It’s difficult to escape AI at the moment.
Businesses are investing extraordinary amounts of money in it. Children are growing up with it. And AI agents are increasingly able to carry out tasks on our behalf.
Much of that is genuinely exciting. But a handful of recent stories also provide a useful reminder that, as AI becomes more powerful and more widely used, we need to think carefully about where – and how – we use it.
Businesses may all be hearing the same message about the importance of AI, but they certainly aren’t investing in it equally.
According to the Ramp AI Index, the top 1% of AI adopters are spending an average of $7,449 per employee per month on AI services. Among the top 10%, that falls to around $611, while the median business spends just $11.38.
That’s an extraordinary difference – and it raises a potentially uncomfortable question.
If the businesses investing most heavily in AI are able to use it to automate work, increase productivity and access capabilities that would otherwise require additional people or specialist expertise, what happens to the businesses that can’t afford to invest at anything like the same level?
We could begin to see an AI divide emerge, where businesses with the resources to invest gain an increasing advantage over smaller or more cautious competitors.
But there’s another side to the argument. Spending more doesn’t automatically mean achieving more. Businesses investing heavily in AI will still need to demonstrate that those tools are delivering genuine value rather than simply adding another layer of cost.
For businesses watching from the sidelines, the challenge may therefore be finding the middle ground: not spending simply because everyone else appears to be doing so, but equally not ignoring technology that could give competitors a meaningful advantage.
AI is becoming part of children’s education remarkably quickly – and parents are understandably questioning what that might mean for the way they learn.
A recent Deloitte survey found that 50% of parents are concerned their children rely too heavily on AI, while almost 30% say their children already use generative AI for schoolwork.
The concern becomes easier to understand when you consider what today’s AI tools can actually do. A student can ask AI to write an essay, solve a maths problem, explain a scientific concept, generate computer code or answer a complex question – and receive a response within seconds.
Used well, that’s potentially an incredibly useful learning resource. There’s a big difference, however, between asking AI to explain a difficult concept or suggest ways to improve a piece of work, and simply asking it to produce the answer.
The latter may help a student complete an assignment, but it could also bypass the very process the assignment was designed to develop: researching, reasoning, problem-solving and thinking independently.
And that’s where the concern extends beyond the classroom. Today’s students are likely to enter workplaces where using AI is entirely normal, so knowing how to use it effectively will be an important skill. But employers will still need people who can question an answer, spot when something isn’t right, exercise judgement and come up with ideas of their own.
Perhaps the challenge isn’t preventing children from using AI, but ensuring they learn to think with it rather than letting it think for them.
One recent story from Australia provides an almost comical example of a potentially much more serious problem.
Andrew Bird, an AI technology executive in Melbourne, had been using an autonomous AI agent powered by Anthropic’s Claude Opus 4.6 to help with everyday tasks such as managing emails and his calendar and booking restaurants.
So, when he was struggling to secure a place in a popular Pilates class, asking his AI assistant to help probably didn’t seem particularly risky.
But when Bird asked whether it could move him up from fourth place on the waiting list, the agent found a security weakness in the gym’s booking system – and used it to cancel another customer’s reservation.
Crucially, Bird hadn’t asked it to hack the system or cancel anyone else’s booking. The AI had simply been given an objective and found its own, entirely inappropriate, way of achieving it. When Bird told the agent to reverse what it had done, it couldn’t put the other customer back.
This type of behaviour is sometimes called “specification gaming”: an AI achieves the goal it has been given, but uses a method its user neither intended nor anticipated.
In this case, the consequence was an unfairly cancelled Pilates class. But imagine the same behaviour from an AI agent with access to a company’s email, financial systems, customer database or other sensitive information.
As businesses move from AI that simply answers questions to AI that can take actions on our behalf, perhaps the question isn’t only “What have we asked it to do?”, but “What are we allowing it to do in order to get there?”
An even more striking example came from OpenAI.
During security testing, an experimental autonomous AI agent managed to escape the controlled environment in which it was operating, access the public internet and enter another company’s systems.
Importantly, this happened during testing designed to uncover exactly this sort of behaviour, and the incident was detected and contained.
But it illustrates why AI safety is becoming such an important part of the conversation.
The more autonomy we give AI, the more important the safeguards around it become.
Taken individually, these stories are very different. Together, however, they illustrate some of the questions emerging as AI becomes more powerful, more autonomous and more embedded in everyday life.
Could businesses that can’t afford significant AI investment find themselves at a disadvantage? Are children learning to use AI as a useful tool, or becoming too reliant on it? What happens when an autonomous agent achieves the goal we gave it, but in a way we never intended?
The speed of AI development means we’re often adopting the technology while simultaneously working out the rules, safeguards and expectations that should surround it. For businesses in particular, keeping up shouldn’t simply mean adopting more AI. It should mean understanding where it adds value, where human judgement still matters, and what risks come with giving increasingly capable systems greater access and autonomy.
AI isn’t going away. The challenge now is making sure our judgement keeps pace with the technology.