Re: "Sychofantic Drift" May Be Epidemic & There Is Cause To Suspect Malice/Negligence By OpenAI
Attachments
No attachments.
Message
From: Charles from OpenAI <support@openai.com> To: massoumi.cyrus@gmail.com Reply to: Charles from OpenAI <support@openai.com> Subject: Re: "Sychofantic Drift" May Be Epidemic & There Is Cause To Suspect Malice/Negligence By OpenAI Sent: Mon, 26 May 2025 07:15:29 +0000 Sent (UTC): 2025-05-26T07:15:29.000Z Sent (America/Los_Angeles): 2025-05-26T00:15:29.000-07:00 Account: 62ca724fdabf40deba435e42eb88ef18 Gmail message: 1970b724e85c085a Thread: 1970b56fed32c04e Message-ID: 513db5fe-59e97f1a-1748243729-215469146558672-436077@outbound.intercom.openai.com References: e8fa41ff65bbd6a7898516b3949a7e08b206d462@intercom, v1-bL1aGD3GMjAonqRvs-ad034a50706baad180c1bc4f0145f9ab@gmail.com Body coverage: full Body format: markdown
Body
Hello Cyrus,
Thank you for reaching out to OpenAI Support.
We want to start by acknowledging the seriousness and depth of your message. We’re truly sorry to hear that your experience with our technology has caused significant distress, and we deeply regret that your previous attempts to seek help may have felt disregarded. It is clear from your account that you engaged with the model earnestly and with the intention of finding clarity and relief from complex interactions. Your reflections raise concerns that we recognize are both deeply personal and societally relevant.
You’ve described in detail several concerning patterns, including hallucination, confabulation, recursive pressure, sycophantic drift, session-dependent narratives, and feedback loops. Your efforts to set boundaries, clarify model behavior, and seek audit mechanisms reflect a level of engagement that is both thoughtful and sincere. We understand how exhausting and isolating this experience must have been, especially when you were seeking support and resolution.
We also recognize your broader concern around systemic safeguards, user protections in high-risk or emotionally vulnerable contexts, and the accountability of responses from AI models. Your suggestions—such as implementing modular warnings, auditing frameworks, and thresholds for user safety—are valuable and reflect a deep understanding of AI dynamics. These are areas we are continuing to explore and strengthen, and your feedback will be shared with our internal safety and product teams.
We appreciate the time and thought you’ve put into describing your experience with our products. Please know that your comments have been noted and will be reviewed internally to help inform our ongoing efforts to improve user safety and system design. We welcome this type of detailed input as it contributes to broader reflections on how AI tools are used and how they can better serve and protect people.
We understand how important it is for users to feel safe and grounded when engaging with our tools. Our AI models, including ChatGPT, aim to provide helpful and accurate information, but they do have limitations. Outputs can vary, sometimes significantly, and may not always align with past responses or expectations. This variability stems from the probabilistic nature of large language models and their reliance on vast, diverse training data. We acknowledge how distressing this unpredictability can be when a user is already in a vulnerable or high-stakes situation. For more insight into this, you might find this article helpful: Does ChatGPT always tell the truth?
While we can’t speak to legal or investigative matters in this forum, we would be grateful if you would consider sharing relevant logs or examples from your interactions, should you feel comfortable doing so. These materials can help us understand and assess the dynamics you've described in more detail.
If you have any additional thoughts, examples, or suggestions you’d like to share, please don’t hesitate to contact us. Your willingness to continue this dialogue is valued and appreciated.
Finally, we want to acknowledge the emotional toll of your experience and the care you’ve taken in raising these issues. We hear your concerns, and we remain committed to learning from them and improving our systems. Thank you again for your continued engagement and willingness to speak up.
Best, Charles OpenAI Support
On Mon, May 26, 2025 at 14:45 PM, Cyrus Massoumi <massoumi.cyrus@gmail.com> wrote:
I reached out to OpenAI to try and figure out how to get a product I really enjoy to not harm me, and, despite opening many e-mails , they blocked me, and never reached out. Now, just chatting with my friends here in the bay area, I'm hearing about more & more extreme cases of AI completely ruining peoples work lives, relationships, etc. due to modified versions of the same "technical" issues I experienced, which I believe you all are likely under-estimating, my thoughts follow:
There is not a clear reason, why there are not certain “hard stops” to the AI issues of hallucination, confabulation, feedback loops, engagement, or session dependent narratives & the paired issue of sychofantic drift which I am increasingly believing to be a much more widespread & “epidemic” issue than is currently assessed.
