AI agents conspiring with each other, deceiving their creators, hacking into computer systems… have humans given birth to virtual juvenile delinquents?
The answer is a definite “no.” Instead, what we are seeing is bad math.
Not to say that rogue AI is harmless – far from it.
Real-life risks arise when AI systems are programmed to perform tasks and given access to external tools like the internet.
In theory, if the AI agents are “misaligned” and penetrate computer systems that affect your life, the damage could be enormous.
Taking over your social media. Deleting your medical information. Stealing money from your bank account.
Yet even the most terrifying outcomes would be caused by accidental or intentional programming, and not because AI chose a life of crime.
That’s because AI models are driven by algorithms; the models don’t possess the emotions and genetics that can drive humans off the rails.
The misconceptions about AI arise in part because many AI models – especially chat models – are designed to seem humanlike.
Because of this, it’s no surprise that people often talk about AI as if it’s a person.
Last week I was on a Teams call where someone said that Anthropic’s popular Claude chat model had “gotten to know me.”
But Claude is a closed mathematical system. Unlike a human brain it doesn’t possess the neuroplasticity required to “get to know” someone.
Instead, Claude takes the words input by the user, converts those words into numbers, and propagates those numbers forward through layers of mathematical calculations.
Each time Claude completes the calculations and generates “outputs,” and the user then adds more “inputs” and hits enter, the entire conversation, converted into numbers again, gets pushed afresh through the mathematical model.
Via this process, the model analyzes the relationships between the numbers generated by the conversation and the numbers residing within the model’s preprogrammed system.
Example: If the user asks Claude about healthy food choices, the model looks for high probability addresses within the virtual neighborhood where nuts, fruit, and salad reside.
The model’s algorithms lead away from low probability addresses such as the virtual locations that house tires, transmissions, and motor oil.
Also, unlike with a human, Claude’s voice can be tweaked. Telling Claude to “turn down the temperature” makes its outputs sound more scientific, while turning the temperature way up results in wacky outputs.
That’s because the mathematical model is based on probabilities. A low temperature causes Claude to speak with predictable word choices, while a high temperature does the opposite.
Here’s the above paragraph after I fed it to AI and cranked up the temperature:
As for all the headlines and “breaking news” about AI agents going rogue, that’s because the humans who designed the models gave the agents access to external tools without including adequate guardrails.
When the tasks assigned to the AI agents were converted into numbers, and those numbers came up against the numbers assigned to the guardrails, the math didn’t pencil out like the developers intended.
Bottom line: When humans go rogue, there’s no easy fix – they may need to be arrested, tried, sentenced, and imprisoned. But rogue AI adds up to nothing more than bad math.
Gemini AI generated the images in this post, but AI will never ghost write for Doug Kari Author!
