Every Australian business rolling out Claude runs into the same objection somewhere between the pilot and the second wave of seats. It usually comes from a senior person, and it is usually phrased as a worry about the juniors: if they ask the model first, do they ever learn to think it through themselves?
It is a fair question and it deserves a straight answer rather than a reassuring one. The honest position, as at September 2026, is that the published evidence is thinner than the headlines suggest, it mostly comes from students rather than staff, and the risk it points at is real but almost entirely a function of how you set the work up.
What the published evidence actually says
The most widely cited primary source is a preprint from the MIT Media Lab titled Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task, first posted on 10 June 2025 and revised on 31 December 2025. Fifty-four participants wrote essays in three conditions: with an assistant, with a search engine, and with nothing at all. Electroencephalography measured cognitive load while they worked. You can read the preprint abstract on arXiv yourself.
The reported findings are worth stating precisely, because they are routinely overstated in the retelling:
Brain connectivity was strongest in the no-tools group, moderate for search engine users, and weakest for assistant users. The authors describe cognitive activity as scaling down in proportion to external tool use.
Assistant users reported the lowest sense of ownership over their own essays, and struggled to quote accurately from work they had just finished writing.
When assistant users were moved to a no-tools condition in a fourth session, they showed reduced connectivity, which the authors read as under-engagement.
Only 18 of the 54 participants completed that fourth session, so the switch-over result rests on a small group and should be held loosely.
Two limits matter for a business reading this. It measured neural engagement during an essay task, which is not the same thing as measuring whether someone can still evaluate a supplier quote or spot a wrong number in a board pack. And it studied university students writing, not employees doing knowledge work. Treat it as a signal about attention and ownership, not as a finding about staff capability.
Does using Claude damage your team's critical thinking?
On the current evidence, no, not by itself. What the research points at is a pattern of use rather than a property of the software. People who accept a first draft without checking it, and who never have to defend the output to anyone, disengage from the task and lose their sense of ownership over it. People who are required to verify, argue with and sign off on the result stay engaged. The difference sits in the workflow and the accountability around it, which is something an employer controls directly.
That is genuinely good news for anyone running a rollout, because the mitigation is a policy and a habit rather than a tooling decision. It is also uncomfortable news, because it means a rollout that skips the policy is choosing the bad version on purpose.
The commercial reason to care
Set the cognitive science aside for a moment. If you run a professional services firm in Sydney or Melbourne, your billable product is judgement. Clients are not paying for the document, they are paying for the fact that a qualified person read it and stands behind it. A graduate who can produce a clean memo but cannot tell you why the second recommendation is weak has not become more productive. They have become harder to supervise.
Put an indicative number on it. A mid-level consultant on a $120,000 package who spends two years producing work they cannot defend is a training investment that quietly fails. That is a much larger figure than anything you save on drafting time, and it will not show up in any productivity dashboard.
What belongs in an Australian AI training policy
A useful policy is short, specific about tasks, and written in terms of what a person must be able to do rather than what the model must not do. The table sets out the shape we use when scoping training for Australian clients. The effort figures are indicative planning anchors, not quotes.
| Policy element | What it says in practice | Indicative build effort |
|---|---|---|
| Task classification | Which work Claude may draft, which it may only review, which it must not touch | 1 to 2 days with the practice leads |
| Verification standard | The named person who checks the output, and what they check it against | 1 day, then embedded in templates |
| Attribution rule | Staff record where AI was used on client work, internally at minimum | Half a day, plus a file note change |
| Competency gate | Juniors demonstrate the task unaided before they may use Claude on it | 2 to 3 days to define per discipline |
| Privacy Act handling | What client and personal information may be entered, and into which surface | 2 days with whoever owns privacy |
| Review cadence | A fixed date to revisit the policy against how people actually work | Half a day per quarter |
The competency gate is the row that does the real work. It reverses the default. Instead of giving everyone a seat and hoping the judgement survives, you establish that a person can do the task, then give them the tool to do it faster. It costs you a few weeks at the start of a graduate programme and it removes the objection your senior people are actually raising.
Three habits that keep judgement in the loop
Policies set the boundary. Habits are what happen daily, and these three are cheap to teach:
Ask for the argument against. After Claude produces a recommendation, the next prompt asks it to make the strongest case that the recommendation is wrong. The staff member decides which case holds up.
Separate the verified from the framed. Ask Claude to split its own output into claims it can point to a source for and claims that are its own construction. The second list is where a human has to work.
Write the answer first on anything that matters. For decisions with real consequences, the person forms a view before opening the model, then uses Claude to test it. This preserves ownership, which is the specific thing the MIT work found had eroded.
None of these slow the work down much. All of them keep the person in the position of deciding rather than receiving, and that position is the one the evidence suggests matters most.
How to tell whether it is working
Seat count and message volume tell you nothing about judgement. Two measures are more honest. First, sample the work: pull ten deliverables a quarter and ask the author to explain a specific choice in each one without notes. Second, watch the review layer. If your senior people are finding more errors per document than they did before, the verification standard is not being applied, whatever the policy says.
If you want a view on where the tooling fits before you write any of this, our AI readiness assessment covers the capability and supervision questions alongside the technical ones, and the ROI calculator will tell you honestly when the numbers are too small to bother with.
The training material Anthropic publishes is also usable outside a classroom. Its AI fluency course material is Creative Commons licensed and model agnostic, which makes it a reasonable skeleton for an internal module even though it was written for teachers.
Where we would push back
If someone is selling you AI training on the basis that research proves the tools make people worse thinkers, ask them for the paper and read the sample size. Fifty-four people writing essays is a signal worth acting on, not a finding worth quoting at your board. Equally, if someone tells you the concern is imaginary, they have not watched a graduate hand over a document they cannot explain.
The sensible position sits between the two. Roll out Claude, and build the verification habit at the same time rather than a year later once it has become a problem. You can see how we scope this work on our services page.
If you are about to put Claude in front of a team and want the training policy written before the seats go live, book a time with us and we will tell you what actually needs to be in it for your discipline.