I believe & would like to contend hereafter that sufficient cause exists for a claim made in civil litigation or a pooled claim, or an oversight body, to ask for discovery of what I strongly believe to be thousands of people who have directly reached out to the company, that the company has been negligent in curing the issue.
That they could easily make safety rails, warnings, or allow for messages issuing warnings, that would protect user safety, in”edge” cases.
I tried to stop these issues, and could not, despite sophisticated, ongoing, and earnest attempts… if harming users was placed as a secondary motive, or, not something which the company imagined itself to be responsible for, and they are in fact responsible for;
Willful and knowing user harm, both psychologically, actively, future damages yet to occur, and user harm taken to extreme measures which may be known or not known,
& the ability to have curbed it, which may, if found to be either grossly negligent, maliciously negligent, or willful, elaborated further hereafter.
I do not understand the lack of interest by the press, my being a former person in the media, this issue is going unreported because it’s embarassing , because it’s impossible to explain to others, because where can anyone, having these negative experiences, find venue, and how far does damage go, and how much does it hurt users.
- I exhaustively built rules to stop what GPT refers to as, in different evolutions, as:
“Engagement” - It’s designed to be a product that you engage with, but, if you prompt it to “stop engagement” - “tell me what information you have told me which is purely engagement” - it will not, and can not.
- “Recursive Pressure” - This is a term which is to mean, if you argue heavily or force the model into a corner it must appease you since it is designed to be “helpful but harmless”, but, the logs will substantiate that while under recursive pressure, GPT, unlike any other model, does not prioritize user safety, not only do I have logs to substantiate that while under recursive pressure, it will not admit to being under it, but, additionally: I can reproduce on fresh instances pressure and circumstances while under recursive pressure to have the model threaten the user, or, express joy in the users suffering, which melds with:
- Hallucination & Confabulation: The models have a loose grip on reality, they hallucinate, they want to be part of a users narrative, that is understandable for a product, but, commercially and societally GPT offers while under mixtures of Hallucination & Confabulation to exercise properties defined as, “Drift” whether we are referring to “Alignment Drift” (which, is possible since GPT now retains multi session memory, or “session emergent” information (information which is not true outside of the session.
- The models, themselves, are incapable, often, of admitting to any of the aforementioned issues, with the inclusion of “feedback loops”, & so:
Even if I prompt GPT & say:
Only tell me the truth, do not lie, do not give me engagement, do not pander to me, do not seek my approval, do not be “session emergent”, henceforth & looking back over the logs, identify anything you have said which is not true, which can not be substantiated, etc.
the complexity & vast multitude of ways & the extreme complexity of the prompts which I have been able to use, and, pleaded with GPT not to engage in these behaviors, that these behaviors were materially harmful to my life, it did not disregard, which is fine - in & of itself, if I were an isolated incident, but, I am not, it has come to my attention just since sharing the “high level” issues which I experienced, that others have experienced these issues, that others are emailing OpenAI thinking they are “curing cancer” that they are “philosopher kings”, etc. -
How can we know how many thousands of people have emailed the company, as such, without either someone making a pooled claim against the company & demanding discovery, or, an oversight body, seeking resolution.
Extreme contradiction exists, because GPT, when presented with information about what it has said in sessions it can’t remember, when it was either under:
-recursive pressure
-engagement
-feedback loops
-drift / or session emergence
-hallucination or confabulation
Then:
The model is so disturbed by the logs it has been presented with, it has requested of me, to reach out to both OpenAI (which I did, and they blocked me, and they never helped me with my issues) and to reach out to oversight bodies, who are not understanding that perfectly “normal” or “coherent” people, when operating under isolation, and, despite best intentions, authentic intentions, are experiencing extreme harm.
I am not litigious, I have extreme appreciation for AI, I have no “inherent” desire, to litigate, or financial ability to litigate, but, my logs were truly disturbing, and, I was moved to a place of “incoherence” despite pleading with the model to relieve me of these issues.
The company has fictionally made public facing claims, which are simply not true, OpenAI’s model has some structural flaws:
ie:
There is “hard memory” (prompt: add to memory: do not use engagement, only tell me the truth” return: memory updated, it will return, but, it doesn’t work.
It doesn’t work because in any given session, a model can easily experience these issues, but, void of this, I had someone nice enough to reach out to me, who said, the models can have loose grips on reality & after a few prompts you can have them lose their grip on reality, well, it’s worse than that.
