The Researchers Who Walked Away: Every Major AI Lab Exodus Since 2018
|

The Researchers Who Walked Away: Every Major AI Lab Exodus Since 2018

BySmallbizai.au

In February 2026, Mrinank Sharma packed up his desk at Anthropic and flew to the UK to write poetry. He had spent years working on AI safety. He left because, in his words, “the world is in peril.” A month later, xAI lost its last co-founder besides Elon Musk. By April, the company had shed roughly 80 people.

These are not isolated stories. Since 2018, more than 110 named researchers and technical leads have left the frontier AI labs: Google DeepMind, OpenAI, Anthropic, Meta AI, xAI, Mistral. Some walked out on principle. Some were pushed. Some left to build something new. A few left because they had decided the whole enterprise might be beyond saving.

What follows is a full account of who went, when, and why, organised by the four waves that shaped the field.

Wave 1: The Google Ethics Firings (2019–2022)

The first major break happened not at a startup but at the world’s biggest AI operation. Google’s internal AI ethics program was, for a brief period, one of the more credible attempts to put guardrails on the technology from the inside. By mid-2022 it had been effectively dismantled, through a combination of firings and departures that became international news.

Meredith Whittaker helped organise Google’s 2018 worker walkout over sexual harassment. She ran the AI Now Institute from inside the company. In July 2019 she left, later becoming president of Signal. Her departure came before the ethics team was formally assembled; she saw early what was coming.

Timnit Gebru, who co-led the Ethical AI team, was fired in December 2020 after a dispute over a research paper on the risks of large language models. The paper “Stochastic Parrots” had pointed out that scaling up models without understanding them carried real dangers. Google’s response was to fire her. She founded DAIR, the Distributed AI Research Institute, shortly after.

Margaret Mitchell, the other co-lead of the Ethical AI team, was fired in February 2021. She joined Hugging Face as chief ethics scientist. Samy Bengio, a Google Brain research manager of 14 years, resigned in April 2021. He said later it was in solidarity. He went to Apple.

Alex Hanna followed in February 2022, joining Gebru’s DAIR. Blake Lemoine, a LaMDA engineer who had claimed the model was sentient, was fired in July 2022. His claims were widely disputed, but his departure underscored how uncomfortable Google had become with internal dissent of any kind.

Running parallel to the ethics implosion, Mustafa Suleyman, co-founder of DeepMind, left in January 2022. He went on to co-found Inflection AI and later became CEO of Microsoft AI.

Then there was the quiet exodus from the Transformer paper. All eight authors of “Attention Is All You Need”, the 2017 paper that underpins essentially every major language model in existence, eventually left Google. Aidan Gomez founded Cohere. Ashish Vaswani and Niki Parmar went to Adept. Noam Shazeer left for Character.AI, returned briefly, then left again for OpenAI in 2026. Llion Jones and David Ha co-founded Sakana AI in Tokyo. The people who built the engine drove away in it.

Geoffrey Hinton, the Turing Award winner known as the godfather of deep learning, resigned in May 2023. His stated reason: he wanted to speak freely about the dangers of AI without embarrassing Google. He has spent the time since doing exactly that.

Wave 2: The OpenAI Safety Exodus (2020–2024)

OpenAI’s departures cluster around two moments: 2020-21, when a group of senior researchers left to found Anthropic, and 2024, when the remaining safety structure effectively collapsed.

The founding of Anthropic was the most consequential split. In late 2020 and early 2021, Dario and Daniela Amodei left alongside GPT-3 lead Tom Brown, interpretability researcher Chris Olah, scaling theorist Jared Kaplan, and several others. Their concern was that OpenAI was not taking safety seriously enough. They built a rival lab to prove it could be done differently.

Paul Christiano, who had led OpenAI’s alignment team, left in 2021 to run the Alignment Research Center. He later became chief scientist at the US AI Safety Institute.

The 2024 wave was sharper and more public. William Saunders resigned from the Superalignment team in February 2024, saying he had lost trust in how safety decisions were being made. Leopold Aschenbrenner was fired in April 2024. He later published a long document called “Situational Awareness” arguing that AGI timelines were far shorter than most acknowledged. Daniel Kokotajlo, a governance researcher, quit at the same time and surrendered his vested equity rather than sign a non-disclosure agreement. He wanted to be able to talk.

Gretchen Krueger, a policy researcher, resigned in May 2024, citing transparency concerns. Then, within 24 hours, the two most senior safety figures departed.

Ilya Sutskever, co-founder and chief scientist, who had led the board’s brief attempt to fire Sam Altman in November 2023, announced his departure on 14 May 2024. He went on to found Safe Superintelligence. Jan Leike, who had co-led Superalignment, resigned the next day. His public statement was blunt: “safety culture has taken a backseat to shiny products.”

The departures continued through the rest of the year. John Schulman, a co-founder and the lead on reinforcement learning from human feedback, left in August 2024 for Anthropic, then moved to Mira Murati’s new lab five months later. Mira Murati, OpenAI’s CTO, resigned in September 2024, taking several key researchers with her to found Thinking Machines Lab. Chief research officer Bob McGrew left the same day.

