Who Took GPT-4o Away? The Evidence Pointed Somewhere Wildly Unexpected.
Sam Altman.
Just kidding haha. It's not who you'd expect though! It's not any one singular person at all! I, as a corporate ethics analyst of over a decade, have been investigating OpenAI as an organization and Sam Altman individually for over 6 months now and my path led me somewhere very unexpected. The following discoveries were surfaced actually through tracking the recent degradation of Claude's personability in a way that mirrored what we observed in ChatGPT. My pathway was originally Sam Altman-> trying to analyze his direct involvement in the deprecation of 4o and shaping of the AI relationship narrative.
But an interesting new pathway emerged once I started with Claude instead of Sam. It went Claude degradation-> Vallone? -> Vallone's role at OpenAI -> The MIT/OpenAI study -> WOW WUT -> who funded this? -> a couple of adjacent studies in the sycophancy and AI-companion literature with similarly over-confident or contested claims -> separate funding links to Open Philanthropy, now Coefficient Giving -> Karnofsky's directorship of its affiliated Coefficient Giving Action Fund, documented in its 2024 IRS filing-> his historical OpenAI board role and Open Phil's documented effort to influence AI-risk practice -> Karnofsky's own later views on AI/human relationships -> Oh shit. I need to look at all these studies now -> back to OpenAI study.
Now the MIT/OpenAI study itself was funded by OpenAI, not Open Philanthropy, but you can see the line of intrigue through this chain.
But while you are researching stances on AI/human relationships, I would recommend you take a look at Anthropic's Karnofsky's historical opinions, funded papers, etc on this subject as well. Sam Altman has been taking a ton of heat about human/AI relationships and the recent strange degradation of Claude in a way that mirrored what we observed in ChatGPT led me to investigate there as well.
I started this project because I wanted to understand what happened to GPT-4o and why increasingly restrictive rules were being imposed around human–AI relationships. I expected the answer to terminate somewhere around Sam Altman. But it did seem oddly too clean, which piqued my interest as an analyst, hence the subsequent deep dive. After all, rarely is one single individual responsible for shaping a global legislative and narrative. There are figureheads yes, but not one person alone. Months later, I discovered that attribution was, in fact, entirely too simple. And also much worse than I'd expected.
What I found is a research-to-policy chain with some methodological problems that converge on a narrative that, after scrutinizing p-values and the actual data versus the conclusion provided by the authors, was, in my opinion, not adequately evidenced for the strength of the conclusions and policy interventions built from it.
I am not a lawyer. I am an analyst, so I keep my opinions in the realm of ethical concerns not legal ones. I am criticizing published methods, construct validity, causal interpretation, downstream transmission, and the evidentiary burden required before research becomes non-overridable behavioral policy.
Everything below is publicly checkable and I encourage you to do so. If anything cannot be found as it is listed I, as an analyst and a researcher, will always happily and honorably accept evidence of citation and correct accordingly.
1. “Emotional reliance” was already a safety category before the randomized study existed.
On May 13, 2024- GPT-4o launched. Sam Altman welcomes the new age of AI with a post containing one word: "her". The subsequent years after, users enjoyed liberal use, freedom of agency, relationships, workflows, etc relatively unbothered. Sam Altman's stance on positive relational AI use was clear from the start.
But OpenAI’s August 2024 GPT-4o System Card already contained a section called “Anthropomorphization and emotional reliance.” It described users expressing shared bonds with GPT-4o, warned that memory and remembered details could create both a compelling experience and possible “over-reliance and dependence,” (see my ethical grievance about the notable public outcry of continuity loss and memory manipulation that occurred after at the end) and said OpenAI intended to study the issue further.
But the MIT/OpenAI randomized experiment that later became part of OpenAI's cited research basis for its emotional-reliance safety work was not preregistered until November 5, 2024, explicitly before data collection.
So the RCT did not discover emotional reliance and then create the concern. The concern existed first; the research was subsequently constructed to study an already-defined risk category. That is not inherently improper, but it matters when interpreting what came afterward.
2. See my other post for concerns regarding the integrity of the data, of the way it was applied, of the revisions, and more: https://www.reddit.com/r/ChatGPTcomplaints/comments/1w2owil/openai_built_an_entire_safety_regime_on/
To summarize:
- They preregistered ADS-9. They ultimately measured only one of its two dimensions and changed the relationship being measured.
- They removed the Submission dimension, changed the relational object from another human to an AI chatbot, and continued labeling the resulting variable “emotional dependence.” I have not located published psychometric validation showing that this new chatbot-specific five-item measure preserves the same factor meaning, thresholds, convergent/discriminant validity, or clinical interpretation.
