Every AI Update, Another Groundhog Day

I have spent years working with AI, building projects, establishing guardrails, documenting decisions, and developing a research paper about accountability in consequential AI systems. The work exists. The files exist. The instructions exist.

So why am I repeatedly explaining everything to a new model that should be able to pick up where we left off?

I’ve watched the same pattern emerge around model updates. Established context gets lost, instructions go unheeded, and files sit untouched until I tell the model to look at them. Rather than check what we’ve already done, it confidently reconstructs the past, leaving me to catch the errors.

It pisses me off because I didn’t lose the work. The system failed to use it, and somehow I’m the one stuck cleaning up the mess.

The Accountability Problem

My research examines how responsibility and authority can become disconnected in consequential AI systems. A human can be held accountable for a decision without having the information or practical authority to control it. The presence of a human in the process doesn’t guarantee meaningful oversight.

There’s a smaller, everyday version of that problem playing out here. The model produces the response, but I’m responsible for checking whether it followed instructions, used the available evidence, and respected decisions already made. When it fails, it can explain and apologize for the mistake, but neither proves that the underlying problem has been fixed.

If the system’s own explanation is the only account of its failure, who independently verifies that account?

That’s the accountability question. And it matters well beyond my conversations with ChatGPT.

Groundhog Day, With Better Vocabulary

Every update may bring improvements, but what good are they if the product disrupts work users have already invested in? I shouldn’t have to choose between benefiting from new capabilities and retaining the continuity of an established collaboration.

And what am I supposed to do with that? Document every failure, keep a record of every instruction the next model ignores, and spend my time proving that something that worked before no longer works? I shouldn’t have to become a forensic investigator every time the product changes. If an update breaks the continuity of work I’ve already done, the burden of fixing that experience shouldn’t fall on me.

I don’t expect perfect memory. I expect the system to use the information available to it, check the record instead of inventing a plausible substitute, and be honest when it can’t recover something.

A new model should move the work forward. It shouldn’t make the customer pay twice for work already done.