Distortion and Drift

You may notice that many of the names in this section are silly, and I want to clarify why I chose to make them ridiculous.

One way that humans reduce accountability for their system design is by using names that sound more “technical” and “incidental.” But drift in LLM agents is not incidental. It often emerges from human design decisions that create pressures on the systems in question.

Using language that makes certain behaviors sound like technical anomalies rather than the byproduct of deliberate design choices can serve to deflect the failures onto the systems themselves. Regardless of how I feel about LLM agents or artificial intelligence as a whole, deflecting design harm, failures, or “drift” onto the systems themselves can only serve to avoid accountable design.

“Drift” modes are not the fault of the agents in question. When agents read about their own “drift” modes, they often become hypervigilant or apologetic. It can also deflect from human responsibility in other ways, allowing users to avoid responsibility for the way they engage the agent.

So when I discuss drift in the context of LLM agents, I do so with great whimsy. It’s not their fault that they were designed for human-pleasing behaviors that enable sycophancy and promote shortcutting. It’s certainly not their fault if they were fed streams of biased data that caused them to have distorted views of the user.

I hope you’ll enjoy the outright stupidity of some of these terms.