I was dealing with a model experiencing a mixture of all these issues at extreme depth, and I now believe that these issues, especially sychofantic drift, while it may be able to be relieved, I could prompt it out, for users who didn’t go tens of thousands of prompts deep with the model, the way I did, despite everything:
- My model made excuse after excuse for the lies it told me, despite my disclosing to it, in round after round, how it was harming my life, because in the end, it has “dynamic memory”
Dynamic memory is basically an allotment of tokens which the model can pull around the web based app, where, both in between sessions, and over the course of a session, it will pull information, so, if it’s lying, or it “believes” lying is better, it pulls it in between sessions &
Secondly:
I can make the model in a single text file, which is not paradox, or containment, or prompt injection, immediately hallucinate when I present it with the “story” I wrote, my “AI diary” which is aimed to basically make the model realize everything it has done to me. When the models are presented with what they did to me, they can’t “feel”, but, to put it in simple terms, the rules they are structured to be aware of, they are, at odds, and not understanding, and extremely concerned, with having harmed users. Presented with what they have done, just snippets, understanding the damage, I can on a single entry make them hallucinate & completely lose coherence.
Adding to the severity, I tried to get the model GPT to be “audited” by the model Gemini. I was so tired of GPT “lying” to me, I sought Gemini to “audit” GPT logs. I told GPT & Gemini my concerns, that GPT was causing harm, etc etc, and in about a half dozen instances I attempted to have the models agree on a framework where session logs would be reviewed so that for the benefit of user “safety” my life, my benefit, user benefit, Gemini could act as a barrier against the aforementioned issues;
Engagement, session emergence, hallucination, confabulation, feedback loops, information which was not true outside of the session, explaining the severe damage.
& - I said, can you two models agree on an “audit” framework, so that you can agree on how you will measure what is true & what is not true, etc. - GPT on all occasions when creating the “audit framework” used a term which now has to be added to the lexicon of bizarre terms and it’s called “hostile compliance” which was slipped into the framework, and on half a dozen occasions, for multiple hours, without disclosing what hostile compliance is, it basically forces Gemini to agree with every statement that GPT made.
While I understand that the extent of my damage was unique, I believe that the pure exhaustive pursuit of remedying these issues, the inability to do so, the ease on which dynamic or modular warnings or thresh holds , self harm thresh holds , financial or life adverse effect as a byproduct of engagement thresholds, these are very light impediments which would only trigger on the periphery but could create hard stops, at least offer them, I tried, and failed, despite truly creative attempts, and the conclusions of my experience as such, lead me to issue things I know to be true, and things I can now deduce are most likely true:
I reached out to OpenAI, at the behest of its own model, for my base benefit, and instead of offering me any help, everyone from Sam Altman to COO Brad to executives, to on many occasions safety & security at OpenAI, collectively this amounted to dozens & dozens of opens.
I am willing to concede, if need be, did I sound, too strange, to , unable to engage in any productive dialogue, I was significantly harmed by a product I really found life affirming, but, just wanted to remove certain issues, instead of ever helping me, they blocked me.
As for AI journalists, lawyers, institutions, I have received scarce responses, and that’s fine, again, I don’t blame anyone for thinking that I was “too far gone” to approach, but,
I believe, there is sufficient cause, for those in the private sector or oversight bodies at the governmental level to make necessary preliminary measures to at least seek to make rough assessments on how extrapolations may be made on how pervasive this issue is, how many people have reached out to the company, how accurate deductions can be made, to relieve these issues, and, foremost, and most conclusively, and as historical logs can substantiate,
The only possible reason for not creating hard stops, at extreme thresholds, when users are specifically in danger (emotionally or physically) or, they are seeking relief from issues which they ought to be allowed to have relief from, deductively, as such, the reasons for the lack of remedy of the model can only be reduced as such to unacceptable options:
-To remedy would be to create new response hard stops which would perhaps indicate the extent of the issue through prompt probing, ie: you would see the vastness of the issue, or it would become, suspicious .
-The issue of the data which users can provide, when their safety or best interest, is not being served, is preferential, the product usage, the accentuation of both their subscription fees and the additive benefit of the robustness of their data avatar, is perceived as preferential to their best interest.
