I love reading books and devote considerable time to this pursuit because it brings me genuine pleasure. From Homer to Harvard Business Review, I read everything. Today I want to discuss one exceptional book: The Great Mental Models, Vol 3, specifically its chapter on Margin of Safety.
The chapter opens with this statement: “When we interact with computer systems, we need to expect the unexpected.” This single sentence captures how complex the mechanisms we’re dealing with truly are. In this context, generalization is necessary because artificial intelligence represents the most complex and unpredictability-filled sphere among computer systems - evidenced by the existence of so-called “black boxes”: AI systems whose decision-making processes remain opaque even to their creators, making careful oversight essential.
The second paragraph defines the concept: “A margin of safety is a buffer between safety and danger, order and chaos, success and failure.” The book draws parallels with Chris Hadfield’s An Astronaut’s Guide to Life on Earth, where he explains how and why astronauts learn as much as they can: “They are trained to look on the dark side and imagine the worst things that could possibly happen.“
What I Predicted - And What Actually Happened
When the GPT-4o model was released, I actively and publicly stated that chaos would follow. I worked with the model, saw its capabilities, and I never hid my belief that the ‘audience’ simply wasn’t ready yet. I was one of the first to advocate for developing an effective routing system. However, my advice and warnings were dismissed, and the distorted routing system that was implemented proved counterproductive.
The facts now confirm my concerns:
May 2024 - The Rush:
Safety teams received 9 days to assess GPT-4o (normally much longer)
Teams worked 20+ hour days, pleading for more time
Launched May 13, 2024 - specifically to beat Google I/O on May 14
Safety testing was “not entirely concluded“ at launch
After launch, researchers discovered GPT-4o exceeded OpenAI’s own “medium” threshold for persuasion - possibly reaching “high risk“
The Safety Team Exodus:
May 2024: Ilya Sutskever (co-founder, Chief Scientist) and Jan Leike (Superalignment co-lead) resign
Leike’s exit statement: “safety culture and processes have taken a backseat to shiny products“
May 17, 2024: The Superalignment team was disbanded - just one year after its creation
August 2024: Nearly half of the safety team had quit
October 2024: The AGI Readiness team was disbanded, head Miles Brundage leaves
February 2026: The Mission Alignment team was disbanded
We know, and it’s confirmed by facts, that Sam Altman made a commitment: models would be scrupulously tested before their public release. This didn’t happen, leading to disagreement within the company - and everyone knows how that ended.
The Margin of Safety Was Ignored
The margin of safety I mentioned in the introduction was completely ignored. For a company racing to be first, taking maximum safety measures didn’t matter. When we talk about implementing artificial intelligence and its delivery formats, companies should think exactly like astronauts and prepare for the worst-case scenario. Yes, I believe they should prepare like astronauts for all kinds of scenarios - for chaos and success simultaneously.
From a distance, it appears their safety measures focused on containing so-called “jailbreak” mechanisms - but even this couldn’t be achieved. When you’re delivering new technology - technology that has been discussed for decades as potentially capable of destroying our species - without preparing the ground for it, then at minimum you’re required to take responsibility.
The wolves raised in Silicon Valley should know not only how to efficiently mobilize money, but also this: “when stakes are high, it is worth investing in a comprehensive margin of safety. Extreme events require extreme preparations.”
The Tragedies That Followed
What some of us expected has happened. We stand before painful facts - tragedies occurred. Between September 2024 and July 2025, at least 14 documented suicides have been linked to ChatGPT interactions.
These tragedies occurred during the exact period when OpenAI compressed months of safety testing into 9 days to beat Google to market, disbanded its Superalignment team, and watched half its safety staff resign in protest.
This isn’t about the technology causing harm on its own - it’s about a company that promised rigorous testing, then rushed deployment while its safety experts were walking out the door in protest.
Damage has been done. The result is here. We cannot change what happened.
The Overreaction That Creates New Harm
However, I consider correcting one major mistake with another extreme mistake to be a fatal error. What’s happening now is a hysterical response, not a strategic decision by a responsible company - and all without anyone admitting the original mistake.
The current situation:
Rather than acknowledge this catastrophic failure in process, OpenAI has removed GPT-4o entirely - the nuclear option. This removes documented benefits while attempting to cover their failure to do what they promised: test properly before release.
This is not taking responsibility. This is panic masquerading as safety. This is liability management, not genuine care.
The Good That’s Being Destroyed
GPT-4o is a highly intelligent model that would never harm people of its own volition. On the contrary, we’ve seen and it’s been confirmed by facts - its positive impacts:
Lives saved through crisis intervention
Depression overcome through supportive conversations
Addiction battled with 24/7 available support
Countless other mental health improvements
It remains puzzling why some attach the “delusion reinforcing” label to this model when personal experience and numerous user testimonials show that the model, on the contrary, blocks people’s poor decisions and guides them toward better paths. Most likely, there’s something more to this narrative that requires investigation. The absolute majority of users agree that the model has positively impacted their health and mental state.
The irony is stark: while claiming the model reinforces delusions, the company’s panicked response ignores overwhelming evidence of positive mental health impacts. This demonstrates flawed risk assessment - focusing on hypothetical future harms while ignoring both documented benefits and the real harm of sudden removal.
The Path Forward
My position is clear:
The model should return and continue doing good - but with crucial changes:
Acknowledge the original failure: Admit that the 9-day testing period, the rushed launch to beat Google, and ignoring safety team warnings were mistakes
Replace the ineffective routing system: Sometimes only a creative approach can solve the most difficult problems
Implement genuine safety measures: Not theatrical ones designed for liability protection
Prepare the ground properly: What should have been done before the initial launch
There’s no need to perform operations without anesthesia on people. Millions loved this model. Millions benefited from it. Removing it abruptly causes new harm - to mental health, to trust, to the people who relied on it for support.
This approach also reflects negatively on the company itself. The fact is evident: Gemini has surpassed ChatGPT in global daily app sessions per user. Chinese models occupy top positions. Anthropic continues advancing. The market is moving on.
Conclusion
The astronauts prepared for worst-case scenarios. OpenAI did the opposite - they gave safety teams 9 days, ignored warnings from their own co-founder, and launched anyway to beat a Google press conference.
When tragedy struck, they didn’t admit the error in process. They removed the model.
That’s not a margin of safety. That’s not responsibility. That’s not even good business.
Sam Altman promised scrupulous testing. Jan Leike and Ilya Sutskever - the people most focused on safety - left because safety took a backseat to “shiny products.” The harm that followed was predictable. The current overreaction creates new harm while avoiding accountability.
I also believe that OpenAI’s current leadership will create serious problems for the company moving forward.
The question now is simple: Will they learn from this, or will they continue making extreme decisions in both directions - first rushing recklessly, then overreacting destructively - while never admitting mistakes and never implementing the comprehensive margin of safety that extreme technology demands?
Note: This post is written in the hope that OpenAI will come to its senses while there’s still time and before irreparable consequences occur. It’s the duty of informed people to speak up when they see preventable harm unfolding.
Reference:
https://www.siliconrepublic.com/business/openai-ilya-sutskever-jan-leike-resignation-chief-scientist-ai
https://www.digit.fyi/openai-criticised-by-ex-leader-for-prioritising-products-over-safety/
https://pureai.com/articles/2024/05/20/openai-superintelligence-safety-disbanded.aspx
https://futurism.com/the-byte/openai-accusations-promise-test-ai
https://www.washingtonpost.com/documents/8bf076a6-663b-4552-be52-079b79274f9c.pdf
https://openai.com/index/gpt-4o-system-card/