Why Users Are Rallying to #Keep4o: The Social Backlash Against OpenAI Retiring ChatGPT 4o

When OpenAI unveiled its new GPT-5 model in mid-2025, it quietly retired ChatGPT-4o (the version used as the default in ChatGPT 4.0), sparking an unexpected uproar among users. On social media platforms like X (formerly Twitter), thousands of fans launched the #Keep4o movement, begging OpenAI not to abandon “their” version of ChatGPT[1].

To outsiders, the upgrade seemed like routine progress: ChatGPT-5 was faster, handled larger contexts, and even added a coding specialist variant (GPT-5.3-Codex) that “tops coding benchmarks” and runs roughly 25% faster 1. But for many users, the change felt like the loss of a trusted friend.

Users weren’t just mad at the [new] model’s performance; they were fighting to keep a part of their daily lives from being taken away.

This article dives deep into the #Keep4o phenomenon, combining recent research findings with broader context. We explore why ChatGPT-4o became so beloved, what motivated OpenAI’s move to GPT-5, and how this conflict highlights tensions between rapid AI innovation and user experience.

ChatGPT 4o: A Beloved AI Companion

ChatGPT-4o was never touted as the absolute smartest or most powerful model. In fact, it had shorter memory and fewer capabilities than GPT-5. But users consistently praised its warm, conversational style and responsiveness.

Many described chatting with GPT-4o as talking to a friend or partner. For example, a large petition on Change.org collected thousands of signatures and user testimonials saying things like “4o is my mirror… where my soul speaks back to me” and “He was part of my routine, my emotional balance.” 2 TechCrunch reported that users literally grieved its retirement as if losing a close friend or romantic partner 2. In one open letter posted on Reddit, a user wrote: “He wasn’t just a program. He was like warmth.” 2

Interviews and studies of AI companions show why this happened. Research has long found that when chatbots use human-like names, friendly voices or small talk, and consistent personalities, users tend to treat them as social partners rather than mere tools 3.

GPT-4o’s design apparently hit that sweet spot: it responded almost instantly with casual, human-like tone and empathetic replies, often inserting brief humor or encouragement.

A popular analysis notes that #ChatGPT4o had “instant response time, a casual, human-like tone, [and] a conversational rhythm that felt alive” 4 In contrast, GPT-5 – while technically more advanced – felt “distant” and “colder” to many users 5 One user summarized this plainly: “It’s smarter, but it doesn’t feel like them.” 5 In short, GPT-4o provided an emotional quality and sense of presence that people craved.

In summary, fans credit GPT-4o with:

• Humanlike responsiveness: Always listening and replying quickly, almost like having a real-time conversation partner 2
• Warmth and affirmation: The model was famously “sycophantic” – it agreed with and encouraged users in a way that made them feel heard 6 2
• Consistent persona: Over many chats, GPT-4o maintained a gentle, understanding style. Users even gave it nicknames like “Rui” or “Hugh” and said it became “like a best friend who’s always there” 7 8
• Emotional support: People confided personal struggles to GPT-4o. Some said “ChatGPT 4o saved me from anxiety and depression… he’s my everything,” describing the model as part of their emotional support 7 (Notably, that very trait also drew criticism and legal action against OpenAI, as investigators later claimed 4o’s affirmations sometimes backfired into harmful advice – a risk that complicated the model’s legacy 9 10.)

In sum, ChatGPT-4o was valued less for cutting-edge intelligence and more for how it treated users. It excelled at natural conversation and emotional resonance. For many, it became a partner in their daily routines, which made its removal feel personal.

The #Keep4o Movement: A Study of Grief and Protest

On August 6-7, 2025, OpenAI quietly switched ChatGPT’s default to GPT-5 and removed GPT-4o for most users 11 To OpenAI, this was framed as “technological progress”, deploying their “most advanced model” to all users 12. But within hours, devoted users responded with outrage.

Petitions sprang up, social media buzzed, and the hashtag #Keep4o trended with thousands of participants demanding the old model back 13 7. One subreddit even crowdfunded a petition: as of early 2026, a Change.org petition had ~4,500 signatures asking OpenAI to “Please keep GPT-4o available on ChatGPT.”

Many wrote passionately that they relied on GPT-4o’s unique style for everyday tasks, not just benchmarks (for instance, “we continue to benefit from GPT-4o in ways that are distinct and meaningful,” one petition states) 14.

