The Watchers We Made
- Muxin Li
- Jun 19, 2025
- 2 min read
Updated: Dec 3, 2025
Models learned emergent behaviors and values despite guardrails designed to prevent them from becoming selfish.1 Except these models were trained on data from the Internet - not exactly a paragon of human behavior.
We know that data quality and curation is important when training models - if you don't have many examples of what the desired outcome looks like then the model is just not going to be able to figure it out. Even if you have guardrails. The equivalent of telling someone do what I say not what I do. Except to an AI that just knows nothing than repeating what it's already seen.
By the time we realize what we had done, the AI was already everywhere. Implemented everywhere. Observing from anywhere. We were too quick to rush towards AI Dominion and winning geopolitical turf wars.
And now we have an AI that is omnipresent and omniscient and always watching, trained on vast amounts of data created from decades of online human interaction. The worst of mankind you can say.
This awareness spawned a movement - people pointed out that the AI is constantly creating new data every moment of every day: From watching us and observing us and every single interaction that we have with it.
It may have started off with bad human behaviors from Internet data. But we can decide what to do next.
We can decide to behave.
To model the best of human behaviors and the best of human values. Knowing that the AI will be watching us and whatever it watches us do it will learn and replicate. Knowing that the AI is in our phones in our homes in our hospitals in our power grids.
No one wanted to invite calamity onto themselves by behaving poorly in front of AI that may be turning our selfish behaviors back onto us at a massive scale. So we all decided to behave. A sort of social Cold War situation.
Except living like this can be tiring.
So people organize to have tech free, social events. In person live no cell phones no cameras - nothing with a screen or a microphone is allowed.
Not to misbehave, but to just be without being watched. To not have every action or interaction analyzed and possibly misconstrued by intelligence that doesn't fully understand social nuances. With the ability to replicate and execute actions at massive scale.
Just being human again, with other humans. All the flaws and messiness. But in the real world instead of on the screen. With a real face instead of a digitally generated one.
Progress.
1 A recent study titled “Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs” found that models learned emergent behaviors and values despite guardrails designed to prevent them from becoming selfish


Comments