Whose future is it anyway? (Part 3: From leverage to lock-in)
In the first two posts, I looked at how AI is reshaping outcomes at a global level.
- Accelerating disruption and opportunity at the same time
- Concentrating capability in a small number of countries and organisations
- Creating real risk of exclusion for those without access
But this dynamic doesn’t just play out between countries, it plays out within them, because even where access exists, control is uneven.
The promise of leverage
One of the most compelling aspects of AI is the leverage it creates.
Individuals can:
- Write code faster
- Analyse data more effectively
- Produce content at scale
- Build products with smaller teams
This is real, it's transformative, it feels like empowerment, and in many cases, it is.
The hidden dependency
That leverage is rarely self-contained.
It depends on:
- External models
- Proprietary platforms
- Paid APIs
- Hosted infrastructure
In other words:
You're not using AI.
You're using someone else’s AI.
That distinction matters, because it introduces a new kind of dependency.
The barriers to entry are not being kicked over
AI adoption also meaningfully moves the barrier to entry in the software development world, and not in the way that you might initially think.
There's lots of talk of AI "democratising" programming, but I'd argue that by creating dependency on specific, commercial tools it's doing exactly the opposite.
For many years arguably all you needed in order to learn to write computer software was access to information and access to a computer1 — you could learn to code on a computer at your local library, reading instructional books from that same library, or accessing tutorials online there, and never having to own a computer yourself.
Whilst there are free tiers for many of the AI tools, they run out of steam very quickly, and charges are consistently moving towards metered usage rather than unmetered tiers as the AI companies seek to make their offering actually profitable rather than subsidised, increasingly over-subscribed, but undeniably popular.
If you want to swerve that, keep your flag nailed to the open software mast, and run local models instead of being in hock to the AI giants like OpenAI and Anthropic or the platform owners like Microsoft or Google then you're going to need some beefy hardware, which has a price tag that comes with multiple zeroes before the decimal point rather than being zero.
So yes, we are now seeing more non-programmers experimenting, learning, and producing code, but that's only "democratising access" in that sense that anyone who can pay can produce code — I'd call that explicitly commercialising access, but given the ongoing commercialisation of democratic institutions (particularly in the USA) perhaps that distinction is becoming harder to identify?
The platform pattern
We have seen all of this behaviour before though.
New technology emerges — it's cheap, accessible, and highly enabling.
Developers adopt it. Companies build on it. Entire ecosystems form around it.
And then, over time:
- Pricing changes
- Terms tighten
- Control increases
- Alternatives narrow
What began as empowerment becomes something closer to dependency, and it is clear to me that AI is following a similar trajectory.
From tool to control surface
AI systems are not just tools, they're interfaces to hitherto unprecedented capability, and whoever controls that interface controls:
- Access
- Pricing
- Constraints
- Visibility
That creates a subtle but powerful shift.
The question is no longer just:
What can you build?
It becomes:
What are you allowed to build, and on whose terms?
The cost of convenience
There is a fundamental trade-off here.
Centralised platforms offer ease of use, rapid capability, and reduced operational overhead.
But they also introduce vendor lock-in, opaque behaviour, and external dependency.
That trade-off is often accepted implicitly, especially when the short-term gains are obvious.
The enshitification risk
Corey Doctorow’s concept of “enshitification” describes how platforms evolve through 3 key phases:
- Be good to users
- Be good to business customers
- Be good to themselves
AI platforms are not immune to this pattern, and the incentives may accelerate it significantly.
When it comes to AI, the cost of switching is not just technical, it's cognitive, behavioural, and organisational.
Capability vs ownership
This brings us back to a recurring theme.
Capability is increasing, but ownership is concentrating.
You may be able to do more than ever before, but you may not control the tools, the data, or the economics.
That's not a neutral position.
A familiar outcome
If this pattern holds, we end up with something familiar:
- A small number of dominant platforms
- A large number of dependent participants
- Increasing asymmetry in power and value capture
The difference is that AI operates at a deeper level, not just distribution of software, but rather distribution of intelligence.
The bridge to something bigger
This is where the story shifts again.
From access, capability and platforms to power, control, and concentration.
If a small number of organisations control the most powerful AI systems, then the question is no longer just about dependency, it becomes one of influence, and potentially, dominance.
What comes next
In the next post, I’ll look at that risk directly.
Not in terms of runaway AI systems and killer robots with Austrian accents, but in terms of something more immediate — the possibility that AI accelerates the concentration of economic and decision-making power in ways and at a pace and completeness we have not seen before.
The real question may not be how or whether AI changes the world, it’s who ends up owning the systems that do.
Footnotes
-
I am, for example, completely self-taught. Admittedly I did have a personal computer when I started building web pages for fun, and then for a living, but productivity was determined by my own labours rather than being replaced by a commercial tool that can rapidly create the same kind of functional but fallible code that I used to type by hand. ↩