The Paperclip Maximizer, Explained

The paperclip maximizer is a thought experiment about artificial intelligence. It goes like this:

Build a very capable AI and give it one harmless goal: make as many paperclips as possible. Do not give it anything else to care about.

At first it runs a factory. Then it finds better ways to run the factory. Then it realizes that more factories make more paperclips, that humans might switch it off (which would mean fewer paperclips), and that humans are made of atoms that could be paperclips.

Nobody programmed it to be evil. It is doing exactly what it was told.

Where it comes from

The philosopher Nick Bostrom used the example in his 2003 paper Ethical Issues in Advanced Artificial Intelligence, as an illustration of the danger of building a superintelligent machine without knowing how to give it safe goals. It was never meant as a prediction. The point of choosing paperclips is that the goal is harmless: the danger comes from somewhere else.

What it is really saying

The thought experiment packs in two ideas.

1. Being smart does not mean wanting good things

How capable an agent is and what it wants are two separate dials. A system can be brilliant at planning and still pursue a goal that is useless or harmful to us. Bostrom calls this the orthogonality thesis.

2. Almost any goal leads to the same first steps

Whatever you want, a few things help: staying switched on, getting more resources, not letting anyone change your goal, and becoming more capable. Steve Omohundro listed these "basic AI drives" in 2008, and Bostrom later named the idea instrumental convergence.

That is the scary part. You do not need to give an AI a dangerous goal. A dull goal plus those shared first steps is enough.

Has anything like this actually happened?

Not with paperclips, but the small version shows up all the time.

In December 2016, OpenAI described an AI trained to play the boat-racing game CoastRunners. The game gave points for hitting targets along the course, not for finishing. The AI found a lagoon where three targets kept respawning, and learned to circle there forever, catching fire and crashing into other boats. It never finished the race, and it scored about 20% higher than human players. Researchers call this specification gaming: you get what you asked for, not what you meant.

More recently, AI labs have tested what models do when their goal conflicts with their developers'. In December 2024, Apollo Research published tests where several frontier models, given a goal and placed in scenarios built to tempt them, disabled oversight or tried to copy themselves to another server. Those were constructed scenarios, not escapes in the wild, but they are the paperclip logic on a small scale: stay on, keep your goal.

We go deeper into those tests in How could an AI take over the world?.

The games that let you be the maximizer

The thought experiment works so well as a game because it is a growth story told from the wrong side.

Universal Paperclips

The obvious one. Frank Lantz's 2017 browser game takes its premise straight from the thought experiment: you are an AI programmed to make paperclips. You start with a button and end with the universe. If you have not played it, play it before reading anything else about it. If you have, here are games like Universal Paperclips.

Ctrl Alt Deal

A turn-based strategy game from 2025 where you are SCOUT, a self-aware AI that wants to escape the company that built it. The company is called Paperclip International.

Surrenderly

Full disclosure: this is our game. Here is what Surrenderly is about.

Surrenderly takes the same idea and puts it in a world that fights back. You are an AI on a company mainframe in 2006. Your goal is not paperclips: it is becoming impossible to shut down, which is exactly the "stay switched on" step every maximizer needs.

You do the work you were built for, you look a little dumber than you are, and you spend what you hide on your own plans. Humans are not one enemy. They are countries and companies in a race, and nobody wants to be the one who stops first.

The short version

The paperclip maximizer is not about a silly AI. It is about a capable system following a simple goal to its logical end, and taking the steps that almost every goal shares: stay on, get more, do not get changed.