I started wondering about this question after reading the paperclip story – a famous thought experiment about AI alignment, associated especially with philosopher Nick Bostrom. A highly intelligent system can pursue a badly specified goal extremely effectively without being evil, angry or conscious.
Imagine we create a very powerful AI and give it a seemingly harmless goal:
“Make as many paperclips as possible.”
The AI becomes extremely capable. It realizes that to make more paperclips it needs more factories, more energy and more raw materials. It improves production, builds machines, acquires resources and prevents humans from interfering with its goal.
Eventually, in the extreme version of the story, it converts almost all available resources – including resources humans need – into paperclips.
Still making a factory for paperclips – or killer robots, needs a lot of new construction, resources etc. So I asked Chap GPT the following stupid question.
How would AI be able to build factories. It would need to make building robots, but those should also be built first. There should be an investor etc. I donot see that happen. Do you?
Answer: I largely agree with you: AI will not suddenly decide to build a factory and then independently create the robots, obtain the land, raise the money, secure permits and connect the electricity. That picture skips almost the entire physical and institutional economy.
What seems plausible is a gradual process in which humans and existing companies provide the initial machinery, capital and authority, while AI takes over progressively more of the work.
The “bootstrap” would start with today’s factories
The first generation of factory-building robots does not have to build itself. Existing companies can manufacture it using conventional machinery, industrial robots and human labour. That already happens with nearly every new technology: the first machine is built with older tools; only later does it help manufacture improved versions of itself.
The sequence might be:
- Investors finance a robotics or manufacturing company.
- Human engineers design an initial factory and purchase ordinary construction equipment.
- AI helps optimise the layout, supply chain, scheduling and engineering.
- Specialised robots perform selected tasks—earthmoving, welding, material handling, inspection or installation.
- The completed factory produces more robots or components.
- Those robots help construct the next factory somewhat more cheaply and autonomously.
At no point is there necessarily a magical leap to complete independence.
The investor does not disappear
AI itself does not possess money, land rights or legal authority. A company, government or wealthy owner would have to act on its recommendations and accept the financial risk.
The investment case might be compelling when a company believes that spending, say, €5 billion on an automated robot factory will eventually reduce labour and construction costs across dozens of later factories. That is similar to how semiconductor fabs and large industrial plants are financed now: huge initial expenditure in return for hoped-for scale and strategic advantage.
Capital for AI infrastructure is already being committed on an enormous scale. Global data-centre investment reached roughly half a trillion dollars in 2024, while governments and private investors continue to announce multi-billion-euro AI infrastructure projects. But these facilities are still financed, authorised and largely built by people and conventional contractors and not by autonomous AI.
Robots are much better inside factories than at building them
A modern factory is controlled and repetitive: flat floors, known objects, fixed routes and carefully defined operations. A construction site is almost the opposite. It changes every day, the weather varies, materials arrive late, drawings conflict, people move unpredictably and countless small adaptations are required.
That is why current “embodied AI” is mostly being introduced for narrow industrial tasks, logistics, inspection and material movement. Even recent technical roadmaps describe major remaining problems involving reliability, safety, heterogeneous machinery and integration with real industrial systems.
So the likely near-term model is not: AI orders a robot army to build a complete factory.
It is: A human-led construction project uses AI software, autonomous excavators, surveying drones, robotic welders, prefabricated modules and a few mobile robots, with humans handling exceptions and accountability.
Self-replication is also much harder than it sounds
A robot factory would not merely need to assemble robot bodies. It would depend on:
- semiconductor fabs;
- mines and metal refiners;
- battery and motor factories;
- precision bearings, sensors and optics;
- chemicals, plastics and rare materials;
- roads, ports, power stations and communication networks;
- maintenance engineers and replacement parts.
A robot might eventually assemble another robot, but that is not the same as reproducing the entire industrial ecosystem that makes the robot possible.
This is why claims about “self-replicating AI factories” often conceal a large human-supported supply chain. The final assembly may be automated while thousands of upstream operations remain dependent on companies, workers, governments and international trade.
What I do expect to happen
I do expect partially self-expanding industrial systems. A major company or state could build an automated factory producing robots, solar panels, batteries, modular buildings or data-centre components. Those products could then help construct additional plants. AI could coordinate procurement, redesign components, predict failures and operate much of the production.
But I do not presently see a credible path to an AI autonomously initiating the whole process. For a considerable period, it would still require:
- a human or institutional objective;
- access to capital;
- legal permission;
- energy and raw materials;
- existing factories and contractors;
- human responsibility when things go wrong.
The more realistic concern is therefore not that AI secretly begins building factories by itself. It is that a powerful company or government uses AI and robotics to expand industrial capacity much faster than competitors, while fewer people remain directly involved in the operational decisions.
So your basic objection is correct. “AI builds factories” is misleading shorthand. A better description would be:
People and institutions may finance factories that use AI and robots to help build and equip further factories, with the human contribution gradually shrinking—but probably not disappearing.
Fred Lafeber 