When you read the word binary, your mind probably goes somewhere first. Gender, politics, the cultural lines we are constantly asked to pick a side of. Put that down for a minute. The binary I am talking about is the computing one: zeros and ones, on and off, true and false. The technical logic that has quietly run underneath our lives for the last fifty years.
AI takes us well beyond that, in every sense of the word “beyond”, and across every meaning of “that”. Our world does not reduce to a clean black or white anymore. We have moved into something closer to a quantum future, where many things can be true at once, where the answer depends on the question, the context, and who is asking. The responsibility for critical thinking has shifted with it, and it sits squarely with us.
You do not honour that responsibility by switching off; you honour it by staying in the room.
In April, I co-facilitated a workshop at the Aus Gov Data Summit, and someone made the comment that AI is not foolproof. My immediate thought was: Well, what is?
After immersing myself in everything AI for the past two years, including building (and rebuilding) my own agentic workforce, it was a bit of a shock to be brought back into conversations where AI keeps getting collapsed into binary categories. Yes or no. Good or bad. Safe or risky. The reality sits in the grey, in textured, complex greys that demand contemplation. These topics are big, amorphous and at times challenging. Refusing to engage with the grey does not minimise the risk profile, it amplifies it.
Right now, a lot of people are dismissing the tool because they either have not engaged with it, or have tried once, not had a great output, and not engaged again. The world being what it is today, disengagement is not a luxury we have. Not everyone needs to become an AI expert, but everyone, particularly knowledge workers, needs to meaningfully engage with it. There is no better way to arrive at an informed view of both the risks and the opportunities.
There is a deep irony with people sermonising the risk of losing our critical thinking skills when often those same people are not engaging with the technology. How can you apply a critical lens to something you have not engaged with? Reducing it to a good-or-bad lens is reductive and does not serve the complexity of our world.
How can you apply a critical lens to something you have not engaged with?
There are real questions to sit with. One often posed is, how will universities and schools adapt meaningfully if kids can get deep answers this easily? Perhaps a more expansive and dare I say useful question might be: how might we help our kids recognise the questions they need to ask to ensure objectivity and protect their critical thinking? How do we actively acknowledge and account for biases including automation bias and confirmation bias? How will we be able to confidently assure ourselves and our communities that we are working from the best knowledge we can access, and have not been walked down a digital garden path by a sycophantic digital persona?
Often where I see people baulk is at the point where they innately recognise that these answers are not clear cut, and their critical thinking functions shut down. This is precisely the point where we should lean in. Regardless of whether we do or not, the rest of the world already is, and that includes both the good actors and the nefarious ones. If we want to continue to have input as a real economic contender, we can’t afford to shut off.
The risks are real, and that is the point
It’s generally at this point people are internally screaming at me: the risks though! We can’t forget about the risks! I do understand. My response: we can hold both the risks and the need to engage at the same time, in service of greater knowledge and capability.
There are many examples of real biases at play today. The conversations at the Aus Gov Data Summit really brought home to me that we have a romanticised view of how safe and in control we were before AI, and how unsafe we have become with it.
For example, we now know and understand that AI uses data which carries in-built bias, but that bias only seems to register as a risk when automation is in play. What about the decisions made today by a human using that same data? There is a false sense of security that because a human has rubber-stamped something, it must be sound. Very few people and organisations have built in a logic or rationale check against the data, and in fact, if the data has told us “x”, then it’s empirically true and touted as the “source of truth”. AI simply brings the limitations to light. Rather than condemning AI, perhaps we should be celebrating. When we know better, we can do better.
AI simply brings the limitations to light. Rather than condemning AI, perhaps we should be celebrating. When we know better, we can do better.
I know data was my undoing in my own AI journey. I started learning the mechanics of how to prompt, gained a little knowledge, and got a bit cocky, right up until the point I tried to build something more complex and it immediately failed. I had the good fortune of healthy competition (although my competitor was blissfully unaware of his involvement). I needed to understand why some things would work and some things would not, or work for a bit and degrade over time. Sure, I could say the AI was “dumb”, but it is merely a tool, only as good as the craftsperson who wields it. We love to knowingly look at each other, exchanging witticisms about “garbage in equals garbage out”, but then what do we do once we’ve hit the very short peak of where good prompts will take us? Worse still, if you are not aware of the limitations of the tools you or others are using, the likelihood for subpar outputs masquerading as detailed research increases exponentially. Take the time to understand the output. Is it true? Believable? How do we test it? Was the data underneath it any good? Why would it only give me one answer after a certain point in time? What was the actual issue? This process, to understand the question, the answer and whether we are comfortable with the output, this is critical thinking. Saying “well, that did not work, so I am not going to engage anymore” is the opposite of critical thinking.
