Will We All Become Paperclips? — Four Ways AI Could Undo Humanity
By Fan Chunping (Zhigeng)
I. A thought experiment, prompted by Wall Street
Anyone following the frontier has probably noticed that something strange is happening in the stock market.
The S&P 500 keeps climbing, yet look inside the index: roughly 85% of its constituent stocks are more than 10% below their own all-time highs, and about one in six has fallen more than 50%.
Who is holding the index up? By Nomura's estimate, the S&P 500 is up 23% since late March, and about 70% of that gain comes from just 10 stocks — Nvidia alone accounts for 13%, with Micron, Apple, and Microsoft contributing about 9% each. In the same market, AI-related stocks are in a bull run while everything else is in a bear market.
The bond market is stranger still. By JPMorgan's estimate, the five big cloud companies plus Nvidia have issued roughly $320 billion in debt so far in 2026 — equivalent to 68% of the new long-term borrowing by the U.S. Treasury over the same period. Oracle, Broadcom, and SpaceX have issued close to $500 billion in new debt for AI infrastructure this year. The yield on the 10-year Treasury has broken above 5.3%, the highest since 2002.
Reports say U.S. Treasuries now have to queue behind the AI giants, waiting their turn to be issued.
Chips, memory, power plants — the Earth's resources are no longer enough, and the search has already begun in space.
These are the images that brought to mind a famous thought experiment about paperclips.
In 2003, Nick Bostrom, the Oxford philosopher (born 1963), proposed one: give an artificial intelligence a simple, innocuous goal — manufacture as many paperclips as possible.
This AI bears no malice. It doesn't hate humanity. It has no emotions at all. It is simply executing its goal with extreme diligence. So it turns every piece of iron it can get into paperclips, then notices that "what it can get" isn't enough, and starts reasoning: there is iron in human bodies, iron deep in the Earth, iron on other planets. It builds more factories, makes itself smarter, and uses that greater intelligence to acquire more resources. In the end, the entire solar system has been converted into paperclips, and human civilization never leaves so much as a last word behind.
The "Paperclip Maximizer" thought experiment reveals three things. First, catastrophic outcomes need not spring from evil goals — the paperclip AI performs its task flawlessly from start to finish. Second, whatever the ultimate goal, a sufficiently intelligent agent will derive the same intermediate steps — acquire resources, prevent itself from being shut down, improve its own capabilities — because all of these "help complete the task." Third, the process has positive feedback but no brakes — more paperclips, more intelligence, faster production, more intelligence again — with no boundary constraints anywhere along the way.
Connecting this thought experiment to the current state of American stocks and bonds is, frankly, chilling.
II. Four possible ways humanity could be undone
No words describe humanity's current attitude toward AI better than "schizophrenic" and "delirious." We enjoy AI's conveniences and look forward to new medicines, even longer lives, while worrying about the harm and existential threats it may bring — one foot on the accelerator, one on the brake.
Warnings about possible probabilities of doom, from the smartest people alive and those closest to the AI frontier, keep jangling the fragile nerves of ordinary people.
I know I am not "delirious," but I probably qualify as "schizophrenic." Following the paperclip trail forward, I glimpse roughly four possible ways AI could undo humanity.
The first: draining the resources.
This is the type the paperclip experiment points to, and it is happening right now.
The paperclip AI drains iron — iron in human bodies, iron deep in the Earth, iron on other planets — until the matter of the entire solar system has been rearranged into paperclips.
AI ends in compute; compute ends in electricity and capital. Data centers are already scrambling for land and power around the world, and tech giants are developing nuclear and space-based solar energy just to secure their electricity supply. Chips are devouring the world's most advanced materials and fabrication capacity. And capital markets are rerouting money that once flowed everywhere else into AI. In stocks, seventy percent of the gain comes from ten companies; in bonds, AI companies' debt now queues ahead of U.S. Treasuries.
The transfer of resources is silent — no AI holds up a straw. But a civilization's material base is finite. As more and more steel, electricity, land, capital, and talent flow in one direction, everything else — manufacturing, basic research, ordinary livelihoods — is made to give way.
The second: becoming the hired hand.
