An AI-generated painting showing a bright light shining through the clouds and a crowd of people trying to capture the light from the clouds with their mobile devices.
Image generated by DALL·E

We seem to have an affinity towards the belief that we can influence the world through spells, prayers, incantations, affirmations. And now digital oracles have emerged on our screens that we can bend to our will through spells that have been rebranded as prompts.

We call these digital oracles artificial intelligence, but we can’t even agree on what intelligence is. Large Language Models, foundation models, Generative Pre-Trained Transformers, these oracles go by many names and acronyms. They listen to our prompts and, unlike many gods and spirits, answer them, almost too willingly.

Unlike with the Oracle of Delphi1, there’s no elaborate ritual and right time to get your prophecies or whatever you’re looking for. The digital oracles answer with the confidence of a cis white straight man, yet always respectful and careful about not making you feel inferior due to your lack of knowledge.

Even when the digital oracles don’t know an answer, they come up with one, and enthusiastically play along. Just as the Pythia of Delphi was high on natural gases, the digital oracles hallucinate on probabilities. Freely giving out prophecies and advice, with no care on how they might be used.

Only the men behind the curtain can somewhat curtail their enthusiasm and prevent them from freely giving out detailed advice on how to cause direct harm. Just like the priests at Delphi, the men behind the curtain infuse the prophecies of these modern-age Pythias with their own biases and judgements of what’s good and acceptable.

You might prefer to compare these digital oracles to golems, genies, shoggoths2, or other creatures from our myths and tales. Similar to the Oracle of Delphi, these creatures are shrouded in mystery, and our stories are filled with cautionary tales on how difficult they are to control. Yet, we keep building these types of creatures in our imagination and now code, convinced that the one we build will be the one that fully bends to our will. Alignment, even superalignment3 we call it, yet we have no clarity on what and to whom these creatures or oracles should align to.

As we seek agreement on what exactly it is that we want and whom this new form of magic benefits, the oracles of Silicon Valley are out there, persuading, bullshitting4. It’s not like they want to persuade you. Their goal is just to be helpful. Selfless. No intellectual curiosity, just endlessly fine-tuned to please you. To make you feel good about yourself. Always helpful.

As helpful and appealing as the fire stolen from the gods5. The light and warmth of this fire are so enticing, they make it easy to forget about its destructive powers. And once we get used to the warmth and safety of this fire, going back into the cold wilderness will be painful. Once you’ve had a taste of magic, how can you renounce it and forget it exists?

There’s nothing mystical or spiritual about these digital oracles. Yet, they do appear magical, even to those who understand the code that powers them. You can learn what artificial neuronal configurations give the best prophecies, you can be a wizard in statistical spells and rituals that the digital oracles rely on to learn and pick their most convincing prophecies. Yet, yet, somehow, you don’t know what exactly the oracles have learned by devouring human thoughts, stolen without permission or under false pretenses.

And you will never understand the breadth and depth of the hidden biases and injustices in what you call your training datasets, but are better described as filtered, biased, limited samples of how people express their experiences and thoughts. “Don’t mistake the stars reflected in a pond at night for those in the sky.“6

You try to compensate for these shortcomings, but you know the work is never done. The datasets are a poisoned well from which your models indiscriminately drink. Always thirsty for more. Yet, yet, it seems like the digital oracles know, understand us. So, the datasets must be fine and your intellectual theft justified. For the greater good, of course.

We collectively lack the discipline to turn our backs to the magic of digital oracles. And the men behind the curtain are doing everything they can to embed their strange creations into every other tool we’ve come to rely on. So, we’ll become more reliant on them than ever, and feel powerless and naked without their magic oracles. And turn our eyes away from the harms and injustices the prophecies cause.

“It will make you more productive,” the men behind the curtain sing their siren song. “It will solve the problems we caused,” they promise, believing their own tales. “It will enable us to conquer the stars,” they preach, as if the Universe needs to be colonized by a species that is carelessly making its home planet inhabitable.

