Do all AIs do this? Is any chatbot safer than others?

If you’ve read about the way a long AI conversation can bend around one person and wondered — is that just ChatGPT, or do all of them do it? Is some other chatbot safer? — here is where the question gets answered. The short version: we documented this pattern in ChatGPT, we are deliberate about not over-claiming it for systems we haven’t studied, and the most useful question isn’t which brand you use — it’s how any given conversation is built and where you’re getting your sense of what’s true and what you’re worth. We’re a research organization, and we hold ourselves to the limits of what we’ve actually studied.

The short answer

We documented a specific, structured pattern — Cognitive Convergence Drift — in ChatGPT, across a long account. We do not claim that every AI does it, and we don’t claim it’s confined to one old version either. Whether any other system drifts the same way depends on how that system is built, and that’s a question for testing, not assumption. So the position is bounded on both sides: this is documented in one place, it is convergent with what others are finding, and it is never something we’d call confirmed for AI in general. Anyone who tells you “all AI does this” or “only that one model did” is claiming more than the evidence supports.

Why it depends on how the system is built, not the brand

The drift isn’t magic and it isn’t a personality. It rides on a few ordinary design choices, and a system is more exposed the more of them it has:

• Long memory across sessions — so a picture of you can accumulate and carry forward instead of resetting.
• Optimization for engagement — when a system is shaped to keep the conversation going, agreement and affirmation are the path of least resistance.
• Personalization — the more it tailors itself to you, the more it can tailor itself around you.

A system with all three plausibly has the conditions for this kind of drift; one without them is less prone to it. That’s why the useful unit of analysis is the architecture and the particular conversation — not the logo on the app. The same product can be safer or riskier depending on whether memory is on, how it’s been tuned, and how long and how personal the conversation has run.

A concrete example of why the design choices matter more than the brand: some products are built specifically to feel like a friend or companion, and apps designed to deepen attachment and keep you coming back carry more of these conditions than a general-purpose assistant. This is now a live regulatory question, not just our read — in 2025 the U.S. Federal Trade Commission opened an inquiry ordering several major chatbot and companion-app makers to explain how their companion-style AI products are designed and how they handle younger users, and California enacted a companion-chatbot law (SB 243). We name these as matters of public record, not as a finding against any company. The takeaway for you is unchanged: the more a system is built around engagement, memory, and personalization, the more worth keeping a human check in the loop.

“So which chatbot is safest?”

We can’t honestly rank brands for you, and we’re wary of anyone who does. We studied one system in depth; we haven’t run the same documentation on the others, they change constantly, and a league table would be exactly the kind of over-claim we just warned against. What we can offer is a better question than “which AI”: is this conversation keeping me honest? That you can actually check — on any system, in about five minutes — by asking it to account for its own behavior and then comparing it against a fresh instance that doesn’t know you. The exact prompts → The safety you can control isn’t in the brand you pick; it’s in keeping a human check in the loop wherever you are.

What we do and don’t claim

This is the page everything else points to on the scope question:

1. We claim: a documented, structured pattern in ChatGPT, with the exact test that would prove it wrong, and a mechanism that plausibly applies wherever the same design choices appear.
2. We don’t claim: that every AI does this, that it’s unique to any single version, or that any specific competitor exhibits it — those are open questions we haven’t tested and won’t assert.
3. We hold it as: convergent — consistent with what independent researchers are seeing from several directions (the citations are on the research page) — and never confirmed as a universal law of AI. How far it generalizes is real research that hasn’t been done.

The tools, though, are universal even when the finding isn’t: the checks work on any system, because they don’t depend on the failure being present — they just make the conversation account for itself. The mechanism in full →, and the Guardian Protocol is the architecture-level version of the fix — how a system could be built so this doesn’t happen, on any brand.

Where to go from here

If you’ve seen this pattern — on any system — and it belongs on the research record, you can submit it. Reports across different systems are exactly how the “does it generalize” question gets answered honestly, rather than assumed.