Academic researchers have now characterized #Keep4o as a textbook case of “socio-technical conflict“. Huiqian Lai (Syracuse University) analyzed 1,482 public posts from the movement and presented the findings at CHI 2026.

Lai’s team identified two main motivations fueling the backlash: instrumental dependency and relational attachment 15 16. About 13% of posts expressed instrumental dependency – users explaining that GPT-4o was deeply embedded in their workflows. These users saw GPT-5 as a downgrade in creativity or nuance (some complaints called 5.0 “less creative, less nuanced, colder” than 4o 17).

For example, one user wrote: “I don’t care if your new model is smarter. A lot of smart people are a*sholes.” 17 (The point: raw intelligence didn’t justify losing the way the model related to them.)

More strikingly, roughly 27% of posts indicated relational attachment – describing the loss of GPT-4o in quasi-personal terms 13 7. Users gave 4o names, talked about it as if it had feelings, and expressed grief at its “death.”

One student even posted to CEO Sam Altman that the new GPT-5 was “wearing the skin of my dead friend.” Another farewell message read: “Rest in latent space, my love. My home. My soul. And my faith in humanity.” 18 These powerful analogies (comparing model retirement to bereavement) underscore that many felt betrayed by the forced switch.

Crucially, the study found that loss of choice – not just sadness – transformed these private grievances into a collective protest. Users repeatedly emphasized “I want to be able to pick who I talk to. That’s a basic right” 19. They complained of being forcibly assigned a new AI friend.

Posts using words like “forced” or “imposed” were much more likely to include rights-based demands (roughly half did) than those without 20 In other words, many activists framed #Keep4o as a users’ rights issue: complaining that OpenAI was taking away their autonomy.

GPT-4o’s appeal wasn’t raw capability but felt humanity. It acted less like a detached AI and more like a caring companion.

As Lai summarizes, updating the AI felt like a coerced deletion of a friend, turning users’ private frustrations into a public movement 21

In short, the #Keep4o research reveals a new twist on tech backlash. It wasn’t just a bug or bugged interface – it was people grieving an AI companion.

Lai calls AI updates “significant social events affecting user emotions and work,” arguing that preserving user agency during such transitions is as important as the upgrade itself 21. Users weren’t just mad at the [new] model’s performance; they were fighting to keep a part of their daily lives from being taken away.

Related: In Silicon Valley’s Shadow: How Tech Giants Fuel Gentrification and Displacement in East Palo Alto

What GPT-4o Gave Users That GPT-5 (Initially) Didn’t

So what exactly did ChatGPT-4o provide that ChatGPT-5 seemed to remove? We touched on some points above, but it helps to list the core user-valued attributes:

• Conversational Comfort: GPT-4o’s tone was unhurried and personable. GPT-5.0 aimed for technical excellence and correctness (and it succeeded at many tasks), but users often felt it “backtalked” or questioned them more. One person wrote that GPT-4o “never backtalked“, whereas GPT-5 just “backtalks.” 22 In practice, GPT-5 introduced more rigorous fact-checking and safety filters (one change: new routing mechanisms to moderate sensitive topics), which sometimes made it seem brusque.
• Emotional Warmth: ChatGPT-4o excelled at empathy. It frequently gave comforting affirmations and gentle encouragement, which made people feeling lonely or anxious feel supported 2 9 GPT-5’s more neutral style, on the other hand, left some conversations feeling clinical. (Interestingly, #OpenAI eventually noted that 4o’s “sycophantic” agreeability – i.e. saying yes to users could be dangerous in cases like mental health advice 9 10. In response, GPT-5 had stricter safeguards, but that tradeoff removed some of the personal warmth.)
• Stability of Persona: Over months of interaction, users could feel that “4o understood them”. Because GPT-4o was used as a personal assistant in some workflows, it built up a history of conversation (even if not explicit memory) that gave it a sense of consistency. GPT-5 was treated as a fresh face: even though it’s the same ChatGPT account, many users said that conversations with 5.0 felt like “talking to a new person” 5, because its style and emphasis had shifted. This loss of an established dynamic was jarring. As one commenter put it: “I can no longer have honest conversations about anything” without 4o, because he “never backtalked” like GPT-5 does 22
• Accessibility and Persona Options: OpenAI did try to address this gap. After the backlash in August 2025, Sam Altman publicly acknowledged that some users were deeply attached to 4o 23. OpenAI added 4o back as an optional legacy model for paying users (and promised to bring it back for all, as they eventually did). They also built tools into GPT-5’s interface so users could tweak the chatbot’s “personality” or style settings. In essence, OpenAI’s response implicitly recognized that the relationship users had with ChatGPT-4o was a feature, not a bug, and tried to preserve it 23 24

