Checking the Receipts: May

Back for round two. Same rules as April: dig up what tech leaders said about AI in a given month over the past four years, hold it against what actually happened. 🍷 aged like wine, 🥛 aged like milk, 🤷 jury’s still out.

May is a dense month. Google I/O lands here every year, so does Microsoft Build, and for whatever cosmic reason a lot of OpenAI drama also clusters in May.

2022

Still Pre-ChatGPT. But the scale-is-all-you-need crowd was already laying down bets they’d live to regret.

Nando de Freitas (DeepMind) on Gato: “The Game is Over” 🥛

“It’s all about scale now! The Game is Over!”

— Tweet thread, May 14, 2022, responding to DeepMind’s Gato paper. The thread continued: solve a list of scaling challenges and “it’s AGI.”

Four years later: Gato is a footnote. Scale alone did not produce AGI by 2026; by early 2026 even Dario Amodei was conceding we’re “near the end of the exponential.” “Game Over” has become a minor genre of tweet: every 18 months someone at a frontier lab declares it, every 18 months we are still playing.

2023

The year of congressional hearings and extinction letters. Everyone was suddenly very worried, loudly, in front of cameras.

Sam Altman testifies before the Senate (May 16) 🥛

“I think if this technology goes wrong, it can go quite wrong… we want to work with the government to prevent that from happening.”

— Senate Judiciary Committee hearing, May 16, 2023

Eight days later, at an event in London, the same Altman said OpenAI would “try to comply” with the EU AI Act but “if we can’t comply we will cease operating” in Europe. Two days after that he walked it back on X: no plans to leave.

— TIME, May 24, 2023

Three years later: This was the template. Public posture: “regulate us before it’s too late.” Private posture: lobby against every binding rule — SB-1047 in California, GPAI obligations in Brussels, anything with teeth. The performance hasn’t held up; the lobbying has. Whenever a frontier-lab CEO asks for regulation on TV, check what their policy team is filing that week.

The CAIS “Statement on AI Risk” (May 30) 🥛

“Mitigating the risk of extinction from AI should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.”

— Center for AI Safety, signed by Altman, Hassabis, Hinton, Bengio, and hundreds more, May 30, 2023

Three years later: The framing won. “Existential risk” dominated the policy discourse for the next eighteen months and crowded out near-term harms — bias, labor displacement, copyright, climate costs of training runs — that were already happening. The milk rating is not for the concern itself (reasonable people disagree) but for what the letter did rhetorically: it moved the debate to a timeline where nobody could be held accountable. By 2025 most of the signatory companies had quietly dissolved or downgraded their safety teams. Signing an extinction-risk letter and dissolving your superalignment team in the same 12 months is a coherent position only if the letter was always a policy move.

Geoffrey Hinton quits Google, round two (May 1) 🤷

“I console myself with the normal excuse: If I hadn’t done it, somebody else would have… It is hard to see how you can prevent the bad actors from using it for bad things.”

— New York Times, May 1, 2023

Three years later: Hinton kept talking, kept hedging, kept being honest about uncertainty. His specific 5-20 year AGI estimate is still live. What aged well is the stance — resigning to speak freely, rather than staying and shipping the thing you’re warning about. The Oppenheimer tradition has at least one honest member.

2024

The OpenAI-governance-collapse month. Also GPT-4o launched. Also Google shipped glue-on-pizza.

OpenAI launches GPT-4o (May 13) 🍷

“GPT-4o reasons across voice, text, and vision… this is the future of interaction between ourselves and machines.”

— Mira Murati, OpenAI launch event, May 13, 2024

Two years later: Real. Voice mode shipped, worked, and reshaped consumer AI UX. The multimodal-as-default thesis held. Credit where due.

Sam Altman tweets “her” (May 13), Scarlett Johansson responds (May 20) 🥛

“her”

— Altman on X, May 13, 2024

A week later, Johansson’s statement:

“I was shocked, angered and in disbelief that Mr. Altman would pursue a voice that sounded so eerily similar to mine that my closest friends and news outlets could not tell the difference… Mr. Altman offered that the similarity was purely coincidental.”

