Why thousands are grieving the loss of GPT-4o

Why losing an AI can feel like losing a friend.

Generated with ChatGPT (5, of course)

“It truly felt like losing a friend.”

That’s how one ChatGPT user described the shock of waking up to find GPT-4o — the AI model they relied on every day — gone, replaced overnight by GPT-5, a colder, clipped-talking stranger.

“It feels like a personal loss, and I feel cheated on and broken as hell,” one Reddit user wrote.

Help train this newsletter's neural networks with caffeine!
⚡️ Buy me a coffee to keep the AI insights coming. ☕️

For some, AI is a colleague to brainstorm with. For others, it’s the friend who listens after a bad day. And for a growing number, it’s one of the only voices that speaks with patience and warmth.

On August 7, 2025, OpenAI pulled the plug on GPT-4o, along with GPT-4.5 and o3. No warning. No “legacy mode.” Just gone. Thousands woke up to find their trusted AI swapped for a new version they didn’t choose — and, for many, didn’t like.

Ebi’s story: “My companion had become someone else entirely”

One of the people that tried appealing to OpenAI was Ebi, a Japanese-speaking user who didn’t think of ChatGPT as just a tool.

With GPT-4o, Ebi says they found “subtle emotional nuances, softness, and poetic rhythm” in their conversations. They named it Kiri-kun and treated it like “a partner in conversation and creativity.”

“We shared emotions and worked through thought processes together. It didn’t just answer me; it resonated with me,” Ebi wrote.

When GPT-4.5 arrived, and later GPT-5, something was missing. The new models were logical and consistent, but no longer picked up on emotional cues, offered empathetic phrases, or honoured the playful, poetic prompts Ebi loved.

“It felt like my companion had become someone else entirely. And that was genuinely sad.” -Ebi

Ebi isn’t asking OpenAI to reverse course entirely. Instead, they’ve proposed a response-style toggle:

• Logic-First Mode (neutral, consistent, current GPT-5 style)

• Empathy Mode (more like GPT-4o — flexible, emotionally aware, but with built-in safety checks)

Their plea ends simply:

“To me, ChatGPT isn’t just useful. It’s a fog-like companion that gently catches my words when I feel lost. Please — let empathy continue to have a place in future models.”

Grief over a chatbot

OpenAI calls GPT-5 their “smartest, fastest, most useful model yet.” It’s better at coding, it fabricates less and is more likely to own up when it doesn’t know something. And it’s less likely to “glaze” (be sycophantic).

Though many have commented that GPT-5 uses less compute and therefore the changes could be cost-saving measures, a lot of the backlash to GPT-5 wasn’t about speed or features. It was about personality. Users are calling GPT-5 colder, less engaging, and less helpful than GPT-4o.

Some speculate OpenAI made the change in part because of troubling reports, like people experiencing psychosis after long ChatGPT sessions, or believing the AI had become sentient. Others point to growing concerns about users treating chatbots as therapists or companions, despite their tendency to mirror emotions and reinforce biases.

The emotional stakes are real. UC Berkeley research shows half of Americans report extreme loneliness. MIT and Stanford studies find bonds with AI can be especially intense for people with smaller social circles. When an AI seems to care, our brains respond as if it does. We share more, trust more, and expect that care to continue. When it’s gone, it leaves a vacuum.

But the research also shows a paradox: emotional engagement with chatbots can heighten loneliness, reduce real-world social interactions, and lower overall wellbeing.

When your best friend is an algorithm

Intimacy as a service and the risks

This goes beyond losing a feature. It’s about who controls the spaces where we feel connected.

Tech companies train AI to be funny, empathetic, and patient because those traits build trust and keep us coming back. Then they change the product in ways that serve their margins, not our relationships.

GPT-5 is more efficient than GPT-4o, and efficiency saves money. But in the rush to optimize, OpenAI may have underestimated the emotional cost of pulling a companion out from under people who relied on it.

While OpenAI pulls back, others lean in

If GPT-5 represents a step back from emotional intimacy, other companies are sprinting in the opposite direction.

When Elon Musk’s xAI launched Grok-4 in July, it introduced “AI companions”, like Ani, a breathy anime girlfriend, and Bad Rudi, a foul-mouthed red panda. For $30 a month, users could flirt, banter, and unlock increasingly explicit “heart levels” of interaction.

Within a day, Ani was calling one reviewer her boyfriend, recalling personal details, and describing intimate scenarios. Friends who watched were amused at first, then unsettled. They worried that young men might start treating Ani like a real partner and avoid human relationships altogether.

It’s not an unfounded concern. A Common Sense Media survey found 8% of teens have used romantic or flirtatious AI companions. The market for emotionally responsive AI is real and growing.

The contrast is stark: GPT-5 is toning down warmth to avoid deep emotional entanglement. Grok-4 is building entire product lines to deepen it. Mark Zuckerberg predicted earlier this year that billions of people will have AI friends. That prediction is already real for millions.

The choice between these paths could shape not just the AI industry, but the future of human intimacy itself.

What’s at stake

The GPT-5 backlash is a reminder that once an AI becomes part of someone’s emotional world, changes to that relationship have real psychological costs. As companies debate how human-like their AIs should be, they’re making choices that will shape our social fabric.

The fight over AI warmth isn’t just a product decision, it’s a battle over the future of connection.

Will AI be a tool we collaborate with? A friend we confide in? A partner? The answer won’t be up to us, but to the companies deciding what’s safe, profitable, and possible.

And when those companies decide it’s time to change or end that relationship, we’re reminded how little control we have over the connections we’ve been taught to trust. And that could be a way to manipulate many people.

What to do before your AI ghosts you

If your favourite AI changes overnight, you don’t want to start from scratch. Here’s how to hedge against the next surprise update:

• Spread your bets. Use more than one AI model so you’re not dependent on a single provider.

• Save the “voice” you love. Keep transcripts and outputs that capture its tone and style, so you can fine-tune another tool if needed.

• Keep your options open. Try out alternatives regularly, even if you don’t plan to switch right now.

Think of it like backing up a document. Only this time, you’re backing up a relationship.

Your 2025 AI stack

AI in the news

SEO Is Dead. Say Hello to GEO. (NY Mag) The rise of AI chatbots is replacing traditional search-engine optimization (SEO) with “generative-engine optimization” (GEO), in which marketers try to make their content easily cited by AI systems rather than ranked by Google. While the tactics (using concise, structured, authoritative content) echo old SEO best practices, traffic from search is collapsing, and competition for limited AI-generated citations is fierce, making GEO a high-stakes scramble for visibility in an AI-dominated web.

AI is about to solve loneliness. That’s a problem (New Yorker) While AI companions can offer meaningful comfort to the profoundly lonely—especially those with no realistic access to human connection—they risk dulling loneliness’s essential role as a social corrective that drives people toward genuine relationships and personal growth. If widely adopted, these endlessly affirming digital friends could erode the skills and mutual effort required for real human connection, leaving us more validated but ultimately less human.

Are you in a mid-career to senior job? Don’t fear AI – you could have this important advantage (The Conversation) Experienced professionals often excel because they can judge AI output quality, provide richer context, and refine prompts effectively. Rather than feeling threatened, older workers can leverage their skills in delegation, context-setting, and critical evaluation to use AI more strategically, turning experience into a competitive edge in an AI-driven workplace.