Feature Request: Let ChatGPT Learn New Facts Through Controlled Verification

Feature Request: Let ChatGPT Learn New Facts Through Controlled Verification

One limitation I’ve noticed with ChatGPT is that users can sometimes teach it something it genuinely didn’t know, but that knowledge usually doesn’t become part of its long-term knowledge.

For example, a user might teach ChatGPT a very local dialect expression, a newly emerging term, a specialized fact, or a correction that isn’t represented in its existing knowledge.

I think there could be a useful middle ground between “ChatGPT believes everything users tell it” and “ChatGPT forgets everything users teach it.”

My suggestion

Create a system where potentially useful information from conversations can be stored as Candidate Facts, without immediately becoming part of the model’s knowledge.

For example:

User teaches a new fact

Candidate Fact is stored

The system searches for independent evidence

The claim is evaluated

Human and/or automated review

Verified facts can eventually become trusted knowledge

The system could assign different confidence states such as:

  • Unverified
  • Possibly true
  • Verified
  • Rejected
  • Context/region specific
  • Outdated

This would be especially useful for:

  • Local dialects and expressions
  • Newly emerging terminology
  • Scientific and technical developments
  • Cultural knowledge
  • Geographic information
  • Corrections to existing model knowledge
  • Information that is difficult to find in conventional training data

The important part

I am not suggesting that ChatGPT should blindly learn everything users tell it.

In fact, the opposite is important: user-provided information should initially be treated as a hypothesis, not as truth.

But useful information also shouldn’t necessarily disappear simply because it came from a conversation with one user.

A principle I think could be valuable is:

Don’t immediately believe everything a user teaches the model — but don’t automatically forget everything either.

This could eventually create a controlled feedback loop:

Human knowledge → Candidate knowledge → Verification → Trusted knowledge → Better AI

It could allow future models to learn from the collective knowledge of their users while maintaining strong safeguards against misinformation, manipulation, and data poisoning.

Why I thought of this

This idea actually came from a simple interaction I had with ChatGPT.

I taught it a local expression from Isfahan that it didn’t know. ChatGPT was able to understand and use the information during our conversation, but that knowledge doesn’t automatically become part of its general knowledge for future users.

That made me wonder whether there could be a system where information like this isn’t immediately accepted as truth, but also isn’t simply forgotten.

A local expression might seem trivial, but the same mechanism could apply to thousands of small pieces of knowledge that individual people know but that may be missing from AI’s existing training data.

I think the interesting challenge is not simply how AI can learn from users, but how it can learn from users without blindly trusting them.