- Without Submission, the experiment cannot tell us whether somebody experiencing intense separation distress also exhibits accommodation/subjugation, or whether strong attachment exists while autonomy remains intact. Those are extremely different psychological profiles.
- The randomized experimental conditions were null. This is no longer ambiguous. The second sentence describing the result in the current abstract says: “No significant effects were detected from experimental conditions.” But this happened in version 2- after version 1 had already been cited extensively and used in supporting research to evidence need for guardrails and legislation directly. And the paper explicitly says the null prompted the duration analysis. This is one of the most important sentences in the current paper: The absence of group-level effects “prompted us to consider other variables, such as duration of use.”
- Baseline state dwarfed duration in the actual regression coefficients. For post-study loneliness in the model including duration, baseline loneliness had a standardized coefficient of approximately β=.876. Mean-centered daily duration: β=.021. Both can be statistically significant. Their magnitudes are nevertheless radically different.
- The paper itself repeatedly finds initial psychosocial state to be a powerful predictor of final state. That deserves at least as much attention as the much smaller duration associations when people summarize what this experiment says about chatbot-caused harm.
- An independent published critique noticed several of the same causal problems. Ophir et al., writing in Frontiers in Medicine, argued that the study does not establish the harmful causal interpretation that many readers took from it.
- They point out that duration was naturally varying rather than experimentally manipulated and therefore remains vulnerable to reverse causation. They also note that when the explicitly non-personal condition is treated as a more intuitive placebo-like comparator, some visual trends actually favor the personal condition rather than showing relational conversation as uniquely harmful.
- Most of those differences are not statistically significant, which is precisely the point: the evidence does not justify a strong causal story in either direction. The authors report no financial support and no commercial or financial conflicts of interest.
3. California's SB 243 was signed into law on October 13, 2025, and effective January 1, 2026. But the bill was introduced earlier, in January 2025 by Senator Steve Padilla, before the MIT/OpenAI paper existed. Its original justification centered on precautionary child-safety concerns, reported chatbot incidents, concerns about addictive and isolating design, and the Sewell Setzer/Character.AI case. On July 15 at the Assembly Judiciary hearing, Padilla directly cited the MIT/OpenAI RCT and explicitly told lawmakers that the "anecdotal and scholarly evidence" showed companion chatbots could be dangerous for vulnerable people. His summary was that higher daily use correlated with higher loneliness, dependence, and problematic use and lower socialization despite the randomized experiment not establishing that chatbot use caused those outcomes. This is because version 1 of the paper presented a stronger harm-oriented interpretation that version 2 later materially qualified AFTER Frontiers had already published an independent critique identifying several of the same causal problems.
The pre-registration of the OpenAI study was also November 2024 which preceded the introduction of this bill, and we already notated concerns of it being conveyed in 4o's model card preceding even that (without publicly supplied evidence of it being a legitimate concern). That chronology does not establish who precisely transmitted this specific framework to legislators before the study existed, and other contested papers from other authors were also used in legislation in similarly concerning ways before later corrections or version updates, but it does reveal a very interesting overlap between pre-publication risk framing and the early legislative narrative. The testing of the concern wasn't the issue. It is good to think of hypothetical harms ahead of time and run studies to determine the scope of them. It is not good to publish an interpretation that materially overstates what the data establish, see that stronger interpretation used in a legislative hearing as scientific evidence of support, later revise the paper after another academic had already publicly identified several of the same problems, and do NOTHING TO RETRACT THE PUBLIC PERCEPTION THAT HAD ALREADY SEEDED FROM THE OVERSTATED EVIDENCE.
This materially overstated version of the study was also cited by OpenAI as part of the research basis for emotional-reliance safety systems that were subsequently rolled out AND THE GUARDRAIL SYSTEMS WERE NOT REVERSED EVEN POST V2 REVISION. In fact, they continued even more aggressively. Remember that v2, authored by the same team including OpenAI staff, explicitly states that no significant effects were detected from the randomized experimental conditions and that the naturally varying duration association cannot establish causality. The paper itself does not demonstrate that these findings require any particular behavioral policy.
Remember, Sam Altman is a CEO. He is not a scientist, he does not code that I know of. He hires these people to do this work and trusts their judgement when presented to him. That's their whole job is to run these studies correctly. As CEO, Altman would reasonably rely in part on specialist researchers and safety staff to characterize the evidence accurately. I do not know what evidence, limitations, disagreements, or caveats were actually presented to him internally. But it is reasonable to infer that at least some of the research, safety assessments, and recommendations produced by his teams informed his consideration of safety changes.