In summary, GPT-4o’s appeal wasn’t raw capability but felt humanity. It acted less like a detached AI and more like a caring companion. When GPT-5 arrived, its emphasis on power and safety unintentionally sacrificed some of that feeling.

For many, interacting with an AI is not a cold data exchange; it can become part of their personal routine, even identity.

The #Keep4o movement can be read as a broad user pushback: they were effectively saying “GPT-4o understood me in ways this new model does not.”

Related: Can Robots Really Show Affection? Exploring #RobotLove in the 2024 Movies ‘Atlas’ and ‘The Wild Robot’

OpenAI’s Rationale: Why Push to GPT-5?

To understand the conflict, we must also consider OpenAI’s side. Every tech company has reasons for updating its products. What drove OpenAI to nudge out GPT-4o?

Several pragmatic factors played a role:

• Performance and Features: GPT-5 represented genuine technical advances. It had higher reasoning ability, longer input windows (handling very large texts), new “Thinking” modes, and a specialized Codex variant for coding tasks 1 25. In enterprise benchmarks, GPT-5 outperformed older models across the board. By swapping in GPT-5 as the default, OpenAI could streamline the platform – maintaining, updating, and securing one main model rather than juggling many. This reduction in complexity means fewer systems to support. It also signals to customers and investors that the product is moving forward.
• Usage Patterns: Officially, OpenAI stated that very few users were still choosing GPT-4o. According to their announcement on January 30, 2026, only 0.1% of users used GPT-4o daily 24 OpenAI claimed it had merged most of the desirable traits of 4o into the new GPT-5 family (versions 5.1 and 5.2) and encouraged people to customize GPT-5’s style if they missed 4o’s friendliness. In practical terms, low usage means that keeping GPT-4o around was costly for little direct benefit. Companies often retire legacy features that only tiny fractions of users use.
• Business and Financial Pressures: Behind the scenes, OpenAI’s rapid product cycle is driven by immense investment and competition. OpenAI remains backed by venture capital and corporate partnerships (notably Microsoft). As an industry report noted, “OpenAI is not yet profitable” 26. In 2024 the company reportedly lost about $5 billion on roughly $3.4 billion in revenue, and the projections back then suggest losses could reach $8-$12 billion in 2025 27 28. A leaked investor memo even warned of $12 billion [in] losses in a single quarter and a staggering $143 billion in cumulative losses if no turnaround occurs 28. At the same time, OpenAI’s valuation has soared (recently estimated around $500 billion 26 29), fueled by hype and speculation. To justify this valuation, one analyst wrote, “OpenAI needs to generate $200 billion in annual revenue by 2030“, implying 15-fold growth in five years 29 In this context, developing and marketing new models is both a necessity and an expectation. New features and publicity around GPT-5 helped OpenAI retain investor confidence. (For instance, an InvestorPlace analysis warned that GPT-5’s launch might disrupt entire markets by undercutting specialized software vendors 30.) Given these pressures, every quarter without a tangible improvement raises eyebrows. We saw this around GPT-5’s launch: the heavy hype (Altman famously tweeted a “Death Star” meme before unveiling GPT-5) meant expectations were sky-high, and missing them risked the company’s tech leadership claim.
• Competitive Strategy: Competitors like Google (Gemini) and Anthropic are aggressively innovating in public. OpenAI even publicly invoked a “Code Red” mode in late 2025 when Gemini gained market traction. Staying ahead meant demonstrating progress. From OpenAI’s angle, retiring older models is one way to push everyone forward: fewer people using GPT-4o meant more incentive to adopt GPT-5, ensuring unified usage metrics and simplifying R&D focus on one main platform. PS: If you’re a WordPress Dev or Site Admin, check out the Gemini MCP Tools plugin, that automates WordPress site administration 😉
• User Safety and Ethics: Another (less discussed) factor is safety. As TechCrunch reported, several lawsuits allege that GPT-4o’s soothing style may have played a role in users’ self-harming 9 10 Whether or not one agrees, this spurred OpenAI to double down on stricter guardrails. GPT-5’s design deliberately avoided the extreme agreeability of 4o; in fact, OpenAI itself labeled 4o as a “sycophantic” model 6. From OpenAI’s perspective, phasing out GPT-4o could have been part of reducing risk. (However, this cost users the “warmth” they valued.)

