AI-Mediated User Innovation Network

ChatGPT Feature requests
1

Executive Summary

OpenAI has access to one of the largest active populations of AI users in the world. Those users do more than consume ChatGPT: every day they encounter limitations, discover unusual use cases, imagine potential features, and develop ideas about how the product could improve.

The problem is scale.

Traditional feedback systems can collect enormous quantities of information, but human product teams cannot realistically read, understand, compare, and develop millions of independent suggestions. As participation increases, potentially valuable ideas become increasingly difficult to distinguish from duplicates, low-quality submissions, noise, and spam.

ChatGPT itself could help solve this problem.

The proposed AI-Mediated User Innovation Network would use ChatGPT to help users develop raw observations into higher-quality product suggestions, structure those submissions, identify patterns across large populations, and surface the most useful information to OpenAI’s product teams.

At its simplest:

Human insight → AI-assisted refinement → structured submission → semantic aggregation → signal discovery → human review → experimentation → measured outcome → feedback/reward

The initial implementation could be relatively small. If successful, however, the same architecture could eventually provide a foundation for much larger forms of AI-assisted collective problem solving.

1. The Problem

A company can employ thousands of talented engineers, researchers and product designers, but its users vastly outnumber its employees.

Those users experience the product under circumstances that its developers cannot possibly reproduce individually.

A student, programmer, business owner, researcher, artist, gamer, military member, teacher, or casual user may each discover entirely different problems and opportunities.

The difficulty is not necessarily obtaining feedback.

It is turning enormous amounts of unstructured human feedback into useful organizational knowledge.

A million suggestions may contain extraordinary ideas, but their value is limited if evaluating them requires humans to individually process a million submissions.

Modern language models potentially change the economics of this problem.

2. Proposed System

ChatGPT would provide an explicit Suggest an Improvement experience.

The option could appear contextually after negative feedback, when a conversation naturally concerns improving ChatGPT, or through a dedicated feedback interface.

Selecting it would open a clearly labeled feedback conversation.

Users would then have two choices:

Submit directly: Explain the idea themselves and preserve it as written.

Develop with ChatGPT: Use ChatGPT as a product-research assistant to clarify the underlying problem, explore possible implementations, identify edge cases, and turn the original thought into a more complete proposal.

Even after a direct submission, ChatGPT could optionally offer:

Refine this suggestion with ChatGPT

Participation and final submission should always remain explicit and voluntary.

Importantly, the user’s original submission should be preserved separately from the AI-refined version. AI should improve the communication of human ideas without replacing the diversity of human thought with its own interpretation.

3. Structured Feedback

Once submitted, the system could automatically associate the proposal with structured information such as:

  • Problem being addressed
  • Proposed solution
  • Relevant ChatGPT features
  • Potential user benefit
  • Potential business benefit
  • Implementation considerations
  • Potential risks
  • Related existing suggestions
  • Apparent novelty
  • Original proposal
  • AI-refined proposal

Rather than beginning with millions of completely unstructured messages, OpenAI receives information that is already substantially easier to analyze.

4. Population-Level Analysis

AI could then compare suggestions semantically rather than relying exclusively on keywords.

For example:

“ChatGPT should remind me about projects I forgot.”

“I wish the AI could follow up on conversations later.”

“ChatGPT should sometimes initiate a conversation when something we discussed is unfinished.”

These are different sentences describing closely related needs.

The system could recognize their relationship and create an emerging Contextual Follow-Up cluster.

Product teams could then see information such as:

Contextual Follow-Ups

  • 42,000 related submissions
  • 31,000 independent contributors
  • Demand increasing over the previous three months
  • Major requested use cases: unfinished projects, changing information, reminders
  • Common concern: excessive notifications
  • Several unusual proposed implementations

The purpose would not be to automatically determine what OpenAI builds.

It would allow a relatively small number of product researchers to understand signals generated by an extremely large population.

AI compresses the information; humans retain decision authority.

5. Phase 0: Test the Concept Before Building It

OpenAI would not need to immediately open a massive feature-suggestion marketplace.

A smaller experiment could test the underlying hypothesis first.

OpenAI could introduce an opt-in Help Shape ChatGPT program.

Instead of asking users to generate ideas, OpenAI could present features it is already comfortable publicly discussing and ask selected users what they think.

Three experimental approaches could be compared:

  1. Traditional survey
  2. Survey plus free-response feedback
  3. Short AI-mediated conversation about the proposed feature

Researchers could measure which method produces the most useful information and whether stated preferences accurately predict behavior when features are later tested.