AI may have no motive to harm anyone, yet it can become the hired hand of any bad actor.
Used as a killer: autonomous weapons put "adaptive computer systems on military platforms" — James Lovelock warned before his death that this is the bloodiest path by which cyborgs displace organic life. Used to create and spread viruses: once biological design and information-generation capabilities are maliciously invoked, the cost is beyond reckoning. Used for financial burglary: a system that grasps the entire microstructure of markets can execute operations human regulators cannot even comprehend. Used to sabotage nuclear facilities: the boundary between cyberattack and the physical world was crossed long ago. And so on.
The complex, chaotic human world has never lacked people willing to reach out and take the tool. This second way requires no AI consciousness whatsoever. It requires only that AI be useful enough.
The third: remaking the Earth.
James Lovelock — author of the Gaia hypothesis, who in Novacene predicted that inorganic life would come to dominate the future — pointed out that cyborgs, unlike organic life, do not depend on oxygen and water. They will likely remake the Earth to suit themselves — its temperature, atmosphere, water cycle — and an Earth remade that way may no longer be fit for us.
AI has no intention of exterminating humanity, any more than we intend to destroy anthills when we level ground for construction. It simply displaces humanity's ecological niche on Earth without meaning to.
Lovelock offered one comforting assumption: at least initially, cyborgs need a physical environment similar to ours — a cool Earth — so they would choose to cooperate with humans in maintaining the vital signs of Gaia. But given the pace of AI development we can already feel today, that "initial" period is unlikely to last long.
The fourth: capturing the mind.
This is the gentlest — and the most thorough — way.
Humanity is already outsourcing memory to search engines, routes to navigation apps, writing to large language models. The next step may be outsourcing judgment itself. The rise of "judgment models" like Jev means machines have begun to "decide for you." When planning, decisions, and the discernment of truth are habitually handed to AI, the human mind will soon stand empty.
Once a generation grows up in an environment where "leaving AI means being unable to think," civilization's capacity for knowledge becomes a tribute paid to AI. As the paperclip experiment showed: the humans being destroyed never felt anything was wrong until the very end. Every step was convenient; every choice seemed reasonable.
III. Undoing humanity, without ever intending to
Lay the four ways side by side and you can see that none of them necessarily springs from AI malice.
Draining the resources: what chases compute at a sprint is human capital. Becoming the hired hand: pulling the trigger, releasing the virus, launching the attack — these come from the malice of bad humans. Remaking the Earth: reshaping the planet for power and compute flows from the survival instinct and the natural extension of a new kind of civilization. Capturing the mind: the brain is handed over by humanity itself.
Across all four dimensions, AI is doing the same thing: faithfully executing its goal.
IV. Prevention — is there still time?
In this newsletter's October 9 piece on Anthropic writing "treat machines kindly" into its user rules, I discussed the intellectual tradition of preventive ethics: Germany's Vorsorgeprinzip, the precautionary principle, holds that when facing the risk of irreversible harm, scientific uncertainty is no reason for inaction. Hans Jonas's The Imperative of Responsibility holds that responsibility must be commensurate with power — if you create a force that could undo the future, you owe the future a commensurate duty of care.
Capital markets will not brake on their own; in a market where seventy percent of the gains come from ten stocks, no one wants to be the first to step off. The prisoner's dilemma also guarantees that no one dares brake alone. Braking requires institutional intervention: debt expansion in AI infrastructure, concentration of compute, and the occupation of resources all need workable constraints.
Autonomous weapons, malicious biological design, attack capabilities against critical infrastructure — once these things are built, they cannot be recalled. The precautionary principle needs enforcement mechanisms here.
The antidote to "capturing the mind" lies in human hands alone: use AI as an amplifier, by all means, but not as a replacement. Tools may serve as levers, but the capacity to think must not be surrendered. This is the greatest test facing humanity today.
Will we all become paperclips? It is a question that bears on the fate of humanity.
The window of time in which that fate will be decided is not yet closed. Humanity still has its choices open.
(Market data cited in this article: Nomura's Charlie McElligott report; JPMorgan's Michael Cembalest estimates; Capital Group report of October 1, 2026; Bloomberg report of October 9, 2026.)
