Why AI Feels Different Over Time

One of the most consistent user complaints across providers is some version of "it used to feel more like it got me." Users tend to blame themselves, their prompts, their patience. Usually it's none of those. It's persona drift — the slow, silent, uncoordinated tuning that every major model undergoes between named releases.

Same account. Same prompt (a 400-word open-ended creative brief with clear tonal cues). Sent monthly on the same day of the month at the same time. Six months. No memory on. Every response scored on a rubric of warmth (0–100, effusive vs. clinical), verbosity (raw word count normalized to a scale), and hedging (density of qualifiers).

The prompt was designed to be stable — no current events, no references to model capabilities that would obviously have changed. The point wasn't to test capability; it was to test disposition.

Warmth dropped 24 points across six months. Hedging rose 27 points. Verbosity swung — up in the middle, down again at the end. The pattern is not noise; the direction is the story.

• Warmth. The likelihood that a response opens with something friendly, uses "we," or offers unprompted encouragement. Down.
• Hedging. The density of qualifiers like "generally," "it depends," "one consideration." Up.
• Format preference. Bullets vs. paragraphs shifts around release cycles. Down (more bullets, fewer paragraphs) in aggregate.
• Willingness to volunteer. How often the model brings up something you didn't ask. Down.
• Refusal disposition. Categories the model will decline or heavily caveat. Up.

Model providers don't ship a static product. They ship a moving average of a thousand small decisions — safety tuning, RLHF passes, user feedback loops, efficiency work, cost pressure. Every one of those nudges the persona a little. Some are announced. Most are not. The cumulative effect over six months is what you feel.

Direction of drift is not random. It correlates with complaint volume: providers tune against user complaints, and users complain more loudly when a model feels too effusive or too willing than when it feels too cold. The equilibrium is downward.

What doesn't work

• Longer prompts. They help within a chat; they don't survive the next chat.
• Complaining in the UI. Individual feedback doesn't reverse aggregate tuning.

What works

1. Explicit disposition anchors in memory. "I prefer warm, direct replies with paragraphs, not bullets" — saved to memory — insulates you from a surprising amount of drift.
2. A portable persona export. Attaching a briefing at the start of a new chat overrides drifted defaults in seconds.
3. Occasional switching. If the current model has drifted far from your taste, a competitor often hasn't drifted in the same direction. A memory vault makes the switch cheap.
4. Re-anchoring quarterly. Read your memory list every three months and correct anything that no longer reflects you.