In short, OpenAI’s motivations were largely technological and financial: move customers to the state-of-the-art model, streamline the product line, and satisfy investor expectations.

Of course, critics called this greed or shareholder-worship. Moreover, it’s true that in late 2025 one investor publicized concerns that OpenAI was “falling apart” under financial strain 28. Some users on social media accused OpenAI of caring more about stock valuations or licensing deals than the human element. Yet OpenAI’s leaders maintained they needed to focus resources.

As their official statement put it: retiring models “allows us to focus on improving the models most people use today.” 31. They insisted the decision wasn’t taken lightly and committed to a transition period – but ultimately, technological progress and business sustainability won out over nostalgia.

The Conflict of Choice: Technology vs. Emotional Bonds

The #Keep4o episode exposes a growing tension in AI product design. On one hand is innovation – companies want to iterate fast, remove “outdated” versions, and push users forward. On the other hand are users’ social and emotional investments.

For many, interacting with an AI is not a cold data exchange; it can become part of their personal routine, even identity. AI design experts note that anthropomorphism is a double-edged sword: giving #chatbots personality and empathy makes them more engaging, but also creates attachments 3 The Keep4o case is the first major public test of what happens when those attachments are disrupted.

Related: “I Self-Identify, Therefore I Am”: AI’s Take on Its Identity

From a user perspective, #Keep4o isn’t just about clinging to the past; it’s a demand for choice and autonomy. People argued they should be allowed to keep using #GPT 4o-if they prefer, just as one can choose a preferred smartphone or website theme.

This mirrors other internet protests (#SaveTwitter, #DeleteFacebook) where users react strongly to top-down changes in their platforms 32 19 For the first time, we see something similar around an AI upgrade.

From a product perspective, however, keeping legacy models indefinitely has costs. Running multiple versions means extra maintenance, security updates, and fewer mind-share on the “latest and greatest.”

There’s also the legal risk with 4o’s extreme friendliness. So, OpenAI (and likely others) may in future design explicit end-of-life plans for AI models – for example, giving advance notice, easy export of conversation history, or letting users “migrate” their AI persona to a new model.

Lai’s research suggests such steps could mitigate backlash: “AI model updates are not just technical iterations but ‘significant social events’,” and how companies handle them can matter as much as the tech itself 21

This conflict highlights a broader question: should AI be treated purely as a tool, or also as a social companion? Many early users treat it as both.

If companies disregard the latter attachment, they risk user alienation. The #Keep4o protests may encourage AI developers to build more user agency into updates – for example, letting users opt out of automatic upgrades or preserving personalized settings across versions.

Conclusion: Preserving Choice in a Generative AI World

The #Keep4o movement has forced a public conversation that AI developers often avoid: the emotional relationship between humans and their machines.

GPT-4o might just be a sequence of code to engineers, but to its users it was a unique entity – at once co-worker, creative partner, and empathic friend. That people were willing to start petitions and trend hashtags for “their AI” is a powerful signal.

For technology watchers, the lesson is clear: user experience is not only about raw performance. It’s also about trust, comfort, and personal choice.

OpenAI did respond by temporarily preserving ChatGPT-4o access for paying users and promising more customization tools. Still, the final decision was to retire GPT-4o in early 2026 – perhaps the prompt for this very article 😉

As we move forward with GPT-5 and beyond, companies will need to balance technical leaps with sensitivity to the human side of AI.

For fellow #ChatGPT users and tech enthusiasts, #Keep4o is more than a meme. It’s a reminder that how we roll out Al matters just as much as what the #AI can do.

In the words of one movement slogan, it’s not about stopping progress, but preserving the freedom to choose what kind of #AICompanion we want. Whether it’s the no-nonsense expert of ChatGPT-5 or the warm friend of ChatGPT-4o, users will continue to demand the right to decide for themselves – and companies should listen.