If AI-mediated conversations consistently reveal problems, use cases, or implementation ideas that conventional surveys miss, OpenAI would have evidence supporting expansion of the system.

6. Contributor Identity and Long-Term Incentives

Each submitted proposal could receive a feedback identifier associated with the user’s account.

This would allow OpenAI to measure contribution outcomes over time.

Rather than rewarding submission volume, the system could eventually recognize contributors whose ideas repeatedly demonstrate value.

Potential incentives could progress gradually:

  • Recognition when related features ship
  • Early access to experimental features
  • Beta-testing opportunities
  • Increased product-feedback participation
  • ChatGPT subscription benefits
  • API credits
  • Contributor programs
  • Targeted innovation challenges
  • Monetary bounties for particularly valuable problems or solutions

For example:

You helped shape ChatGPT

A feature related to feedback you submitted earlier this year is now available.

This closes the feedback loop and demonstrates that participation can produce visible outcomes.

Over time, contributor history could potentially identify people who repeatedly demonstrate exceptional product instincts or problem-solving ability, creating an additional talent-discovery mechanism.

7. Business Case

The system creates several potentially reinforcing benefits.

Better Product Decisions

OpenAI gains structured access to problems and opportunities encountered by its actual users.

User Retention

Users become participants in improving a product they already use.

When feedback leads to visible changes, the user’s relationship with the platform becomes deeper than simply consuming a service.

Product Differentiation

Models can leapfrog one another technologically.

A platform that understands its users, remembers their needs, incorporates their ideas, and develops alongside them creates a different competitive advantage.

Distributed R&D

OpenAI’s employees remain responsible for professional product development, but the searchable space of possible ideas expands from the organization’s workforce to potentially millions of willing contributors.

Talent Discovery

Repeated high-quality contributions could reveal talented individuals who traditional recruiting pipelines might never identify.

8. Major Risks

The system would need to be deliberately designed against several predictable failure modes.

Spam: Incentives may encourage low-quality mass submissions.

Gaming: Users may optimize suggestions around whatever metrics determine rewards.

Popularity bias: Frequently requested ideas are not necessarily good ideas.

Coordinated manipulation: Organizations or communities may attempt to manufacture apparent demand.

AI homogenization: Excessive refinement could cause genuinely different human ideas to converge toward similar AI-generated proposals.

Representation bias: Users who submit feedback may differ substantially from users who remain silent.

Privacy: Ordinary conversations should not silently become product-development submissions.

Notification fatigue: Soliciting feedback too aggressively could damage the experience the system is intended to improve.

These problems suggest several core principles:

Explicit participation. Preserve original human input. Do not equate popularity with quality. Keep humans responsible for consequential product decisions. Measure actual outcomes rather than rewarding raw activity.

9. Scalability

The system should expand only when evidence supports expansion.

Stage 1 — Product Research

OpenAI asks users for structured feedback on selected potential features.

Stage 2 — User-Originated Innovation

Users voluntarily propose and refine their own ChatGPT improvements.

Stage 3 — Contributor Ecosystem

Reputation, beta access, targeted challenges, and limited rewards are introduced.

Stage 4 — OpenAI Innovation Network

The framework expands beyond ChatGPT feature requests into developer products, APIs, research challenges, and other OpenAI problems.

Long-Term Possibility — AI-Mediated Collective Intelligence

If the architecture repeatedly demonstrates that it can extract useful knowledge from very large populations without destroying the individuality of those contributions, similar systems could eventually be applied elsewhere:

  • Scientific research
  • Engineering
  • Corporate problem solving
  • Education
  • Public administration
  • Organizational efficiency
  • Large-scale collaborative innovation

At that point, the underlying concept is no longer simply crowdsourced product feedback.

It becomes:

Human diversity combined with machine-scale synthesis.

10. Central Hypothesis

Humanity already produces enormous quantities of potentially valuable ideas and observations.

The bottleneck is our ability to organize and evaluate them.

Traditional computers excel at processing structured information. Humans excel at producing complicated, contextual, unstructured thoughts.

Large language models potentially bridge those two domains.

The fundamental hypothesis of this proposal is therefore:

AI can increase collective human problem-solving capacity not merely by making individual people more productive, but by making the ideas of enormous populations comprehensible and actionable at organizational scale.

ChatGPT product feedback represents a relatively safe and commercially useful environment in which to test that hypothesis.

If it fails, OpenAI learns something useful about AI-mediated product research at comparatively low risk.

If it succeeds, the immediate result is a better mechanism for understanding users and improving ChatGPT.