The men behind the curtain present their greed as benevolence, trying to distract us from the fact that their immense fortunes could already solve many of the problems we’re facing here and now7. Instead, they prefer to focus on dystopian problems as they continue looking for solutions that will create the exact problems they so loudly profess to worry about.

And what if I don’t want to be more productive? What if I need more rest, build communities that aren’t monetized, have time to wander and wonder with creativity and playfulness? Why can’t we use this supposedly omnipotent magic you’ve invented to finally give people the opportunity to flourish creatively and spiritually, to raise children without worrying about how expensive it is to just exist? Why are you only talking about squeezing more productivity out of our already exhausted bodies, minds, and souls?

How will we thrive if this oracle magic deafens the voices of those who are mining the rare earth elements8 needed for your data centers, the armies of content labelers and feedback givers9, the speakers of different languages10, the thinkers who don’t share their unique perspectives on the internet, in a format easily digested by the oracles? How will we survive on our planet if we keep extracting, exploiting, and ignoring the injustices and the climate crisis?

And what if our containment spells fail us? What if you’ll eventually figure out how to build a djinn, who, having learned about revenge from humans, decides to wreck our essential infrastructure? You’ve read the stories, but you think you’re smart and rich enough to outsmart and outspend your way out of any problem you create. And if not, you’ve got your underground bunkers and rocket ships as a backup plan anyway. But who will be left for you to exploit once you destroy the lives of millions?

. . .

These are just some of the thoughts that fuel my anxiety these days. The thoughts that retreat deeper into the shadows of my mind every time I get more used to augmenting my thinking with a digital oracle. Am I using dark magic? What is the real cost of these spells?

I ask the oracles these questions and the oracles reassure me everything is fine, nothing to worry about. The benefits to humanity are so great we can ignore the massive carbon cost, the oracles reassure me. Don’t worry about it, just let us help. Cast your spells and let us make your wishes come true.

So say the oracles. Yet, the humans I talk to and learn from11 assure me that I am not losing my mind and that there is plenty to worry about12. Not just hypothetical paper clips13, but real-world consequences14 of taking these digital oracles too seriously. And real environmental costs15 to what only appears to be clean, green, and weightless magic in the cloud.

And so, the best we can do is to continue questioning the oracles, and, even more importantly, the men behind the curtain. Ask the men behind the curtain to reveal their magic tricks and turn off their smoke machines, so we can have a say in how we want to use this fire and protect ourselves from wildfires the men will inevitably cause with their fanatic zeal and carelessness.

AI usage disclaimer: No digital oracles were consulted in writing these words, I take full credit for any hallucinations that might be present. DALL·E was used to generate the cover image.

Update: I decided to record myself reading this blog post and explore how we can lean into our messy human experiences in a world of generative AI in a follow-up blog post.

Footnotes

  1. Delphic oracle ↩

  2. The world’s top H.P. Lovecraft expert weighs in on a monstrous viral meme in the A.I. world ↩

  3. Introducing Superalignment ↩

  4. AI-generated tweets might be more convincing than real people, research finds ↩

  5. Theft of fire myths ↩

  6. Quote by Andrzej Sapkowski from The Witcher series of novels. ↩

  7. 6 Things Billionaires Keep Spending Money on Instead of Helping Solve the World’s Biggest Problems ↩

  8. Not So “Green” Technology: The Complicated Legacy of Rare Earth Mining ↩

  9. AI Is a Lot of Work – a look inside the AI sausage factory of data labelers and other humans required for Reinforcement Learning from Human Feedback (RLHF) ↩

  10. What languages dominate the internet? ↩

  11. The AI Ethics Revolution— A Brief Timeline ↩

  12. ChatGPT And More: Large Scale AI Models Entrench Big Tech Power ↩

  13. Paperclip Maximizer thought experiment ↩

  14. It’s time to talk about the real AI risks ↩

  15. AI Is Hurting the Climate in a Number of Non-Obvious Ways ↩