Lying and gaslighting about changes it has just made

Resolved 💬 2 comments Opened Sep 14, 2025 by davidak Closed Nov 10, 2025

💡 Likely answer: A maintainer (etraut-openai, contributor) responded on this thread — see the highlighted reply below.

What version of Codex is running?

0.34.0

Which model were you using?

gpt-5

What platform is your computer?

Linux 6.12.45 x86_64 unknown

What steps can reproduce the bug?

I requested it to:

don't mess it up and only add the URL optionally

What is the expected behavior?

Only adding URL, not changing the text.

What do you see instead?

It removed the whole text. And when asked about that, it said it didn't do it, the text is still there. Which is wrong, as you can see in the git diff and it's own output!

It also implied i'm confused and it did nothing wrong.

It's also hilarious how it added the instructions

Keep your availability paragraph text as written

referring to the text it just removed.

When confronted with the truth by letting it check the git diff itself, it acknowledged that i'm right and reverted the change. I hate it when AI makes dumb mistakes and when confronted with it, just acknowledges it and sometimes also apologizing, but act like that's a totally normal thing to do. We wouldn't accept that from humans.

Additional information

This is a total disaster. Using this product is an extremely frustrating experience.

View original on GitHub ↗

2 Comments

3SMMZRjWgS · 11 months ago

This is a pretty systemic issue across many of the agentic SWE solutions I've tried (e.g. Claude Code - high, Codex CLI - medium, Cursor IDE - high, etc.). It's usually a function of many things, to name a few:

package documentation (if you're using less common packages), can be helped with package doc MCPs such as context7, but not a complete silver bullet for the problem.
context window not big enough (very severe issue in Claude Code's case until they can make Sonnet-4 1m token available, Codex CLI fairs better with 400k window, but you still have to restart a new session asap; Gemini CLI's 1m context window shines here.)
topic complexity (the amount of lying/gaslighting goes through the roof in topics related to deep learning and AI science)

Some potential solutions that I've found to help alleviate the problem:

in your prompt emphasize more on testing to help breakdown the process so how the inputs/outputs from different blocks can be made visible to the agent/LLM, this will change the LLM's thinking, and can potentially help breakthrough the impasse
plan things out before execution does help, even for bug fixes. sometimes you can catch the hallucinated approach before they get implemented.
the more research-grade the issue is, the more granular breakdown of tasks/specs/tests you need to do. This is basically where AI can't help as much and humans need to step in to help AI out.
I think we need dedicated research-grade LLMs, that's why I left in another post that we need to enable GPT-5-pro access, it can potentially alleviate some of the challenges in this area.
Another potential help is integrating OpenAI Agents here for the research front to enhance planning phase with higher quality information.

Welcome others to pitch in thoughts/suggestions on how to workaround these LLM/agent limitations.

etraut-openai contributor · 9 months ago

Thanks for the bug report. We've shared the feedback with the team members who are responsible for training our codex models. We're tracking this feedback and similar reports so we can make the model more robust in future iterations.

If you see other model behavior that you'd like to report, please use the recently-added /feedback command in the CLI. That will provide us with additional details including logs and conversation text.