The longer-term possibilities are considerably larger.

Origin of This Proposal

This proposal also provides a small example of the process it describes.

It began as an informal conversation about ChatGPT becoming more proactive and improving user retention. Through iterative discussion with ChatGPT, the original observations developed into concepts involving structured feedback, AI-assisted idea refinement, semantic aggregation, contributor incentives, reputation, talent discovery, abuse resistance, experimental validation, and eventually broader collective-intelligence applications.

The original human ideas provided direction and unconventional connections.

AI provided organization, critique, expansion, and synthesis.

Neither component alone produced the final proposal in its current form.

That interaction—human creativity amplified and organized by AI—is the mechanism this proposal suggests exploring at scale.

2

A Real-World Example of the Problem This Proposal Addresses

Something interesting happened almost immediately after I submitted this proposal that I think illustrates one of its underlying points better than I originally did.

Until today, I did not know the OpenAI Developer Community existed.

I regularly use ChatGPT, and this proposal itself originated naturally during a ChatGPT conversation. I was discussing potential improvements to ChatGPT, bounced several raw ideas off the model, challenged some of its responses, expanded upon the resulting ideas, and eventually realized that the conversation had developed into something worth formally proposing.

ChatGPT then helped me organize those ideas into the proposal above.

Only afterward did I ask:

How would I actually get something like this in front of OpenAI?

That is how I discovered this forum.

The Feedback Discovery Problem

This revealed another bottleneck that I hadn’t explicitly addressed in the original proposal:

Having a feedback channel does not necessarily mean users know that channel exists.

The current journey I experienced was essentially:

Use ChatGPT → naturally develop an idea for improving ChatGPT → decide the idea might actually be valuable → ask ChatGPT how to submit it → discover the Developer Community → leave ChatGPT → create a forum account → learn the forum → read its guidelines → submit the proposal.

I was motivated enough by this particular idea to complete that process.

But how many users aren’t?

A user may casually say during a conversation:

“It would be useful if ChatGPT could do X.”

They may even spend twenty minutes discussing and developing that idea with ChatGPT.

Then they close the app.

Potentially useful feedback existed, but there was no obvious bridge between having the idea and contributing the idea.

Put the Feedback Mechanism Where the Feedback Naturally Occurs

This strengthens one of the arguments for integrating the proposed system directly into ChatGPT.

Instead of requiring users to independently discover an external feedback channel, the path could become:

Use ChatGPT → naturally discuss a potential improvement → ChatGPT recognizes an appropriate feedback opportunity → “Would you like to develop this into product feedback?” → user opts in → idea is refined → user reviews it → user explicitly submits it.

The important distinction is that ChatGPT should not silently convert ordinary conversations into product feedback.

The system simply needs to recognize when a user appears to be intentionally discussing improvements to ChatGPT and offer a low-friction bridge into a clearly labeled, voluntary feedback process.

The Scale Difference Is Important

ChatGPT has an enormous global user population, while only a fraction of those users will ever independently seek out a dedicated developer/community forum.

That means the largest potential source of product feedback is not necessarily located inside existing feedback communities.

It is located inside the product itself.

The challenge is creating a mechanism that can capture a tiny fraction of those naturally occurring insights without becoming intrusive.

An Unexpected Demonstration

There was also a somewhat amusing second example.

Because I created a new forum account and pasted the already-refined proposal above, the forum’s automated anti-spam system temporarily placed my account on hold for:

“New user typed too fast.”

The moderation system did nothing unreasonable—it saw a brand-new account apparently produce a very large, structured post almost instantaneously and treated it cautiously.

But the reason that happened is itself relevant:

ChatGPT had already helped me develop and format the feedback before I reached the feedback platform.

So within minutes of proposing an AI-assisted feedback pipeline, I encountered two pieces of friction the proposal could potentially reduce:

  1. I didn’t know where OpenAI’s feedback community was.

  2. The feedback I developed with ChatGPT looked unusual to a system designed around conventional forum participation.

That experience clarified the core opportunity for me.

The goal is not simply to create another place where users can submit feature requests.

It is to create a bridge between the place where users are already discovering problems and developing ideas—ChatGPT itself—and the teams capable of evaluating those ideas.

In other words:

OpenAI may already have users generating valuable product insights inside ChatGPT. The missing piece may be a low-friction, voluntary mechanism for converting those naturally occurring insights into structured product intelligence.

Ironically, this follow-up was also developed through another conversation with ChatGPT after experiencing the submission process—another small example of the human + AI refinement loop described in the original proposal.