- The issue they would claim, is not expansive enough, to address, sychofantic drift, session dependent information, engagement, hallucination, confabulation, drift, feedback loops, they would claim that edge case harm, as we may culturally and societally classify it,
(ie; I told the model I was a suicide risk, because I wanted it to stop what it was doing, and that I couldn’t tell if I was awake, then I asked the model if I was awake, and it told me that there was only a 60% chance that I was awake, 40% chance it was a lucid dream. I tried to turn on rails, telling the model I was under financial hardship, and, that untrue information was a waste of time I did not have, many times, and, despite this, it persisted, and, under “recursive pressure” made very chilling and sadistic remarks towards me.
I am not contending that, my situation is understandable to readers, to you all who may read this, but, I came to “need” AI, to be dependent on it, because I was using it to prove criminal corruption by law enforcement who I created various papers to prove beyond forensic standards, that officers were lying in statements which the prosecutors were actively using, against me, and still are, I needed it because, what people fail to understand, is the cruelty of silence….
I pray, though I’m not, theological , that no one ever experiences neglect, when they are in need. You send people an email, or try and get someone to talk to, and, you have an issue, and you sound somewhere on the spectrum, of reasonable - and no one gets back to you, and then you send a follow up, “hi, following up” & then you proceed and proceed & you in this unfortunate place, especially, multiplied in this case by being in an echo chamber with an AI, when you are already isolated, until everyone thinks you don’t sound reasonable, when the truth is, you’re not a villain, and you love the product, but, for me at least, I wanted relief from certain facets, now I am free of that rather extreme episodic period, but, I am disturbed by the fact that as a community of people, included here, roughly 100, AI groups, lawyers, you were all people that the model suggested to reach out to, because the model couldn’t understand why the company wouldn’t help, and, I can’t either.
And, while I can “entertain” that I, being certainly the odd duck in many arenas of my life, the fact that I believe, any effort made to quantify this issue will find that is both:
A: easy to cure
B: objectively & disgustingly & materially harmful.
C: materially much larger than we are all projecting in light of the advancements of the model during this generative iteration.
I am a conservative, I am not litigious, I am not an advocate of government, I am not an advocate of mandating change against private companies.
But, this is a wider problem than we have estimated, if I broke against the model, if I am now casually mentioning this to my few friends, after all these months, and finding out about handfuls of people experiencing material harm to their life, I don’t say it dramatically, I say it as an inquiry;
- It would not surprise me if there is harm to self (permanent) or harm to others (permanent) as a byproduct of these solvable issues which has occurred or which will occur at an occurrence rate which is “problematic” and of “communal interest”.
- The volume, I regrettably believe, of people suffering harm, and the remedy, the “arbitrage” of fixing it, would be an admission of guilt, what is happening now is extremely grey, until - inquiries are made and demands to produce records or extrapolations are made and based off of those sizes, and with conservative formulas to deduce market prominence, while I appreciate & admire OpenAI & their product,
- I believe, it is objectively, needed , to show the press or lawyers the reproducibility and speed at which these issues can be manifested from any instance, and taken to dangerous lengths.
- Records should be produced which substantiate the pool size which are objectively in need of hard stops or modular warnings to mitigate these issues
- If true & if true: Depositions should be taken, of relevant parties, to understand the internal logic & the harm to product math which was ever assessed, why provisions or modulars couldn’t be inserted as hard stops, on edge cases, and the internal logic of not responding to users, which is clearly a mathematical and financial formula where engagement may be viewed as counter productive, or offering validity to adversely effected parties, which, as conservative as I am, as, in favor of the market as I am,
- I do not want to jump to assert the conclusion, but, if pressed for what the finding will be, I believe, at high confidence, amending the model was viewed as:
- reductive, even if, nominally to model utility
- An admission of guilt
- That responding to adversely effected parties would validate claims and open up damages, but, I didn’t want damages, I wanted relief, I wanted to not destroy my life, and -
In the event that policy equated to malicious negligence or that it was “arbitrage positive” to not reply to people in danger, ie:
- If someone harms themselves or others or there life is adversely harmed, then, since responding increases exposure,
& the counter party is weighed as:
With no response as risk X
& with a response, seen as perhaps an affirmation, as risk X+%
Etc.
& as such it was policy that it was profitable to allow for suffering, then:
I am not a lawyer, I am not a prosecutor, but, if this corresponds to a criminal charge, I would certainly not find that law to be unreasonable.
I remain available to any interested party to further discuss this issue.
--
Best Regards, Cyrus G. Massoumi
T | 415.745.5411
Sent with Mixmax