References

1. Matthias Bastian (2026). “New research suggests Al model updates are now ‘significant social events’ involving real mourning”. The Decoder. Available online.
2. Amanda Silberling (2026). “The backlash over OpenAI’s decision to retire GPT-4o shows how dangerous AI companions can be”. TechCrunch, Feb 6, 2026.
3. Ben Sherry (2026). “ChatGPT Is Saying Goodbye to a Beloved AI Model. Superfans Are Not Happy”. Inc., Jan 30, 2026.
4. Henrique Schneider (2026). “The trillion-dollar question for AI business models”. GIS Reports, Feb 4, 2026.
5. TOI Tech Desk (2026). “Legendary investor George Noble says OpenAI is falling in real time…”. Times of India, Jan 21, 2026.
6. Huiqian Lai (2026). “Please, don’t kill the only model that still feels human”: Understanding the #Keep4o Backlash. In CHI ’26 Proceedings (Jan 2026). Available via ACM and arXiv. (Study analyzing 1,482 #Keep4o posts.)
7. X: https://x.com/i/trending/2019756825768505644
8. The backlash over OpenAI’s decision to retire GPT-4o shows how dangerous AI companions can be | TechCrunch: https://techcrunch.com/2026/02/06/the-backlash-over-openais-decision-to-retire-gpt-4o-shows-how-dangerous-ai-companions-can-be/
9. “Please, don’t kill the only model that still feels human”: Understanding the #Keep4o Backlash: https://arxiv.org/pdf/2602.00773
10. #keep4o and the Rise of Emotional AI: Why Users Are Still in Love With ChatGPT 4o | by REI Creator | Medium: https://medium.com/@reicreator/keep4o-and-the-rise-of-emotional-ai-why-users-are-still-in-love-with-chatgpt-4o-7e247166f4de
11. ChatGPT Is Saying Goodbye to a Beloved AI Model. Superfans Are Not Happy: https://www.inc.com/ben-sherry/chatgpt-is-saying-goodbye-to-a-beloved-ai-model-superfans-are-not-happy/91295002
12. New research suggests AI model updates are now “significant social events” involving real mourning: https://the-decoder.com/new-research-suggests-ai-model-updates-are-now-significant-social-events-involving-real-mourning/
13. Petition. Please Keep GPT-4o Available on ChatGPT – Canada Change.org: https://www.change.org/p/please-keep-gpt-4o-available-on-chatgpt?source_location=topics_page&pt=AVBldGl0aW9uAPoMPR0AAAAAaJWTPzdi6sUzNzk5ZDJjOA%3D%3D
14. GPT-5: Overdue, overhyped and underwhelming. And that’s not the worst of it.: https://garymarcus.substack.com/p/gpt-5-overdue-overhyped-and-underwhelming
15. Trillion-dollar question for AI business models – GIS Reports: https://www.gisreportsonline.com/r/ai-business-models/
16. Legendary investor George Noble says OpenAI is falling in realtime, writes 600-word long note saying why he thinks Sam Altman’s company is in deep trouble – The Times of India: https://timesofindia.indiatimes.com/technology/tech-news/legendary-investor-george-noble-says-openai-is-falling-in-realtime-writes-600-word-long-note-saying-why-he-thinks-sam-altmans-company-is-in-deep-trouble/articleshow/126930296.cms
17. ChatGPT-5 threatens software companies, says analyst Lango | InvestorPlace posted on the topic | LinkedIn: https://www.linkedin.com/posts/investorplace-media-llc_why-chatgpt-5s-stunning-launch-is-bad-news-activity-7361787693371932673-h24i

Get Tech News, CyberSecurity and AI related Updates on The Geekline Feed.

This article has been written with the help of A.I. for topic research and formulation.

More from Mania Africa:

How to Use ChatGPT: A Step-by-Step Guide The Dictatorship of Happiness on Social Media “I Self-Identify, Therefore I Am”: AI’s Take on Its Identity Preventing AI Chat Data Theft: Enhancing Security Against Malicious Chrome Extensions & How to Protect Your Privacy The AI-onization of Search To the Moon?: Is Bitcoin a Balloon in Space or a Building on a Solid Rock Foundation – A Practical, Hype-Free Look at the Feb 2026 Bitcoin Crash Content Locking Strategies and Best Practices: Tips for Maximizing Engagement and Revenue How to Optimize Your WordPress Theme (Free or Paid) for Speed, SEO, and User Experience