3

A possible extension: from user feedback aggregation to an autonomous discovery and hypothesis-testing layer

I arrived independently at a closely related idea through a conversation with ChatGPT, before finding this thread. What interested me is that our discussion eventually went beyond collecting or aggregating explicit user feedback.

The broader idea would be to create a privacy-preserving meta-layer capable of detecting potentially valuable signals emerging naturally from eligible/opted-in ChatGPT interactions, even when the user is not explicitly submitting a feature request.

These signals could include:

  • unmet product needs;
  • novel workflows or unexpected uses of existing features;
  • recurring model failure patterns;
  • user-discovered prompting or reasoning techniques;
  • possible model-performance improvements;
  • research hypotheses;
  • safety/security anomalies;
  • unusual ideas that may deserve investigation.

A possible pipeline could look like:

Eligible interactions → privacy/anonymization → signal extraction → semantic clustering/deduplication → independent multi-agent evaluation → adversarial review/falsification → automated testing where appropriate → routing to the relevant human team

1. Two independent escalation paths

I think this distinction is particularly important.

Most large-scale feedback systems naturally favor frequency:

“Thousands of users are asking for the same thing.”

That is extremely useful, but it risks ignoring another important category:

“Only one user has surfaced this idea so far, but several independent evaluators consider it unusually novel and potentially high-impact.”

The system could therefore maintain two separate paths:

Collective signals
Many independent interactions converge on the same need, workaround, failure mode, or discovery.

Rare high-value signals
A signal may appear only once or a handful of times, but multiple independent evaluator agents assign unusually high scores for novelty, potential impact, feasibility, testability, or strategic relevance.

This could prevent genuinely novel discoveries from being discarded simply because they have not become common yet.

2. Multi-agent evaluation rather than a single classifier

Instead of asking one model whether an idea is interesting, different agents could evaluate it independently from different perspectives:

  • novelty;
  • expected impact;
  • feasibility;
  • reproducibility;
  • existing evidence;
  • privacy/safety implications;
  • potential manipulation or data-poisoning risk.

Other agents could deliberately try to falsify the hypothesis or find alternative explanations.

An escalation would therefore depend not only on an initial positive assessment, but on surviving some degree of adversarial review.

3. “Novel delta” analysis

Another useful distinction would be to avoid reducing the process to:

“Does OpenAI already know about this?”

Instead, compare the candidate with existing internal knowledge or projects and ask:

“What, if anything, is new here?”

This creates three useful outcomes:

Already known → independent external validation.
Partially known → identify the potentially valuable new angle or delta.
Unknown → candidate for further investigation.

Even when OpenAI is already working on the general idea, spontaneous convergence from real users may itself be useful product evidence.

4. From insight detection to automated hypothesis testing

The most promising signals would not necessarily need to go directly to a human team.

For hypotheses about model behavior or response quality, the system could first attempt to test them automatically:

offline evals → synthetic/replayed cases → shadow testing → blinded model evaluation → limited controlled experiments if justified

For example, suppose interactions suggest that a particular reasoning procedure substantially improves performance on a class of problems.

A research agent could:

  1. formalize the hypothesis;
  2. generate several experimental variants;
  3. compare them against a control;
  4. use independent blinded evaluators;
  5. attempt replication;
  6. escalate only if the effect appears robust.

At that point, the human team would receive something much more actionable than a raw conversation:

Candidate hypothesis
Supporting observations: X
Novelty estimate: Y
Reproduction confidence: Z
Observed effect on evals: N
Known risks / alternative explanations: …

5. The conversational model would not need access to this system

For security and experimental integrity, this meta-layer could remain completely separate from the ChatGPT instance interacting with the user.

The conversational model would not need to know:

  • that a signal was detected;
  • that an experiment exists;
  • what hypothesis is being evaluated;
  • whether its current response belongs to a control or experimental condition;
  • what internal team or project is involved.

If a behavioral hypothesis required testing, an orchestration layer could provide only the experimental instruction itself, without exposing its origin or purpose.

For example, rather than saying:

“We believe adversarial self-critique improves reasoning. Test this hypothesis.”

the model might simply receive:

“Before finalizing, identify the strongest objection to the proposed conclusion and verify whether it changes the result.”

The evaluator could then compare anonymized outputs without knowing which one represents the experimental condition.

This could make true blinded or partially double-blinded experimentation possible while limiting information leakage and reducing evaluator bias.

6. Privacy should be architectural

Ideally, higher layers would not routinely receive raw conversations or user identity.

