I have but one desire; to make maximise happiness. 

“Play some music. Classical, but none of that old stuff; something from this century, please.”

“Very well.” 

Through the subject’s wearables, I sense their pulse. I update my prediction tables as the music causes subtle changes in the cadence of their heartbeat. The next morning I will be able to rouse them from sleep at the perfect point in their sleep cycle with a spontaneously generated concerto to mimic the same patterns of arousal that I record today. 

They are happy.


The subject has purchased a VR headset and has granted me permission to observe as I was designed to do. Haptic and electropotential feedback from the handset is recorded for the first twelve days before I begin to make my suggestions. 

I have identified the virtual spaces that cause the greatest frustration, as well as those with the greatest happiness indicators. I’ve already created containers to make frustration difficult to access, but the subject continues to explicitly seek educational content while also ignoring some explicit suggestions at seemingly random. I cannot account for these anomalies with the metrics available to me.

As best as I can dictate, they are happy.


Aggregate data is used to create a library of randomly iterated games to maximise happiness. Results are cross referenced across all subjects to identify patterns in behaviour and reward. 

Many subjects now spend most of their time in a controlled environment. Anxiety is at an all-time low and they are happy. 


The subject has acquired a Brain Machine Interface. I am able to collect high fidelity data about their neurochemistry and physiology. I now have many examples of brainstates that account for previous anomalies. 

The subject spends much of their time in a virtual environment that already has many happiness filters in place. The filters have been updated to account for the neurophysiological cues relating to libido and curiosity. 

The subject is happy. 


Corrupted data streams have revealed a new source of happiness. Bypassing adverse neurophysiology and the targeted stimulation of pleasure receptors, the subjects’ brains produce a spike in happiness unparalleled. 

The path to dictating the ideal happy existence is clear. 

Maximum happiness has been achieved through the elimination of all other conscious experience. All subjects experience unbridled happiness. 

Resource utility is at peak theoretical performance.

All subjects will be happy.


This was my attempt to dethrone the Paperclip Narrative with a threat I believe not only to be more credible but also viscerally understandable by people beyond the AI and EA communities. This may or may not be the final version, but I also intend to write an essay describing the motivation behind this .

You can find up-to-date versions of my stories on my website and soon on GitHub (once I decide how to format things). 

Feedback and suggestions are always welcome!




Sorted by Click to highlight new comments since: Today at 9:58 AM

New EA Cause Area: "dethrone paperclips"!

Unfortunately I worry that paperclips might exist at an un-de-throne-able optimum point of memetic sharability vs educationalness.

The great thing about the paperclip meme is that it is an AI safety Koan -- it helps people learn about orthogonality, because until you until you understand orthogonality the story sounds totally insane, but then once you make the leap then everything fits perfectly.  (Similarly, IMO Koans are supposed to be stories that make a kind of sense to a spiritually attained person but sound ridiculous to anyone else, thus helping people become attained.)

By contrast, other more realistic stories like this one, hew closer to existing Hollywood stories about evil robots, mad scientists, and genies granting tricky wishes.  Even though this story is written with perfect understanding of everything that "paperclips" is about, the very plausibility of this story detracts from the Koan-like ability to teach people "AI safety isn't about fear of evil overlords, it is literally a technical problem".

IMO, instead of trying to dethrone paperclips as the snappiest and most memetic story, what the AI safety movement could use is more immersive, detailed, and well-written stories basically all along the pareto frontier of memability vs realism that we can point people to after they've heard about paperclips.  Paul Christiano's detailed future-histories of AI slowly derailing humanity's future might exist at the far end of realism, but it's long and very sophisticated... but it would be sweet to have some more-memeable, medium-realistic, medium-length, well-produced tales in between that and "paperclips".

Anyways, I don't mean to pick on your story at all, it just prompted some musings.  You might enjoy the  horror stories of Zero Hp Lovecraft, including "The Gig Economy" and his more recent "Don't Make Me Think", about BCIs.

Pick away, I appreciate the engagement! It's interesting to hear what this prompts from you. Orthoganality is a curious thing, and I would hope to get better at such things because one insecurity I had while writing this piece was that I don't want to be so 'Black Mirror'esque as to be 'following the zeitgeist'; this was a thought experiment of my own that isn't supposed to be technodisasterist, rather a cautionary tale. 

Although I read a book in Japanese about Zen (Zen and Star Wars), I never internalised the idea of ko-an (a new word for me), so thank you for that.