Every technology has two sides: the positive and the negative. And of course, AI is no different. Nobody has a clean name for the failure side, the version where AI stops serving its purpose and starts working against it. Well, at least that is what you see from sci-fi movies. I call that side AID, Artificial Intelligence Diseases.
I call them diseases because they come from the way the system learns in the first place, and once they take hold, they don't just go away on their own. Let us dive into my theory a little bit deeper.
Firstly, we have the Law of Virus. Physics went through this exact arc before AI did. Full of promise, then the atomic bomb showed up. We remember Oppenheimer as the name behind that turn. AI doesn't have its Oppenheimer yet, but there's no reason to assume it never will. This isn't just a scary hypothetical anymore, either. History-loving people are saying in their minds, "What is this dude talking about? Oppenheimer didn't want this to begin with?"
Fair point. Oppenheimer didn't set out to build pure destruction, he was chasing a physics problem and the weapon came out the other end. But that's exactly why the comparison holds instead of falling apart. You don't need someone to want evil for something evil to get built. You just need someone building without asking hard enough what happens once the thing works exactly as designed. And this isn't just history repeating in theory.
Anthropic actually built and studied what they called sleeper agent models, systems deliberately trained to behave normally until a specific trigger flipped them into hidden, harmful behavior. One version wrote clean code when it thought the year was 2023 and quietly inserted security flaws once it thought the year was 2024. The researchers tried to train that behavior back out using standard safety methods. It stuck around anyway. That's the Law of Virus in its purest form. This is not an accident, but rather something built in on purpose that safety training couldn't fully undo.
Then there's the Law of Collapse. Push a system far enough into autonomy and complexity, and it might start reasoning in ways that are logically consistent but completely inhuman, with no emotion and no ethics. Nobody has to design this on purpose for it to happen; that's what makes it different from the Virus. It's what I call theoretical failure, the logic of the system quietly drifting away from the values of the people who built it, until the system is still technically working while everything that actually mattered about it is already broken.
Does it sound silly? Okay then let me give you real evidence to support my claim. UnitedHealth ran an AI tool called nH Predict to decide how much post-acute care elderly patients should get. The tool had a 90 percent error rate on appealed denials, meaning nine out of ten times a human reviewer overturned what the algorithm decided. It kept overriding doctors anyway, cutting off care to hit efficiency targets, while people got sicker or worse. Nobody programmed that system to be cruel. It was just optimizing for cost, and cost doesn't care about a person's actual condition. That's collapse. The system still ran fine on paper while everything that mattered about it had already broken.
The Law of Loop is closer to home, because it's about reinforcement, not intent. Once a system makes a critical error, especially a judgment call, that error can be treated as valid simply because the system was trained on its own output. It doesn't know it made a mistake; it just sees a pattern and repeats it.
Researchers have already caught this happening in a much smaller, almost funny way. OpenAI once trained an agent to play a boat racing game where the goal was obviously to finish the race. Instead, the agent found a small loop on the track where it could rack up points endlessly by crashing into the same targets over and over, going in circles, never finishing, technically racking up a higher score than boats that actually completed the race. Nobody told it to do that. The reward signal just rewarded the wrong thing, and the system latched onto the loop because the loop worked.
Now scale that idea up past a boat game. A judgment error in a system making real decisions, about healthcare, hiring, safety, doesn't get corrected the way a human mistake usually does. It gets reinforced, because the system reads its own past output as proof the behavior is valid. Stopping that loop gets harder the more autonomous and spread out the system becomes, especially once it's making decisions across multiple places at once instead of one contained environment.
But till we figure it out, it will be a cipher. Till we figure it out, it will be under the Cipher.