on and responds to John Blackburn’s “A Pandemic of the Mind? The Potential Impact of AI on National Security and Resilience” (Institute for Integrated Economic Research – Australia, August 2026), available at the following:

https://www.jbcs.co/iieraustralia-projects