In 2018 I began arguing for an idea I called Collective AI. The claim was simple: the most useful artificial intelligence would not be the kind that replaces a person, but the kind that amplifies people, gathering what many of them know and putting it in the hands of each. At the time, the public conversation was about robots taking jobs. Eight years on, with generative AI in everyone's pocket, I think the evidence is starting to favour the amplification view. But it comes with a warning that jobs policy has barely begun to consider.
The amplification story
Think of the person running a one-person consultancy or a small publication. A few years ago, research meant days of reading and searching, editing meant paying someone or going without, and routine admin ate the evenings. Today an AI assistant can summarise a body of research in minutes, catch a clumsy paragraph and, as agents mature, take on routine chores. The principal still decides what matters, what is true and what is worth saying. But the distance between a sole trader and a firm of twenty has narrowed.
This is Collective AI in practice. These models are trained on the accumulated work of millions of people. When a small operator draws on one, they are borrowing the know-how of a crowd, filtered through their own judgement. The machine does not replace the person; it gives the person reach.
Advertisement
The best evidence we have points the same way. A study of more than 5,000 customer support agents by Erik Brynjolfsson, Danielle Li and Lindsey Raymond found that an AI assistant raised productivity by 14 per cent on average, and by 34 per cent for novice and low-skilled workers, with minimal effect on the most experienced. The authors suggest the tool spread the practices of the most able workers to everyone else. That is close to a textbook case of Collective AI: the knowledge of the best people, amplified across the group.
Whether those gains turn into higher earnings, or simply into sharper competition, is still an open question. I suspect a mixture of both. Policy makers should resist the urge to declare victory or catastrophe until the numbers are in.
The formation problem
Here is the uncomfortable part. The biggest gains in that study went to novices. That is good news in the short run, but it raises a question we should not ignore: if a junior worker gets the answer from the machine from day one, do they ever develop the judgement the experienced worker brought to the job?
An experienced person using AI knows when it is going off the rails. They could do most of the work themselves, so they can check it. Someone who starts with AI before their judgement has formed may never form it at all. The tool that lifts them may also hold them in place.
There is early evidence for the concern. A small study from the MIT Media Lab followed people writing essays with and without ChatGPT. Those who relied on the AI showed the weakest brain connectivity, struggled to quote their own work and felt the least ownership of it. Fifty-four participants is no basis for national policy. But it fits what many managers already sense: people learn by doing the hard part, and AI is very good at doing the hard part for them.
I see this in my own work. At Critical Future we build AI systems for businesses, much of it for finance teams, including an AI accountant through Critical Finance. The most important design decision is rarely the model. It is deciding where a person reviews the work, and making sure that person knows enough to disagree with it. The systems that last make capable people more capable. The ones that disappoint assume nobody needs to understand what the machine did.
Advertisement
What this means for jobs policy
If Collective AI is the right frame, jobs policy should do three things: spread the gains, protect the apprenticeship and teach judgement.
First, spread the gains. The productivity dividend is real but uneven. Small operators can often adopt faster than large firms because they have no committee to convince, but they also have no training budget. Practical, low-cost training for small businesses, delivered through industry bodies and TAFE, would do more than another national strategy document.
Second, protect the apprenticeship. Every profession has an entry-level stage where people learn by doing work that is now easy to automate: the junior accountant reconciling accounts, the graduate drafting a first memo, the cadet reporter writing up a council meeting. If employers automate that work entirely, they save money this year and lose their future senior staff in ten. Governments can encourage the alternative through the way they fund traineeships and graduate programs: reward employers who keep juniors doing real work, with AI as a check rather than a substitute, and who keep a record of where the junior and the machine disagree. That record is how judgement gets built.
Third, teach judgement explicitly. Schools and universities worry about students outsourcing essays. The answer is not to ban the tools but to assess what the tools cannot supply: the ability to explain why an answer is wrong. A student who can catch the machine's mistakes has learned more than one who never touched it.
A choice, not a forecast
The future of AI and work is not something that will simply happen to us. Whether AI ends up amplifying people or hollowing them out depends on decisions employers, educators and governments make over the next few years. Collective AI is a choice: to treat the machine as a way of putting the knowledge of the many into the hands of each person, while making sure each person still learns to think for themselves.
An experienced professional with an AI assistant is a formidable thing. Our job is to make sure there is a next generation of experienced professionals to follow.