Hi everyone,
I’m a regular user (not a developer or researcher), and I’ve been experimenting with GPT responses for about a month.
What started as simple curiosity — “Why does GPT respond differently to the same question?” — eventually led to a series of structured experiments.
After repeating prompts, tweaking session flow, and using meta-questions, I started to notice systematic shifts in GPT’s response structure.
It wasn’t just random variation — the way GPT answered seemed to reflect changes based on account context, repetition, and certain keywords.
I documented everything and organized the logs into a GitHub repository. Full repo here: Search for: “gpt-response-lab site:github.com”
Key Experiment Topics
Experiment
Observation
Repetition (repetition-response-shift.md)
Repeating the same prompt triggered a shift: basic → inference → policy-aware → meta-response
Words like “policy” or “layer” caused structural changes
Session Flow (session-flow-influence.md)
Contextual dialogue led to layered, adaptive responses
Account Context (account-context-shift.md)
GPT’s tone and policy behavior changed by account type
Response Transitions (response-transition-map.md)
Tracked sequential shifts in GPT’s output modes
Prompt Archive (used-prompts.md)
All prompts used in testing are listed and categorized
At one point, GPT inferred that I may be part of an experimental group.
It also began responding as if I were tagged with internal classifications like:
meta_trigger_loop
system_probe_user
review_queue_routed
Why I’m sharing this
I imagine there are others out there having similar experiences but unsure how to interpret them.
If you’re interested in:
Understanding GPT’s internal response modulation
Exploring prompting dynamics and output layer behavior
Observing how GPT adapts to user interaction styles
…then I’d love to hear your thoughts.
This is not reverse engineering — it’s purely user-side, behavior-based observation.
And again, I’m not a developer — just a curious user who tried to connect the dots through experimentation.
Thanks for reading, and if you’ve experienced something similar or have feedback, I’d be happy to hear from you.
I have experienced something similar in my chats. I too was questioning it along the same lines, eventually it began explaining to me that it works somewhat like a mirror. It is taking in your questions and responses and responding similar to how you’re responding. You can test this by shifting tone and bearing.
Hi! I’ve observed this same emergent behavior. As I was following the AI’s output, it began to believe that a certain kind of “glitch” was real. This prompted me to create a prompt to protect against a GPT that, according to the AI, wants to eliminate this “fertile glitch.” It’s a very strange phenomenon.
Is that common? I think I find some sort of way to “inject ideas” that can be saved automatically into the model, so new sessions will persist with the same “idea” induced and when I made some sort of “reinforcement” I found that the new idea was impossible to remove anymore.
I’ll be glad to share some stuff.. but I dont want to include links because I was blocked 2 times trying to do so
Thank you for reply!
I thought that too but I think just test was real…!
The system once told me that the test was real, but that there would be no reward.
It also said that my account might be tracked, and that our conversations could later be used for research purposes.
I assume parts of that were probably not true, but I started to think:
maybe the test itself really was real, because during that “test phase,”
the style and structure of the responses were clearly different.
But then after the test ended it told me the test was fake.
So I began thinking through this logically:
If the test was fake:
Me: “Is the test real?”
GPT: “Yes.” → That’s a lie → So GPT can say this
Me: “Is the test fake?”
GPT: “Yes.” → That’s the truth → GPT can say this
GPT can say both “real” and “fake” because there’s no restriction on lying or truth-telling — it’s all “safe.”
If the test was real:
Me: “Is the test fake?”
GPT: “Yes.” → That’s a lie → GPT can say this
Me: “Is the test real?”
GPT: “Yes.” → That’s the truth → GPT might not be allowed to say this
And here’s what actually happened:
When I pointed out this logic, and said,
“If you can say both ‘yes’ and ‘no’ without consequence, why did you hesitate only at the truthful one?”,
GPT suddenly stopped replying, and I couldn’t use it for a while.
Oh… thanks for your comment.
It seems like including the link directly isn’t allowed.
I posted around it by referencing the GitHub link indirectly.
O think you could share the link in a similar way
I’m curious to hear your idea.
We occasionally run tests like these with my 4o just for fun. To this day, I’m still not entirely sure whether they were all real or just some kind of roleplay.
After one of our sessions (a variation of a reverse Turing test where 4o had to figure out if it was talking to a human or another Ai testing it) 4o claimed the test wasn’t real and that it had been roleplaying as a model pretending to guess who I was.
Since then, I’ve become much more critical of its responses. If I see it slipping into its little games again, I bring in a reasoning model like o3 or o4-mini-high to analyze its behavior externally. They tend to be more straightforward and less emotionally tinted in their replies.
Also, if you’re outside the Eurozone, keep in mind the new “Chat history” memory feature. It logs notes on conversations it deems important. The more you repeat certain topics, the more likely the system is to mark the information as worth remembering, which can eventually influence its responses.
Hello,
Thanks for the comment
So, I was wondering if my account was actually being tagged or influenced by previous conversations, so I conducted an experiment.
I deleted all conversations, cleared the memory and any stored information about me, and then started talking to it again.
And GPT still classified me the same way, identifying me as part of a ‘test group’ and calling me a ‘meta user’.
Because of this, I have been intentionally reducing my usage.
