Roger Clarke’s recent article on avoiding GenAI's traps examined what comes out of these tools.

Last month I joined an AI leadership course run by Eric Nguyen, with several colleagues from the ACS Management Committee.

The exercise was simple: show each other how we actually use AI.

It was more revealing than I expected, mostly because of how differently capable people use the same tools.

Tools, not agents

I am a heavy user, but a deliberately conservative one.

I use tools rather than agents. I have no need to hand autonomy to software, at least not yet.

Each day I move between Claude, ChatGPT, Copilot and others, and I treat them like smart friends, each with a talent I lack.

I use them to generate ideas, to check my messages and tone before they reach a minister, a member, or a fellow director, and to research areas of interest.

My most common use is the least glamorous.

I write or dictate a long, rambling stream of thoughts, then ask the model to give it structure.

The ideas are mine. The model gives them shape.

From author to editor

AI is extremely good at taking a huge amount of data and making sense of it.

Work that used to take me days or weeks now takes a few hours. My productivity is perhaps tenfold, and I think it could be better.

But it has changed what I do. I have become an editor rather than an author.

I no longer worry much about formatting, and I can focus on the things that matter: the judgement, the message and the decision.

It also helps with breadth. At any time, I am working through problems for several organisations at once and AI helps me keep them separate and move between them without losing the thread.

Still checking everything

Roger is right that the prose is smooth and authoritative whether or not it is correct.

So, I still check all the output. I verify every source personally.

If I haven't checked it, it doesn't go in.

Linking models together is a great start, but it doesn't remove that responsibility.

I still make mistakes, and the efficiency makes it easier to make them faster. Being quicker means the discipline of checking matters more, not less.

Same tools, different purposes

I talk with Prins Ralston, our CEO, almost every day, and AI comes up constantly, from a policy, strategic and tactical perspective.

Prins is a heavy user from a research and fact-checking perspective using multiple models to ask questions around the same challenge in subtly different ways, and yet he hand writes all his documents the old fashion legal way.

I like to call him a token hog, because if anyone is going to hit a research usage limit, it is Prins, and he does it faster than anyone I know.

But we use AI for vastly different purposes and for different reasons.

A president and a CEO sit on opposite sides of the governance line.

Prins is accountable for running the organisation, and I am accountable for its oversight and direction.

So, the questions we bring to AI differ, and so does the risk we carry. ‘

Hearing his approach to researching and fact checking has sharpened mine, and I suspect the reverse is true.

The risk on the way in

My own caution has always been about the question, not just the answer.

Whatever you type into a public model leaves your control.

So, I have one hard rule: nothing sensitive or confidential goes into a public model. Not board papers, not legal matters, not personnel issues.

The cost is that the material where I would most value a second opinion is often the material I can't share.

A late-night experiment

Then I stumbled across a YouTube video on running a model offline.

One night I tried it, using llamafile to run an open-weight model on my own machine, with no internet connection involved. It took some mucking around to get the settings right, but it worked.

That raised the obvious question: what if I ran several models side by side? I built a simple web interface that runs locally and puts three models in one window: Qwen3 4B Thinking from Alibaba, Gemma 3 4B from Google, and Llama 3.1 8B from Meta.

I can put the same question to all three and compare the answers.

The result is what I call my council.

The whole thing runs from a 32GB USB drive that is completely air-gapped from my network. I can load sensitive material, ask questions, and know it goes nowhere.

I gave each model a different brief, drawing on the roles and skillsets you find around a board table.

One looks through a risk lens, another a commercial one, another as a sceptical outsider.

I also use them to review code and suggest improvements to style and formatting.

Three models are still just models. Where they agree, I gain a little confidence. Where they disagree, I have found the question worth asking.

But agreement is not proof, because models can share the same blind spots.

The council tells me where to look. It does not tell me what is true.

These local models are also far smaller and less capable than the best commercial ones, and I accept that trade for the privacy.

A long way to go

I am still at the start of this journey, and there is a long way to go.

The tools will change, my habits will change, and I expect to look back on today's setup with some amusement.

For leaders starting out, three lessons stand out.

Build the habit before the autonomy. Set your data rules before you need them. And match the use to the role.

Roger describes GenAI as an explosive. Used carefully, it is powerful. Used carelessly, it harms others and you.

The harder challenge, as he notes, is organisational use.

That will need an open conversation about what we use these tools for, and what we are prepared to feed them.