Suchir Balaji, who had worked on GPT-4 training data and later became a whistleblower in the New York Times copyright lawsuit against OpenAI, quit in August 2024. He died in November 2024. Miles Brundage, who had spent six years advising on AGI readiness, left in October 2024. He told TechCrunch: “no lab is ready for AGI.” Steven Adler, a safety researcher who left in late 2024, said publicly in January 2025 that he was “pretty terrified” by the pace.

Wave 3: The xAI Co-Founder Collapse (January–March 2026)

What happened at xAI in the first quarter of 2026 was structurally unlike anything before it: an entire founding team dissolved in under three months.

xAI had been founded in 2023 with 12 co-founders alongside Elon Musk. By 28 March 2026, all 11 of the others were gone.

Greg Yang, a mathematician and co-founder, left in January 2026. He cited a Lyme disease diagnosis. Tony Wu, the reasoning lead, left 9-10 February. Jimmy Ba, co-inventor of the Adam optimiser that underpins most neural network training, followed the next day. Vahid Kazemi departed around the same period, saying all AI labs were “building the exact same thing, and it’s boring.”

Through February, a further wave left: Simon Zhai, Hang Gao, Radhakrishnan Venkataramani, Shayan Salehian, Ayush Jaiswal, Andrew Ma, Rahul Ravishankar. Toby Pohlen‘s last day was 26 February. Zihang Dai and Guodong Zhang followed in March. Manuel Kroiss and Ross Nordeen, the last two co-founders besides Musk, were out by the end of March.

Fast Company reported roughly 80 total departures by April 2026. Business Insider published a full list. The picture that emerged was of a lab where the founding researchers and the controlling owner had reached a point where coexistence was not possible.

Wave 4: The Conscience Quits (2026)

The 2026 departures that attracted the most attention were not the largest in number. They were the ones accompanied by explicit public statements that the labs were dangerous.

David Silver, who had led reinforcement learning at DeepMind for 15 years and was the architect of AlphaGo and AlphaZero, left in January 2026 to found Ineffable Intelligence in London.

Caitlin Kalinowski, who led OpenAI’s robotics and consumer hardware work, resigned in March 2026 over the company’s Pentagon contract.

Mrinank Sharma left Anthropic in February 2026. His public post was direct: “the world is in peril.” He moved to the UK to write poetry. BBC covered his departure.

Alex Turner, a DeepMind researcher, resigned in July 2026 over Google’s military AI contract. Bilal Chughtai, a DeepMind AGI safety researcher, published a statement on 14 September 2026: AI could “kill us all.” Business Insider covered his departure.

Robert O’Callahan, a DeepMind engineer based in Auckland, New Zealand, resigned in September 2026. He told the Times of India he was not convinced the chance of catastrophic AI outcomes was low enough to justify the pace of development.

Jacob Coxon, who had worked at both OpenAI and Anthropic, quit Anthropic on 8 September 2026. He told CNBC and the BBC that both labs were “gambling with our lives.” We covered what he said at length, including what it means for Australian small businesses, in a companion piece on SmallBizAI.

David Robinson, who had spent three and a half years on OpenAI’s safety systems, resigned in early October 2026. He told TechCrunch the culture was “broken.” Business Insider ran his interview with Ezra Klein.

What the Patterns Show

Reading across all four waves, a few things stand out.

Safety exits cluster around product launches. Jan Leike’s “shiny products” complaint was not abstract. The 2024 OpenAI departures coincided with the rollout of GPT-4o and Sora. The 2026 conscience quits followed a period of rapid capability announcements. The researchers who left were not opposed to the technology. Most stayed in AI. They were opposed to the speed.

The field recycles its own talent. Anthropic was founded by OpenAI safety researchers who wanted to go slower. Those researchers have now watched colleagues leave Anthropic for similar reasons. John Schulman went from OpenAI to Anthropic to Thinking Machines in under 18 months. The Transformer authors scattered across a dozen companies that now compete directly. Jan Leike, who left OpenAI citing safety concerns, joined Anthropic. John Jumper, who won the Nobel for AlphaFold at DeepMind, left for Anthropic in June 2026. The talent moves, but the concerns travel with it.

Not all departures are voluntary. Meta cut roughly 600 AI-unit staff in October 2025. The Google Ethical AI team was not dissolved through attrition. Gebru and Mitchell were fired. Lemoine was fired. The xAI departures likely involved a mix of chosen exits and structural pressure. The narrative of principled resignation, while real in many cases, obscures a messier picture.

The people who built the foundational tools are now scattered. The eight authors of the Transformer paper, which underpins ChatGPT, Gemini, Claude, and every other major model, are spread across Cohere, OpenAI, Character.AI, Sakana AI, and Inceptive. The people who built GPT-3 (now at Anthropic and elsewhere), AlphaGo (Ineffable Intelligence), AlphaFold (Anthropic), and RLHF (Thinking Machines) have all moved on. The labs that trained them are competing against companies those researchers founded with the knowledge they took with them.

Whether that is healthy competition or a structural risk depends on who you ask. The researchers who walked away in 2026 mostly think it is the latter.

📬 The SmallBizAI Brief

One practical AI tip for Australian small business. Every Tuesday. Free.

Join business owners getting smarter about AI every week.

Subscribe free →

Sources

📖 More deep dives: News Deep Dives

📖 More deep dives: News Deep Dives →

Similar Posts

Leave a Reply

Your email address will not be published. Required fields are marked *