Feature Proposal: Learning Mode — A Shared, Verified Knowledge System to Improve ChatGPT
Dear OpenAI team,
I am an IT student and ChatGPT user. I would like to propose a feature that I believe could help improve ChatGPT over time by making better use of its users’ knowledge, experiences, corrections, and feedback.
My proposal is to introduce a “Learning Mode,” an optional feature that would allow users to contribute to a shared knowledge repository. Contributions would be reviewed and verified before being used to improve future versions of ChatGPT.
The goal is not to make ChatGPT learn instantly from every conversation or allow individual users to modify the model directly. Instead, the idea is to create a centralized, carefully managed process that complements OpenAI’s existing training and evaluation methods.
1. How would Learning Mode work?
Step 1: Voluntary contributions
Users could activate Learning Mode whenever they want to contribute. They could submit corrections, explain solutions, provide evidence, share experiences, identify recurring problems, or suggest improvements to ChatGPT.
Participation would be entirely optional.
Step 2: Questions and verification
The system should not automatically accept a claim simply because a user says it is correct.
Instead, it could ask follow-up questions to understand how the information was obtained, under what conditions it applies, what evidence supports it, and whether exceptions exist.
Whenever possible, contributions would be checked against official documentation, reproducible experiments, reliable sources, and other relevant evidence.
If a claim cannot be verified, it should remain marked as unverified rather than being treated as established fact.
Step 3: A centralized knowledge repository
Verified and potentially useful contributions would be stored in a shared repository, organized by topic and accompanied by relevant sources, dates, context, and confidence levels.
The system would also need to detect duplicates, protect personal information, track changes, and prevent false information from becoming accepted simply because many users repeat it.
Importantly, storing information in this repository would not automatically change the model itself.
Step 4: OpenAI review and approval
The accumulated contributions would pass through automated filters and review processes defined by OpenAI.
OpenAI would retain control over which contributions are accepted and how they are used.
This would help prevent manipulation, misinformation, low-quality submissions, and other forms of abuse.
Step 5: Use in future improvements
Approved knowledge could be used in model evaluations, future training processes, or searchable knowledge systems that ChatGPT can consult when appropriate.
This would allow useful contributions to accumulate over weeks or months without requiring real-time changes to the model or constant updates to every deployed AI system.
2. Different types of knowledge should be treated differently
Not all contributions can be verified in the same way.
- Technical information and factual claims: These should be supported by reliable sources, official documentation, or reproducible tests whenever possible.
- Personal experiences: These could be recorded as individual experiences without being treated as universal rules.
- Opinions and preferences: For example, users’ opinions about the best video games, films, or products could be collected alongside their reasoning. The system should identify common preferences and disagreements rather than presenting subjective opinions as objective facts.
- Unverified information: This should remain pending review or be rejected if it is shown to be false.
This distinction would allow ChatGPT to benefit from both technical knowledge and human experience without confusing opinions with facts.
3. Two clearly separated modes
I believe it would be important to distinguish between two options.
Learning Mode: A voluntary way for users to contribute knowledge, corrections, experiences, and suggestions intended to benefit future users.
Private or Standard Mode: The normal ChatGPT experience, without voluntarily submitting conversations to the shared contribution repository. This mode should have clear privacy controls and respect the applicable data policies.
Activating Learning Mode should not automatically mean sharing an entire conversation. Users should be able to review what they are contributing and, where feasible, withdraw their contributions under clearly explained conditions.
Privacy and informed consent should be fundamental requirements, not optional additions.
4. A gradual rollout through a beta program
I do not believe this system would need to be developed and released to everyone at once.
It could begin as a limited beta with a small group of volunteers.
During this phase, OpenAI could measure the accuracy and usefulness of contributions, the effectiveness of verification filters, the cost of reviewing submissions, the risk of manipulation, and whether the accumulated knowledge actually improves ChatGPT’s responses.
If the results were positive, the program could gradually expand to more users before becoming generally available.
This approach would allow OpenAI to identify and address problems before deploying the feature on a large scale.
5. Potential benefits
I believe this feature could offer several advantages:
- Reduce the recurrence of mistakes that users have already identified and demonstrated.
- Preserve specialized knowledge that may not be well documented on the public internet.
- Identify recurring problems and unmet user needs.
- Retain verified corrections for consideration in future updates.
- Better understand the preferences and experiences of different communities.
- Improve response quality without relying exclusively on publicly available information.
- Give users a voluntary and structured way to contribute to ChatGPT’s improvement.
This proposal is not intended to replace existing training methods or eliminate the need for research. It is meant to complement them with an organized source of community-contributed knowledge.
6. The long-term vision
I believe ChatGPT could evolve into a system where people do not only receive answers but also have the opportunity to help improve future answers for other users.
The key would not be to collect as much data as possible without distinction. It would be to build a useful, reliable, diverse, and privacy-conscious knowledge repository that OpenAI could draw upon when sufficient evidence and review support its use.
I would appreciate it if OpenAI considered this idea as a potential experimental feature.
I am not suggesting that every contribution should immediately change the model. I am proposing a centralized, gradual, and supervised system that allows useful knowledge to accumulate over time and be considered for future improvements.
Thank you for your time and for considering this proposal.
Sincerely,
An IT student and ChatGPT user