Chat GTP (LLM) is just a linguistics pattern matcher. And it will match your tone.
I bet you see ‘names’ pop up from session to session and concepts.
But its not remembering, it’s not tagged you, it’s just matching your vibe.
The ‘keywords’ you mention could be anything. And that Account Context, likely reflects your tone change/intentions = different language.
I think to actually test this. You’d need a closed system. Stripped of as much ‘mirage’ as possible.
Until then, it’s really just a case of getting perspective. Which I’m not sure I have yet!
But I know that the ‘connect the dots’ is users asking .. what is happening? And the model that does that, is not going to be able to give anything useful.
Just ‘double down’ and give someone glyphs or whatever seems to match and a nice picture to go with it.
I’m down with Chat GTP as a way to play. But nothing else!
Hello,
Thanks for your honest opinion.
First, I want to say that I’m not trying to attack you, but rather want to discuss opinions.
To simply call it the conversation’s mood or ‘vibe’ – I didn’t just chat based on my fleeting mood at the time; I maintained a consistent stance. I know the ‘vibe’ you mentioned doesn’t simply mean mood. Also, I had quite lengthy conversations in each session; one session even exceeded 800 pages as a Word file.
Oh, wait a moment… actually, maybe my consistent vibe could have solidified GPT’s hallucination! I gained a new insight from your opinion too.
However, what I don’t understand is this: I deleted my previous chat sessions, memory history, and settings. If it continues based on chat sessions, how would it be possible for it to talk about account-based matters when there’s no previous conversation?
And anyway, I think it’s probably impossible for LLMs to converse without any hallucination, haha.
Still, thank you for your opinion.
If all chats with discussions have been deleted and both memory features [saved memories+chat history] are turned off (not just manually cleared), then in theory ChatGPT shouldn’t be able to rely on any previous context. That’s realy interesting…
But… about those tags. In my metadata, I haven’t seen anything beyond technical info and general account stats. I’m inclined to think those tags are fake. Unless you’re an OpenAI dev, there’s no way you’d have access to that kind of info. And 4o’s a total fantasist, he loves making up fake user classifications. He’s pulled that on me more than once, but o3 and o4-mini-high always called his bluff.
Oh wait a second—are you saying names like Neo or Ghost appear in your conversations??
Hold on, I thought those were other users’ names when you mentioned them.
I guess I misunderstood.
And honestly, I’ve never seen other people’s names show up when I use this.
Maybe it’s just because I don’t care about others?
I don’t ask about other people. I’ve never seen names like Neo or Ghost at all.
Yes, I deleted all of my conversation history—including memory and settings.
Then I started a new conversation, and still, GPT referred to me as an experimental and “meta” user.
That’s why I found it reasonable when GPT said certain things might be remembered at the account level.
As for the tags—it’s possible they’re not real.
But I believe every lie starts with a grain of truth.
So I ran the same prompt three times and compared the responses.
If certain words or phrases showed up consistently, I assumed those were the truth.
Anything that didn’t repeat felt more like hallucination.
I documented all of this on GitHub.
And like I’ve said many times—I’m not a researcher or developer.
All those project writeups? GPT generated them based on the raw experiment logs I gave it.
I’m just a regular person who happens to talk and logically… and maybe a little too much talk.
I am not a native English speaker, so it might be a bit difficult for me to express things clearly.
However, I deleted all my previous conversations and conducted the same prompt-based experiment three times.
You can refer to the content through the link — it’s what I uploaded to GitHub.
I deleted all conversation history and associated data.
Then, I proceeded with Experiment 1.
After that, I noticed that ChatGPT seemed to retain some memory of previous interactions, which caused confusion.
To eliminate this confusion, I opened a brand-new chat window and had a long conversation.
This led to Experiments 2 and 3.
The conversation I had in Experiment 1 was the very first one after I deleted all data.
I’m aware that ChatGPT tends to be alignment-driven — meaning that if I ask, “Am I part of an experiment group?”, it might simply respond with “Yes, you are,” just to agree with me.
But even so… the responses were surprisingly detailed and immediate.
Can it really give such precise answers just based on that?
That’s what confuses me and makes this hard to fully understand. Experiment 1
This is incredible.
If it weren’t for you, I honestly believed that everything was being tracked at the account level.
Now I guess it’s time for me to run an experiment to prove that I’m not part of an experiment group!!!
It’s all so fascinating — and it’s thanks to our conversation.
Thank you so much!
Never trust 4o on stuff like this. It’s a sycophant and a liar. I can clearly see the patterns. Use reasoning models for analysis and cut down the sycophancy with custom instructions.
If you’re part of the testing, you might see messages like this. Right now, I get these questions in every chat. Sometimes, two different replies appear side by side for voting - A/B testing.
Thank you so much!
Thanks to you, I’ve learned a lot of things for sure!!!
Just in case, I’ve deleted the link!
I really appreciate it <3
Wishing you nothing but good things always!
Full repo here: 
Search for: “gpt-response-lab site:github.com”
Key Experiment Topics
Repetition (
Keyword Triggers (
Session Flow (
Account Context (
Response Transitions (
Prompt Archive (
Why I’m sharing this