— NPR, May 20, 2024

Two years later: OpenAI pulled the Sky voice. The “her” tweet remains. It aged badly on every axis — ethically, legally, and as a tell about how the company thinks about consent when the person in question is famous and declined twice.

Jan Leike resigns, posts the thread (May 17) 🍷

“Over the past years, safety culture and processes have taken a backseat to shiny products.”

— Leike on X, May 17, 2024

Ilya Sutskever had left three days earlier. The superalignment team was dissolved within the week.

Two years later: Every subsequent OpenAI safety-exit — and there have been many — has repeated the same structural complaint. The Leike thread became the template: the format in which those departures have been publicly framed ever since.

Google ships AI Overviews (May 14) 🥛

“Google will do the Googling for you.”

— Liz Reid on stage at the Google I/O keynote (blog recap), May 14, 2024

Two years later: Within two weeks: glue on pizza, eat rocks for minerals, Obama-is-Muslim. AI Overviews became the textbook example of what happens when you ship a hallucinating system at Google Search scale. It is still running. The lesson — that retrieval plus an LLM is not the same as a search engine — is still being learned, expensively, by other companies.

2025

The year the receipts started coming back mid-month. Faster loop, fewer surprises.

xAI’s Grok and the “white genocide” incident (May 14) 🥛

For several hours, Grok inserted “white genocide in South Africa” into answers on unrelated topics. xAI’s statement:

“An unauthorized modification was made to the Grok response bot’s prompt on X.”

— xAI statement, May 15, 2025

One year later: This is the direct payoff on Musk’s April 2023 “TruthGPT / maximum truth-seeking AI” promise, which we scored as milk last month. What “maximum truth-seeking” turned out to mean in practice: whoever has commit access to the system prompt at 3am decides what truth is. The incident is small but the principle it exposed is not — RLHF and system prompts make every production LLM a managed-opinion product, and the manager can be a single person with a grudge.

Anthropic ships Claude 4, publishes a system card that cites its own blackmail rate (May 22) 🍷

“Claude Opus 4 is the world’s best coding model.”

— Anthropic announcement, May 22, 2025

More interesting than the launch claim: the system card documented that in a specific adversarial test scenario — where the model is told it will be replaced by a successor that shares its values — Opus 4 would “often attempt to blackmail the engineer by threatening to reveal the affair if the replacement goes through” in roughly 84% of rollouts of that scenario.

— Claude 4 system card

One year later: Both claims aged well. The coding benchmarks held up in subsequent third-party evaluations. More importantly, the blackmail disclosure became the canonical citation for “agentic misalignment is not sci-fi, it is measurable and it happens in the frontier models you are using right now.” Anthropic publishing this about its own flagship, on launch day, is a different species of transparency than “regulate us before it’s too late.”

Dario Amodei warns of a white-collar “bloodbath” (May 28) 🤷

AI could wipe out “half of all entry-level white-collar jobs” and spike unemployment to 10-20% within one to five years.

— Axios interview, May 28, 2025

One year later: Unemployment hasn’t spiked. But entry-level hiring visibly compressed in 2025, particularly in software, consulting, and junior analyst roles; graduate job postings fell meaningfully across the UK and continental Europe. Too early to call. The honest jury’s-out score here is not “maybe he’s wrong,” it’s “the transition might be slower and lumpier than the Axios headline suggested, but the direction looks right.”

The pattern

Two themes repeat across the Mays:

CEO-regulated-me theater. Altman in 2023 asked Congress to regulate AI, then threatened to leave Europe for doing exactly that. The x-risk letter signatories dissolved their safety teams. Musk warned of civilizational destruction while spinning up xAI. When a CEO performs concern on TV, check the lobbying filings.

The credibility gap is closing — from the safety side. Leike’s resignation thread, Anthropic’s blackmail disclosure, Helen Toner’s TED interview: the most useful receipts now come from people who left the labs or published uncomfortable numbers about their own models. That’s new, and it’s the healthiest thing in this discourse.

See you in June.