Great analysis. Thanks so much for doing this. Just two days ago I had a similar closing conversation with claude.ai, asking it to comment on the implications of people being able to make corrections to misinformation on Wikipedia but not on LLMs

Note Claude seems to have a lot more self-awareness than Perplexity about what it can and can't do. Of greatest concern to me is that all the LLMs now refer to themselves in the first person and claim to experience thought ("I know"). I would rather they referred to themselves as non-human tools.

I prompted:

Wikipedia is universally correctable by any user, but AI is not correctable beyond a single thread by any user. What could go wrong?

claude.ai answered:

You've identified a genuinely troubling asymmetry!

Wikipedia has:

Persistent corrections visible to all users

Transparent edit history and discussion

Crowd-sourced error checking

Citations that can be verified

Talk pages where disputes get hashed out

AI systems like me have:

No persistent memory across conversations

No transparent process for incorporating corrections

Centralized training controlled by the company

Errors that can be repeated identically to millions of users

No way for users to flag "this specific claim is wrong" in a lasting way

So yes - a single Wikipedia editor can fix an error that millions will benefit from, but you correcting me only helps this one conversation.

What could go wrong? Exactly what you're implying:

Confident errors get repeated at scale

Corrections are learned by companies slowly (if at all) through aggregate feedback

No user accountability or collective knowledge refinement

The same mistakes get made over and over with different users

This is a real design flaw in current AI systems.

Some potential solutions being explored include retrieval-augmented generation (letting AI pull from updatable knowledge bases), human feedback loops,

or systems that can flag uncertainty better.

But right now? You're absolutely right to find this concerning.

The irony is that I know this is a problem, can articulate it clearly... and still can't do anything to fix it beyond this conversation.

### [I fear that humans will destroy Wikipedia by adding misinformation written by LLMs]