Update

When a technology first arrives, people invest mental effort in learning its quirks, limits, and workarounds. Once they’ve adapted, productivity goes up because the “mental overhead” drops. But when the underlying system changes, even if the update is technically better, the mental model is broken, forcing users to relearn, rebuild habits, and sometimes re-invent old workarounds.

It’s not the existence of new technology that frustrates users; it’s the reset of the adaptation curve. That’s why the recent GPT-5 and Claude changes caused outsized backlash: the people most affected were the ones who had already adapted deeply, and who suddenly had to start that adaptation process again without consent or warning.

If a company knows that breaking a user’s learned workflow resets the adaptation curve, they can soften the blow by:

1. Announcing the change well in advance so users can prepare mentally and operationally.

2. Offering a transition period where both old and new systems are available in parallel.

3. Explaining what is actually changing in plain, practical terms instead of marketing language.

4. Documenting migration paths for common workflows, so power users don’t have to re-discover them.

This isn’t just politeness, it preserves user trust and reduces productivity loss. OpenAI’s quick reintroduction of GPT-4o after backlash was essentially an emergency version of #2. If they’d offered that from the start, the outrage would’ve been much smaller.

Or they could treat their own tech like a compatibility layer.

If GPT-5 behaves differently from GPT-4o in style, length, or reasoning patterns, you could run a thin AI layer that “translates” a user’s prompt into something GPT-5 will answer in a way that feels like GPT-4o would have, and optionally reformat GPT-5’s output to match the old style.

That would let:

Casual users experience GPT-5 improvements without noticing much difference.

Power users keep their finely tuned workflows intact.

The company phase out older models without ripping away user adaptations overnight.

It’s basically the same idea as running old code in a new OS via emulation (like VM Ware) just here, you’d be emulating behaviour instead of instruction sets.

Users, of course, could get smarter by using all LLMs simultaneously thereby getting used to moving around and adapting to different qwerks. Just like upgrading your phone each year prevents it becoming an ordeal when you’re forced to.

The interesting aspect of this blog is the observation of the adaptation of human intelligence to LLM AI. Ideally they move towards each other; that’s when we know we are heading towards AGI.