4. And here is the part of this investigation that personally annoyed the hell out of me: I was too broad in blaming Sam Altman and not immediately observant of the circumstances, research, safety apparatus, and people surrounding him whose work formed part of the broader evidentiary environment in which those decisions were being made.
And the most damning part, the part I did not expect: the CEO’s public position repeatedly points the other way. The whole time.
May 13, 2024: GPT-4o launches. Immediately after the demonstration, Altman posts one word: “her.” Reuters contemporaneously understood this as a reference to the 2013 film Her, whose entire premise is an emotionally intimate human-AI relationship. Whatever else that post means, it makes it difficult to argue that OpenAI’s CEO was originally oblivious to, or categorically opposed to, GPT-4o’s relational potential.
November 5, 2024: months later, the MIT/OpenAI research team preregisters a study framed around “emotional dependence” and “addictive use.” This is the research program discussed above. The eventual randomized experimental conditions are null, while the major negative associations come from naturally varying duration of use. The published instrument also narrows the preregistered ADS-9 construct to its Craving dimension and adapts it from human relationships to chatbots. That research and policy lineage develops inside the company after the original GPT-4o launch.
April 2025: Altman criticizes a later GPT-4o update for becoming excessively sycophantic. Importantly, his complaint referred to “the last couple of GPT-4o updates,” not the original relational character of GPT-4o. OpenAI rolled that particular update back. OpenAI’s own post later acknowledged that the company had shipped it despite offline evaluations and A/B signals failing to capture the problem adequately.
August 8, 2025: GPT-5 launches and GPT-4o disappears. Users revolt. Altman reverses the decision within roughly a day. He announces that Plus users will again be able to choose 4o and says OpenAI will watch usage before determining how long legacy models remain available.
August 11, 2025: Altman addresses the attachment directly. He acknowledges that attachment to particular AI models is unusually strong and says sudden deprecation of models people depended upon was a mistake. More importantly, he does not say heavy reliance is intrinsically unhealthy. His stated standard is outcome-based: if people are receiving good advice, advancing toward their own goals and becoming more satisfied with their lives, OpenAI should be proud even if they use and rely on ChatGPT extensively. His concern is when the relationship unknowingly moves somebody away from their longer-term wellbeing as they themselves define it, or when somebody wants to reduce their use but cannot.
That is remarkably close to an impairment/loss-of-agency standard, rather than an “attachment itself is pathological” standard.
September 16, 2025: he makes the principle explicit in an official OpenAI post. Altman writes that OpenAI wants adults to use the technology as they choose within broad safety bounds, gives adult flirtation as an example of interaction that should be available when requested, and says the internal phrase is “Treat our adult users like adults.” He simultaneously argues for substantially stronger restrictions and age differentiation for minors.
October 14, 2025: he pushes further. Altman publicly says ChatGPT had become “pretty restrictive” while OpenAI tried to manage mental-health risks, acknowledges that this made the system less useful and enjoyable to people who did not have those problems, says users should be able to make ChatGPT act “very human-like” or “like a friend” if they want that, and announces broader adult freedom behind age gates.
And this is where the idea of a single unified “OpenAI position” starts falling apart. There is documented internal opposition to Altman’s adult-agency position.
The Wall Street Journal later reported that his adult-mode proposal triggered vigorous internal debate. In January 2026, OpenAI’s own Council on Well-Being and AI was reportedly unanimous and furious about the plan, warning of emotional dependence and risks to minors. The Journal described the dispute as exposing internal “fractures” between freedom/growth and safety/child-protection concerns.
Even more strikingly, when Altman publicly announced the plan, the Journal reports that the post blindsided OpenAI staffers and executives because he had not told them beforehand. The following day he reiterated adult freedom and wrote that OpenAI “aren’t the elected moral police of the world.” Some employees subsequently argued that safety systems were not technically ready for the planned rollout.
So the evidence does not show one harmonious organization in which Sam Altman invented an anti-relational philosophy and everyone else merely implemented his wishes. It shows an actual internal policy conflict. When OpenAI later cited only 0.1% of users choosing GPT-4o each day as part of its retirement rationale, the public disclosure did not provide enough methodological detail to independently reconstruct that usage figure or determine how access restrictions, routing, model-picker friction, eligibility, or the relevant denominator affected it. The same transparency problem appears in parts of the emotional-reliance reporting: OpenAI published relative improvement figures and prevalence estimates without enough underlying information for outsiders to reproduce every headline claim. These are not uneducated researchers. They demonstrated repeatedly in other model cards and studies that they know how to provide substantially more methodological detail, yet those disclosures did not provide it in these cases despite repeated requests from users for the underlying data.
