New Idea, for more fun, and knowledge

ChatGPT Feature requests
,
1

Feature Request: A Community-Verified Knowledge Layer for ChatGPT

The idea

I would like to suggest a feature that could make ChatGPT much more powerful over time: a community-verified knowledge layer, where users could contribute information that ChatGPT does not currently know, and where that information could eventually become useful to other users after appropriate verification.

The idea is somewhat similar to Wikipedia, but adapted to an AI system.

A real-world example

This idea came to me after asking ChatGPT about the origin of an image.

ChatGPT searched for the image and could not determine where it came from. However, I personally knew the source because the image came from someone I know.

I could tell ChatGPT:

“This image actually comes from [source].”

But under the current system, that information is essentially useful only within my own conversation (or potentially my personal memory). It cannot simply become knowledge that helps the next person who asks ChatGPT the exact same question.

That feels like a missed opportunity.

What if users could contribute knowledge?

Imagine a system where ChatGPT could respond:

“I don’t currently have a reliable source for this information. If you know the answer, you can contribute it.”

The user could then provide the information.

ChatGPT would not immediately treat it as fact. Instead, the contribution could enter a verification process.

For example:

User contribution:
“This image originates from [X].”

Status:
Unverified user contribution

Then ChatGPT could attempt to verify it using available sources, compare it with other user contributions, and potentially look for independent confirmation.

The information could eventually progress through states such as:

  • Unverified
  • User-reported
  • Independently corroborated
  • Verified
  • Disputed

This would allow ChatGPT to benefit from human knowledge without blindly trusting everything users tell it.

Why this could be valuable

There is an enormous amount of information that is difficult or impossible to find through conventional web search.

Examples include:

  • The original source of obscure images
  • The identity of an artist behind a little-known work
  • References in old or obscure media
  • Details about video games that are poorly documented online
  • The origin of memes or images
  • Local knowledge
  • Small creators or independent works
  • Very specific historical details
  • Information that exists only in communities or personal experience

Millions of ChatGPT users collectively know an enormous amount of information.

Today, that knowledge is largely fragmented across individual conversations.

A controlled system could potentially turn some of that fragmented knowledge into a collective, continuously improving knowledge base.

Important: verification and trust would be essential

I would not suggest simply allowing ChatGPT to “learn whatever users tell it.”

That would obviously create enormous problems with misinformation.

Instead, contributions should have a confidence level and a verification history.

For example:

Claim: This image originates from X.
Submitted by: Community member
Independent sources: 2
Other users confirming: 7
Status: Corroborated
Confidence: High

ChatGPT could then make the distinction clear when answering:

“According to verified sources…”

versus:

“This information has been reported by users but has not yet been independently verified.”

That distinction would be extremely important.

A possible long-term system

The system could work something like this:

User asks a question → ChatGPT cannot find a reliable answer → User contributes information → ChatGPT verifies it → Community corroborates it → Knowledge receives a confidence level → Future users can benefit from it.

It could even detect when a user is providing information that may be useful beyond the current conversation and ask:

“Would you like to submit this information as a community contribution?”

Nothing should ever be shared automatically. The user should explicitly approve the contribution and be able to see exactly what would be submitted.

Why I think this is different from simply improving the model

This would not necessarily require ChatGPT’s underlying model to be retrained every time someone contributes a fact.

It could instead function as a separate, continuously updated knowledge layer that ChatGPT can consult.

That would make the system more dynamic while also allowing incorrect information to be corrected, disputed, or removed without having to retrain the entire model.

In other words, ChatGPT could become not only an AI that knows things, but an AI connected to a living, community-verified knowledge base.

The bigger idea

Wikipedia demonstrated that millions of people can collectively build an incredibly useful knowledge resource.

ChatGPT has something Wikipedia does not have: the ability to interact naturally with each individual contributor and understand the context in which information is being provided.

Combining those two concepts could be extremely powerful:

Wikipedia-like collective knowledge + AI reasoning + automated verification + human oversight.

I believe this could be a very interesting direction for the future of ChatGPT.

And, perhaps most importantly, it would allow useful pieces of knowledge discovered by one user to potentially benefit everyone else — rather than remaining trapped inside a single conversation.

2

I think there is a genuinely interesting idea here, although I suspect the useful implementation would look less like a shared memory for ChatGPT and more like a community-maintained claim/evidence graph.

My main concern is that the difficult part is not storing what users tell ChatGPT. The difficult part is deciding when a claim is allowed to become shared knowledge.

For example, image provenance is already a much harder problem than ordinary factual retrieval. Finding an image on a website does not prove that the website originated it. The earliest indexed occurrence may not be the creator either. Proper verification could require reverse-image search, archives, metadata, publication chronology, creator testimony, etc. Some of those capabilities may also intentionally be restricted for privacy and security reasons.

There is also an important distinction between different kinds of claims.

“The Eiffel Tower opened in 1889” is an externally verifiable factual claim.

“This obscure image was originally created by X” is a provenance claim.

“Game X has better movement than Game Y” is opinion.

“I personally know the person who created this image” may be completely true, while still being impossible for the system to independently verify.

So I do not think a generic progression such as:

unverified → corroborated → verified

would be sufficient. The verification policy would have to depend on the type of claim.

I would also be careful with community corroboration. Seven users confirming something does not necessarily mean seven independent sources. They could all be repeating the same Reddit post, belong to the same community, or even be controlled by the same person. Likewise, two websites are not necessarily two independent sources if one copied the other.

That creates a serious knowledge-poisoning / Sybil problem.

This is why I think a knowledge graph—more specifically, a claim/evidence graph—might be more useful than directly promoting submissions into “ChatGPT knowledge.”

Instead of storing:

Image → originates from → X

store something closer to:

Claim: Image originates from X
→ asserted by contributor A
→ supported by evidence B
→ supported by evidence C
→ contradicted by evidence D
→ confidence / verification state
→ valid from / until
→ supersedes earlier claim

Then the model can reason over the evidence rather than treating the community-created triple itself as truth.

There are further issues around privacy, defamation, anonymous creators, local information, temporal facts, coordinated manipulation, source independence and correcting previously “verified” claims.

So I like the direction, but I think the core research problem is actually epistemic governance, not shared memory: Who can submit a claim? What counts as evidence? What counts as independent corroboration? Who has authority to promote or reject it? How are disputes represented? How does a claim get revoked later?

If those questions can be solved, then a community knowledge layer could indeed be extremely valuable—especially for long-tail information that conventional search engines cannot recover. But I would strongly avoid allowing the model itself to decide, “I searched, therefore this user contribution is now verified.” The model should probably be a consumer/reasoner over the evidence layer, while admission into trusted shared knowledge is controlled by a much stricter external policy.