The merge-in-turn problem: why UK Government’s AI strategy is fundamentally broken

Here in the UK, when two lanes become one on the road, we call this a ‘merge in turn’ junction.

Most Brits ignore this advice and instead see ‘merge in turn’ as an opportunity to demonstrate their fondness for the art of queuing and sitting in traffic. Should you be as bold as to use the junction as intended, you will be met with the sort of rage and vitriol that would make the Daily Mail blush.

In America, this isn’t a problem. They call it a zipper.

Words matter.

When Kanishka Narayan, UK Minister for AI, took the stage at the Founders Forum earlier this week to set out his vision for the future, Narayan framed AI – as many do – as a tool to be wielded.

In doing so, Narayan belied a critical misunderstanding of what AI actually is – and it is one which could have disastrous consequences.

The elephant in the room

AI is not something you hang up in your shed and swing a few times a year in your garden.

AI, as the name demonstrates, is an intelligence to be applied.

It is human knowledge with a Bayesian probability layer on top: it is a library which talks back and, increasingly, has the capability to apply the knowledge within without human hand holding.

When ChatGPT 3.5 launched in 2022, it could work on a problem continuously for 36 seconds. This time last year, Sonnet 3.6 could run for an hour. Last month, GPT 5.2 could run for as long as 6 hours.

As of last week, Opus 4.6 is capable of continuously working on a problem for two weeks.

Does this sound like your garden variety sledgehammer to you?

A core rationale for this mismatch of terms are the people who are building AI – engineers. While it’d be unfair to say engineers do not apply intelligence, it would be accurate to say that tools are terminology associated more strongly with the trade.

Unfortunately, this description carries weight with the rest of us and has led to widespread confusion and miscomprehension as to how we should interact with AI.

The difference maker is our approach

Narayan clearly looked to draw favourable comparisons between his vision and Harold Wilson’s one for the Open University when he announced: “our programme to support AI skills is now targeting 10 million workers – almost a third of our workforce – skilled in AI by 2030.”

Narayan mentions Wilson’s role in the foundation of the Open University, but neglects that the OU fundamentally reinvented how we educate to be successful. As such, the AI Skills Hub the UK Government has deployed to deliver these skills will not meet the challenge.

There is a plethora of problems with the courses presented.

First, these courses reinforce cognitive offloading – ie. handing everything to the machine – which not only has been shown to erode cognitive ability and skills, but rob people of the very thing Narayan wishes to create: human agency.

Second, traditional course structure simply does not match the speed of AI. While Euan Blair’s AI apprenticeship company Multiverse is admirable, the courses it provides take up to a year.

To put this in perspective, this time a year ago, the big thing was Deepseek. Now, we are witnessing the death of a profession thought as a safe path to a stable future – software engineering - while $285 billion got wiped out in a day this week by a blog post and a 200-line markdown file.

In the world of AI, a yearlong course will be mostly defunct in its first week, and entirely obsolete by the time it finishes.

Finally, the course providers themselves are IT gurus, ML honchos, and the very engineers who can’t tell a tool from intelligence.

These are the wrong skillsets to teach AI.

We are not teaching people how to use a word processor.

We are coaching them on how to interact with intelligence.

In Narayan’s speech, he asks whether technology is wielded by humans or if humans serve machines. The 20-minute prompting courses on the AI skills hub answers that question – just not in the way he intended.

Let me show you what will.

Humanity’s firmware update

At the start of 2025, I found myself at a crossroads. Recently diagnosed with AuDHD and made homeless by my ex-fiancé, my personal troubles had manifest into problems at work which threatened my employment. Fortunately, I had a great boss who showed me compassion and gave me a chance to turn things around.

I knew AI could support this endeavour but had no idea how. To navigate this uncertainty, I turned to my ancestor – Thomas Bayes.

His theory on probability, after which this newsletter is named, is core to how AI actually works.

One of the central tenets of Bayesianism is that you actively seek out evidence to support your potential routes forward and quickly dispose of previous beliefs which the data shows are wrong. At this time, research was beginning to show that the cognitive offloading model enshrined in the UK Skills Hub was incorrect. Consequently, it was the first to go.

Instead, I gravitated toward the idea that AI is a mirror – it can help you cut corners but can also help you achieve exponential impact.

I drew on existing knowledge I had from psychology and neuroscience, and began reading any and all papers I could find on technology adoption and AI enablement, while throwing a heavy dose of stoic philosophy at it for good measure.

Within a couple of months, I had developed a system of symbiotic metacognition with AI which allowed me to maximise the strengths of AuDHD and mitigate the weaknesses.

At work, we were discovering that all the barriers we encountered were people-shaped problems – exactly the sort of thing my methodology was designed to overcome.

I was handed the baton, and added my communications background, writing skillset, and business methodology such as the Toyota Way to my approach.

It was an incredible success.

The ROI from my enablement programme recorded an astonishing 286%. For our ChatGPT cohort, it added the equivalent of 72 additional FTEs at the company. I had people I’d never met coming up to me and shaking my hand for changing their life – something I was not accustomed to as a comms man.

Central to the approach is empowerment.

People right now are terrified. Leaders have been caught out and fear being shown up by their lack of AI knowledge. Managers don’t have the vision from on high and are dragged down by anxieties from below. Those at the bottom of an organisation fear for their very livelihoods.

I was the guy looking people wracked by fear square in the eye, giving them an AI-shaped crowbar to take to their biggest problems, and returning to them agency and a way to chart the uncertainty.

I can achieve this impact in a short conversation without any powerpoints because I’m not teaching a course, I’m creating a cognitive shift in how people perceive AI. Once you demonstrate that AI is an intelligence that you collaborate with – that the best teacher of AI is AI itself - the penny drops and learning becomes self-reinforcing.

The road ahead

This approach matches the speed and scale we’re now seeing in AI – it is the route to becoming AI native – and is needed now more than ever.

In AI safety circles, there is a concept known as takeoff – the point at which the AI we know today begins to rapidly gain momentum towards Artificial General Intelligence, or AGI. This is the point at which AI surpasses man as the most intelligent entity on the planet.

The exponential gain in autonomous problem solving I mention before is one of many metrics we can use to track takeoff. My current assessment is that if we are not already in takeoff, we are accelerating down the runway, seatbelts fastened, feeling the G-force as the engines roar.

We, as a society, are nowhere near ready for the implications of AGI, which makes Narayan’s vision and goal all the more urgent.

The winners in the AI race won’t be the companies who wait for AI to be ready for their systems – it won’t. It’s like waiting for Microsoft to push an OS to your typewriter. The 5% who are killing it are making rapid gains because they have transformed their systems to be ready for AI.

We, as individuals, must do the same, otherwise we’ll all be stuck in traffic while AI overtakes us.