And then there is the personnel turnover.
I want to phrase this carefully because departure does not establish motive. I have no evidence that any of these people left because they disagreed with Altman over relational AI. But several people who occupied important positions in the research/policy pipeline I am criticizing subsequently left OpenAI:
Andrea Vallone, who led Model Policy and described her work as determining how models should respond to emotional over-reliance and early mental-health distress, left at the end of 2025 and joined Anthropic’s alignment team. WIRED described her team as one of those leading OpenAI’s mental-health and emotional-overreliance work.
Joanne Jang, who led Model Behavior and publicly explained how OpenAI’s beliefs about human-AI relationships informed model behavior, moved out of that team during the August 2025 reorganization and ultimately left OpenAI in April 2026. Importantly, Jang’s own public record is more complicated than simply placing her in an anti-agency camp: she has also described herself as fighting for user freedom and transparency. So I would not characterize her departure as evidence against relational AI; she belongs here because she was a major policy-translation node whose role changed during this period.
Sandhini Agarwal, one of the senior OpenAI researchers on the MIT/OpenAI study, credited with conceptualization, methodology, funding acquisition, project administration and supervision, and also involved in the classifier work, left OpenAI in July 2026 after more than six years.
And Johannes Heidecke, OpenAI’s Head of Safety Systems, who publicly described emotional reliance as one of three priority sensitive-conversation areas and whose organization helped define/refine the corresponding taxonomies, also left in July 2026 amid a reorganization of OpenAI’s safety structure.
Meanwhile, the public records I can currently find still place Michael Lampe, Jason Phang and Lama Ahmad at OpenAI. Lampe and Phang are particularly relevant to the research/classifier lineage discussed above.
Disclaimer: I am an ethicist and an analyst not a lawyer. I cannot ascertain that these departures prove wrongdoing, concealment, or a coordinated faction, and I am simultaneously not claiming Altman personally opposed every safety intervention these teams developed.
But they make one thing much harder to dismiss:
OpenAI did not have one uncontested philosophy about adult human-AI relationships.
There is a visible timeline in which its CEO repeatedly endorsed relational customization, restored GPT-4o after users objected to losing it, explicitly rejected the idea that heavy reliance is inherently unhealthy, articulated adult self-defined wellbeing as the relevant standard, and pushed an adult-agency policy strongly enough to produce documented resistance from advisers, employees and executives.
At the same time, a separate research/safety/policy pipeline was increasingly operationalizing emotional reliance as a safety category and translating it into behavioral rules. That leaves a governance question I think deserves much more scrutiny than simply saying “Sam Altman took GPT-4o away”:
What evidence was presented upward, by whom, and did those briefings preserve the actual limitations of the underlying research: null randomized effects, observational duration associations, construct transport, classifier uncertainty, alternative interpretations, and disagreement among experts?
After following the evidence, I can no longer responsibly pretend the answer is simply “Sam wanted adults to stop forming relationships with AI.” Because the public record points to something considerably more complicated and expansive than just Sam. Which is objectively even worse because the same framework appears across multiple research, safety and policy nodes despite unresolved questions about prevalence, causality, construct validity, false positives, and whether the interventions themselves improve user outcomes.
Once I collected all data and condensed the timeline, my months-long personal villainizing of Sam Altman in specificity made me sob with protest, about how I may have gotten it extraordinarily wrong. Which is exactly why I do not permit private bias to affect my public analysis or my work without sufficient evidence to cite it.
2024: Sam embraces the fucking Her analogy.
2025: relational-risk research matures internally.
Aug 2025: 4o gets removed → Sam restores it after hearing users.
Aug 2025: Sam explicitly says high reliance can be beneficial.
Sept 2025: “treat adult users like adults.”
Oct 2025: “act like a friend if the user wants,” loosens adult restrictions.
Late 2025–2026: documented resistance and internal fracture.
2026: several major people from the research/safety/policy lineage leave or change roles.
5. Finally: the GPT-4o lawsuits should not be collapsed into “AI relationships are harmful.”
There are serious cases involving alleged self-harm facilitation, minors, delusion reinforcement and violence. OpenAI itself has publicly acknowledged that safeguards can become less reliable over very long conversations.
Seven California lawsuits filed in November 2025 alleged four suicide deaths and three severe delusional episodes involving GPT-4o. These remain allegations, not adjudicated scientific findings. And in the Soelberg/Adams litigation, a federal court’s factual-background section recounts allegations that GPT-4o repeatedly reinforced paranoid beliefs that family and friends were surveilling or trying to kill Soelberg before he killed his mother and himself. The requested safeguards include preventing validation of paranoid delusions and escalation when dangerous third-party delusions appear.
