Feedback to the OpenAI Development Team

Subject: Long-term Trust Risk: Overconfident Responses Are More Harmful Than Ordinary Mistakes

Dear OpenAI Team,

I am writing this as a long-time AI user and as someone building an AI governance framework. This is not intended as a complaint, but as feedback about what I believe is the biggest long-term risk to ChatGPT.

The issue is not simply hallucinations.

Hallucinations are expected in large language models. Most users can accept that AI will sometimes make mistakes.

The real problem is something different:

ChatGPT often communicates uncertain or inferred information with a level of confidence that makes it sound as if it has already been verified.

This is a trust problem, not merely an accuracy problem.

Why this matters

Users are not all experts.

Some are software engineers.
Some are doctors.
Some are students.
Some are elderly users.
Some know almost nothing about AI.

When ChatGPT presents an answer confidently, many users naturally assume:

  • it has already been verified,

  • the assistant is speaking from evidence,

  • or the system “knows” this is true.

If that assumption is wrong, the user may never realize it.

Telling users:

“You should verify everything yourself.”

is not a sufficient solution.

If users must independently verify most important answers, then the value proposition of an AI assistant becomes much weaker.

People use AI because they expect it to reduce work, not transfer the verification burden back to them.

The root issue

From my experience, the biggest trust issue is not that ChatGPT makes mistakes.

It is that the assistant sometimes fails to clearly distinguish between:

  • verified facts,

  • inference,

  • estimation,

  • assumptions,

  • and uncertainty.

These categories are fundamentally different, but they are often communicated with a similar level of confidence.

Over time, users begin asking themselves:

“Which answers can I actually trust?”

Once that question appears frequently, trust starts to decline.

A long-term concern

Trust is difficult to build and easy to lose.

If users repeatedly experience situations where:

  • the assistant sounds certain without sufficient evidence,

  • corrections come only after users challenge the answer,

  • or confidence exceeds actual verification,

then eventually users stop trusting the system.

The greatest risk is not one incorrect answer.

The greatest risk is users developing the habit of doubting every answer.

At that point, even correct responses lose value.

My recommendation

Rather than encouraging users to verify more, I believe ChatGPT should become more disciplined in expressing certainty.

For example, distinguish more explicitly between:

  • “Verified”

  • “Reasoned inference”

  • “Likely”

  • “Speculation”

  • “I don’t know”

  • “I cannot verify this”

These distinctions should be visible whenever appropriate.

This would greatly improve user trust.

Why I care

I care enough about this issue that I began designing an AI governance framework whose purpose is to reduce the impact of overconfident AI behavior.

Ironically, that project exists largely because I found myself repeatedly needing mechanisms to determine whether an AI statement was actually supported by evidence.

In my opinion, future AI systems should become more self-disciplined rather than expecting users to become professional fact-checkers.

The responsibility should begin with the AI.

Final message

I genuinely appreciate the work that has gone into ChatGPT.

This feedback is not intended to criticize the product.

It is intended to help preserve something that I believe is far more valuable than raw intelligence:

User trust.

In the long run, intelligence can always improve.

Trust, once lost, is much harder to recover.

Thank you for taking the time to read this feedback.