Artificial intelligence researcher Jacob Coxon has resigned from OpenAI and Anthropic, accusing leading technology companies of acting irresponsibly and gambling with human lives as they rush to build superintelligent systems.
Coxon announced his departure on Wednesday evening in a post on social media platform X, formerly known as Twitter. He revealed that he had spent the past three years working in artificial intelligence pre-training across both companies, warning that developers are accelerating toward self-improving technology without adequate safety controls.
The announcement, published around 9:00 p.m. Brasilia time on Wednesday, Sept. 8, generated massive attention online. By Thursday, the post had accumulated more than 64.9 million views and nearly 500,000 likes from users around the world.
Private fears and public statements
Coxon stated that the artificial intelligence industry is fully aware of the catastrophic risks advanced technology poses to humanity. He said developers inside major research laboratories sincerely believe that artificial intelligence could cause human extinction by the end of the decade, adding that such warnings are not marketing stunts.
According to Coxon, top executives and renowned researchers deliberately use sensible and measured language when speaking to news outlets and the public. Behind closed doors, however, he said those same leaders constantly express deep fear regarding the future trajectory of the technology.
He urged observers not to underestimate what artificial intelligence will soon be capable of doing. Coxon warned that superhuman systems will soon have the capability to hack any computer network, transform any industry overnight, and acquire real-world power and financial resources.
In computer science, superintelligence refers to hypothetical software systems whose cognitive abilities far surpass those of human experts across virtually all disciplines. Pre-training is the foundational stage of building these models, during which algorithms process vast amounts of text and data to learn linguistic patterns before undergoing specialized fine-tuning.
Calls for temporary bans and government oversight
To prevent potential disasters, Coxon argued against continuing the current competitive push toward advanced artificial intelligence. He proposed implementation of temporary bans on enhancing model capabilities, giving researchers necessary time to understand safety mechanisms and mitigate negative outcomes.
His proposal reflects growing apprehension across the technology sector regarding rapid deployment. In July, more than 1,200 technology industry employees signed a public petition named Pacing the Frontier, urging the United States government to create regulatory tools capable of managing the pace of advanced artificial intelligence development.
The Pacing the Frontier initiative includes workers from major AI developers OpenAI and Anthropic, as well as Meta, the parent company of Facebook, and Alphabet, the parent company of Google. The organization notes on its website that while artificial intelligence holds immense potential to improve society, positive outcomes are by no means guaranteed.
OpenAI, headquartered in San Francisco, California, is one of the world's prominent artificial intelligence laboratories and developer of ChatGPT. Anthropic, also based in San Francisco, was founded by former OpenAI research leaders with an explicit focus on building safe and alignment-focused artificial intelligence architectures.
Decisions inside private companies
Coxon strongly criticized technology executives for entering what he described as the final phase of an arrogant technology race. He stressed that decisions affecting the ultimate survival of humanity should never be made internally over Slack, the workplace messaging platform commonly used by private tech firms.
He closed his message with a direct appeal to laboratory researchers across the field. Coxon asked scientists to consider what the coming years will look like, questioning whether they truly want to launch superintelligent reinforcement learning projects without a rigorous understanding of artificial minds.
Reinforcement learning is a core machine learning technique where artificial intelligence models learn complex decision-making through trial and error, guided by digital reward signals. As tech firms attempt to combine reinforcement learning with large-scale pre-training to create autonomous agents, debate continues over whether researchers can effectively align superintelligent minds with human safety before they are deployed.