Start with a deficit, not a tool
There is a pervasive belief that everyone needs to be doing AI the same way, and I do not think that is true. Where I have seen the best effect of AI and automation is when someone identifies a genuine deficit in their own workflow and focuses on a small change to fix it. A small, measurable change supported by AI can make an immeasurable difference.
I have supported transformational change in IT for over twenty years and we saw the same dynamic when SharePoint was introduced. People asked for training. Trainers came in and explained what the system could do. Most of that information was useless, because the only thing that mattered to the person being trained was how the system could support them. The same is true now. Until you identify a deficit and then get into the tool and have a good look around, you do not really know what you need.
The information about getting started is actually not that helpful either. You are told to experiment, but to be careful. Most resources, public and private, tell people to start by asking AI to summarise a document. That seemingly basic action is not a great place to start for many reasons, not least of which is that the action itself is inherently risky. The moment you touch data, you are increasing risk, and without good prompting mechanisms, the output of that simple prompt may not be helpful, or it could minimise core elements that are the crux of the issue. A far more useful prompt would be to feed the tool a number of supporting or background documents which relate to the document you need to read. Ask your AI tool of choice to act as an executive advisor (role) and provide the context that you need to review the document as a product owner or policy holder, and that you need a 1 to 2 page summary based on the attached documents as they are critical background information which led to the creation of the document you are reviewing. Goal, context, source and output: all the hallmarks of a great prompt.
The public servant stuck in the gap
I spend a lot of time with public servants who understand and agree with the government’s position to increase AI use, and who still do not feel confident enacting the policy. Despite having read the policies and participated in training, the fear of the unknown risk can become paralysing, especially in the wake of RoboDebt. And even where it is not fear driving it, the amount of information that is out there can paradoxically make it harder to know where to start. The general consensus is that people understand this technology provides both risks and opportunities, but they do not get support to contextualise how it might help them at work today.
I also want to acknowledge that this is not for lack of trying from the supporting agencies. It is a reflection of the fact that adoption of AI should look different for everyone.
To start to bridge that gap, I have built a series of free, low-risk prompts for public servants to trial. They introduce people to a series of high-value, low-risk prompts to try, as well as some explanation as to what makes them lower risk. The majority of these use cases take credible information from publicly available sources to deliver results that are useful from the get go. These include an individualised media brief, or an approach to learning fundamental information about a framework, concept or policy area you want to understand better. Instead of starting with information from inside the agency, the prompts use trusted information from the public domain and deliver meaningful results without internal exposure. A private sector version has also been created, with more information being added regularly.
What good actually looks like
One thing I would like to see in the AI discourse is positive examples, where AI has been implemented well.
At the same Aus Gov Data Summit I chaired a panel featuring a representative from Whitehorse City Council, who have rolled out a range of citizen-centric accessibility initiatives. One of them is an AI-driven chat that lets residents engage with council in their own language. The benefit on both sides is real. The council reaches people who might otherwise have been quietly cut out of the conversation, and the resident gets to ask a question in the language they actually think in.
Earlier this year I led a social research project interviewing people with intellectual disabilities about their experience accessing government services. One of the strongest findings was that people often just need a little more time to hear and process what they are being told. What if AI could give them that time, instead of stretching an already overworked call centre operator with a growing queue and a looming service level agreement infringement?
These are not hypothetical. They are happening now, and they answer the “what does good look like” question more clearly than any policy document.
Resist the urge to go too big
We sometimes try to be too big with this technology. Taking the entire NDIS assessment process and reducing it to an automated system is, sure, ambitious, but is that not exactly where the human value lives, and where it should continue to live? Ask a tech company whether their tool can solve a big, audacious problem you are facing, and they will tell you, with full conviction, that it can. And that’s not to begrudge tech teams, I’m sure it can, with the right data and the right time, and the right money. Anything is possible! But is it the most effective approach? Just because we can, doesn’t mean we should.
Developing your own experience in building and working with these tools gives you an opportunity to learn the seemingly simple outputs they are quite terrible at, and the applications they are phenomenally good at. Without a clear understanding of how and when to use them, we risk spending serious time and money automating the complex things that need empathy, when smaller changes would make a more meaningful difference to someone’s life.