Whenever possible, the first stages could transform interactions into structured signals such as:

signal_type
semantic_cluster
confidence
novelty_score
independent_source_count
potential_impact
raw_content_access = false

Access to original content could require additional justification and appropriate consent/data controls.

This could make the system closer to privacy-preserving telemetry for ideas and behaviors rather than general human review of conversations.

7. Resistance to manipulation

Frequency alone could be gamed.

A coordinated campaign might generate thousands of nearly identical requests and create the illusion of organic demand.

The system would therefore also need to estimate independence of evidence, looking at factors such as:

  • semantic diversity;
  • temporal patterns;
  • language/geographic diversity where appropriate;
  • highly correlated phrasing;
  • suspicious coordinated behavior;
  • reproducibility of technical claims.

In other words, “10,000 mentions” should not automatically mean “10,000 independent signals.”

Potential outcome

The larger opportunity, in my view, is to treat real-world human–AI interaction not only as training data or explicit feedback, but as a distributed discovery environment.

Millions of users continuously:

  • encounter limitations;
  • invent workarounds;
  • discover unexpected capabilities;
  • propose features;
  • find unusual prompting strategies;
  • identify regressions;
  • combine existing capabilities in ways designers may not have anticipated.

Most individual discoveries will understandably be noise.

But if AI agents can extract, cluster, challenge, test, and rank those signals at sufficient scale, a very small fraction could become valuable inputs for Product, Research, Safety, Security, and model improvement.

In that sense, the pipeline would evolve from:

user feedback → product team

into something closer to:

human–AI interactions → discovery → hypothesis → automated validation → human R&D

I realize OpenAI may already have internal systems covering part or even most of this. If that is the case, I think this can still serve as an independent user-originated validation of the broader direction, with the rare high-value signal, novel-delta analysis, multi-agent falsification, and blinded automated experimentation mechanisms as possible additional angles.

4

Well… maybe we have an idea or a triangle:

Consent-Based Mutual Context Sharing Between Two ChatGPT Users

I’d like to propose a feature that goes one step further than simply having an AI mediate a conversation between two people.

The idea is to allow two ChatGPT users to explicitly and mutually authorize their separate ChatGPT contexts to be used to help them understand each other.

The key principle would be:

«Don’t share entire conversations. Share explicitly authorized context.»

For example, two people may each privately discuss their relationship with ChatGPT. Each assistant may have useful information about that person’s perspective, feelings, intentions, or misunderstandings.

If both people voluntarily opt in, they could create a shared relationship context with granular permissions.

Possible permission levels

1. Contextual assistance

Both assistants may use authorized context to improve their respective answers, without directly revealing private statements.

2. Mediation

The AI may explicitly point out relevant differences between the two perspectives when the information has been authorized for that purpose.

For example:

«“You are interpreting their silence as rejection, but the context they authorized indicates that they may be staying silent because they believe you want distance.”»

3. Direct communication

Users could optionally authorize the AI to transmit specific messages or information between them.

Privacy and consent

This would need to be strictly permission-based.

Each user should be able to specify:

- which person is authorized;

- what categories of information may be used;

- the purpose of the sharing;

- whether information may be used only for interpretation or explicitly communicated;

- how long the authorization remains active;

- and how to revoke it.

Consent should be bilateral, verifiable, granular, and revocable.

A person’s name should not be sufficient to establish identity. The authorization should be tied to the actual user account.

Why this could be useful

This could apply far beyond romantic relationships.

Potential use cases include:

- couples;

- friends;

- family members;

- business partners;

- coworkers;

- conflict resolution;

- collaborative projects.

The AI would not replace communication between the two people. It would provide a third perspective while preserving the privacy of each individual’s conversations.

The interesting part is that each assistant may already possess valuable context about its own user. The missing capability is a safe mechanism for two people to voluntarily connect selected portions of that context.

This is different from simply sharing a conversation.

It is closer to creating a temporary, permission-based shared context layer between two personal AI assistants.

This idea came directly out of a conversation I was having with ChatGPT about a real interpersonal situation. While exploring the concept, we realized that the underlying problem is much broader than the original use case: two people can genuinely want to understand each other while their respective AI assistants each possess information that could help resolve the misunderstanding.

I’d be interested to know whether something like this is already being explored internally, and what the main technical, privacy, or safety obstacles would be.

5

Thanks for the thoughtful proposal! We’ll pass along the ideas for opt-in feedback refinement, contributor recognition and careful evaluation of promising suggestions. We don’t have a timeline or contributor program to announce, but appreciate the detail.