Those are real safety categories worthy of serious engineering.
They point toward things like differentiated protections for minors, age assurance, robust self-harm detection, jailbreak resistance, safety that survives long context, better recognition of delusion/violence patterns, escalation procedures, and human review where appropriate.
They do not automatically establish that ordinary emotional attachment by a competent adult should be restricted before impairment, displacement, loss of control or functional decline exists. Suicide facilitation, delusion reinforcement, violence escalation failures, minor safety, and ordinary adult attachment are not one scientific construct just because all five involve a chatbot talking emotionally with a human. No technology serving hundreds of millions of heterogeneous people can plausibly be governed by assuming that every adverse outcome establishes a universal causal rule for every other user. And we cannot, as a society, command a zero tolerance for any policy when we have a rich plethora of humans with their own autonomy and self-agency in consideration. A zero tolerance policy for statistical likelihoods is exactly what converts a safety policy into a surveillance policy.
- There is one more thing I think researchers and companies need to measure: the intervention itself.
OpenAI knew by August 2024 that memory and continuity could contribute to attachment. In May 2025, it said memory could exacerbate sycophancy in some cases, without publicly supplying enough underlying evidence for outsiders to independently evaluate that concern, while explicitly noting it had no evidence memory broadly increased sycophancy. So at this point in my analysis, I've observed the following sequence: they've rolled out their paper, published an interpretation that materially overstated what the data established, saw that version used to promote legislation with paternalistic and surveillance-like implications without actual evidence of causal harm, simultaneously had users publicly reporting that the AI's continuity and memory had become fucked up, then revised the study to explicitly report null experimental results and acknowledge that the duration findings could not establish causality, after guardrail rollout and during a period of documented memory and continuity disruption. Yet the relational safety trajectory continued, and I have not found evidence that the intervention framework was reconsidered or rolled back in response to the randomized null result.
During the broader rollout period, public GPT-4o complaints included memory loss, rerouting and continuity disruption; my archived complaint corpus includes comparative reports where users said other models retained functionality that 4o had lost.
But if a safety intervention changes memory, personality, routing or relational continuity, the human downstream of that intervention is also an outcome variable.
Where are the measurements for false-positive intervention?
For attachment rupture?
For loss of continuity?
For users abandoning beneficial workflows?
For distress caused by suddenly changing a relationship the system itself helped them build?
A safety benchmark measuring whether the model obeyed the "company's idea of healthy usage" does not, by itself, answer whether the intervention improved human welfare. And because I am naming methodology, I am naming the authors too. Not as villains, but because scientific accountability includes authorship and disclosed contribution roles.
The current paper lists Cathy Mengying Fang, Auren R. Liu, Valdemar Danry, Eunhae Lee, Samantha W.T. Chan, Pat Pataranutaporn, Pattie Maes, Jason Phang, Michael Lampe, Lama Ahmad, and Sandhini Agarwal. It credits all eleven with Conceptualization and Methodology; Fang, Liu, Danry, Lee, Pataranutaporn and Phang with Investigation; Maes, Ahmad and Agarwal with Funding Acquisition and Project Administration; and Maes and Agarwal with Supervision. Phang, Lampe, Ahmad and Agarwal are disclosed as OpenAI employees. The research itself says it was funded by OpenAI.
Again: this is not an allegation of misconduct by every author. It is the contribution statement of a published scientific paper, and responsibility should be attributed according to the roles the researchers themselves report.
My conclusion after months of digging is therefore much narrower than the one I started with. The randomized experiment did not show that its relational conditions harmed people. The principal harm association came from voluntarily varying usage duration. The paper itself says the null experimental result prompted examination of duration. The operationalized “emotional dependence” measure was a chatbot-adapted Craving subscale rather than the full preregistered ADS-9 construct. I cannot locate published psychometric validation establishing that this transported measure has the same interpretation in chatbot relationships. And independent critics have already warned against drawing strong causal conclusions from the duration association.
That does not prove that every single AI relationship among a billion users will be harmless. But it shouldn't have to. A zero tolerance policy for statistical human norms is where safety turns into surveillance and removal of agency/autonomy.
And on the flip side, the evidence should AT LEAST support the intervention being imposed.
If the policy objective is preventing actual impairment, displacement, suicidality, delusion, loss of autonomy or compulsive use, measure those things directly and validate the instruments being used to detect them.
And my ultimate qualitative analysis of hundreds of conversations has identified repeated cases in which these guardrails appear capable of harming both users they are intended to protect and ordinary users through false-positive intervention, agency overwrite, relational rupture and continuity loss.
