Artificial Guarantees Episode III: Revenge of the Truth
Google DeepMind enters the chat
Today we’re updating our ongoing project, Artificial Guarantees, which is a collection of inconsistent statements, baseline-shifting tactics, and promises broken by major AI companies and their leaders showing that what they say doesn't always match what they do.
We’ve been continuing to collect examples of this type of behavior, and below have a new selection of interesting cases to share with you. You can get our full list at controlai.com/artificial-guarantees
Do you have an example that we missed? Join our Discord to add to the collection.
Google
What they say:
Google commits to "publish reports for all new significant model public releases" when models are "more powerful than the current industry frontier".
What they do:
Publicly deploy their “most intelligent AI model” without a model card. Model cards are supposed to “provide essential information on Gemini models, including known limitations, mitigation approaches, and safety performance”.
July 2023 - Google signed the 2023 White House's Voluntary Commitments for Safe, Secure, and Trustworthy AI, where they clearly promised to "publish reports for all new significant model public releases" when models are "more powerful than the current industry frontier".
25 March 2025 - Google DeepMind launches Gemini 2.5 Pro, “our most intelligent AI model”.
3 April 2025 - TechCrunch notes that Gemini 2.5 Pro was publicly deployed without a model card, noting the commitments they signed in July 2023.
Their excuse?
[Google’s director and head of product for Gemini, Tulsee Doshi] told TechCrunch that the company hasn’t published a model card for Gemini 2.5 Pro because it considers the model to be an “experimental” release. The goal of these experimental releases is to put an AI model out in a limited way, get feedback, and iterate on the model ahead of a production launch, she said.
9 April 2025 - Fortune highlights this as an apparent breach of their safety commitments.
17 April 2025 - Google finally publishes a model card.
What they say:
Google’s AI won’t be used for military purposes
What they do:
Drop their ban on AI for weapons and surveillance
2014 - When Google acquired DeepMind in 2014 and CEO Demis Hassabis joined Google, he got Google to pledge that DeepMind’s AI would never be used for military or weapons purposes.
2018 - This was formalized in Google’s “AI Principles” in 2018, where they pledged not to develop AI for weapons or to “cause overall harm”.
February 2025 - Google updated its “AI Principles” to remove its commitment not to use AI for weapons or surveillance.
A Google spokesperson when asked about this, referred to a recent blog post coauthored by Demis Hassabis.
OpenAI
What they say:
OpenAI will require fine-tuned versions of their models to be safety tested.
What they do:
OpenAI quietly alters their Preparedness Framework to restrict this only to cases where the model weights are released.
December 2023 - OpenAI publishes its Preparedness Framework, which describes OpenAI’s processes to “track, evaluate, forecast, and protect against catastrophic risks posed by increasingly powerful models”.
The document specifies that fine-tuned versions of their models need to be safety tested:
April 2025 - OpenAI quietly reduce their safety commitments, changing the Preparedness Framework to no longer require safety testing of finetuned versions of their models, and omitting this from the Preparedness Framework change list.
They now only suggest that this would be done if the model weights are to be released.
What they say:
OpenAI launches a “superalignment” team, pledging 20% of their compute to be dedicated to this team’s efforts over the next four years
What they do:
OpenAI dissolves the superalignment team less than a year later.
July 2023 - OpenAI launches “Superalignment” a team co-led by Chief Scientist Ilya Sutskever, and Jan Leike. The goal of the team was to solve superintelligence alignment, ensuring that “AI systems much smarter than humans follow human intent”. They also pledged to “dedicat[e] 20% of the compute we’ve secured to date to this effort”.
May 2024 - OpenAI dissolves its superalignment team in the aftermath of superalignment co-leads Sutskever and Leike announcing their departures from OpenAI.
What they say:
OpenAI commits to "publish reports for all new significant model public releases" when models are "more powerful than the current industry frontier".
What they do:
OpenAI launches GPT-4.1 with no plans for a system card.
July 2023 - OpenAI signed the 2023 White House's Voluntary Commitments for Safe, Secure, and Trustworthy AI, where they clearly promised to "publish reports for all new significant model public releases" when models are "more powerful than the current industry frontier".
15 April 2025 - OpenAI launched a new family of AI models, GPT-4.1, with no plans for a system card. A system card includes important information about safety performance and mitigations.
OpenAI spokesperson Shaokyi Amdo said they didn’t need to release a system card, because GPT-4.1 isn’t a frontier model. The model does however, make advances on efficiency and latency, which policy analyst Thomas Woodside told TechCrunch makes having this safety report all the more important.
If you find the latest developments in AI concerning and the latest steps towards better AI security exciting then you should let your elected representatives know!
We have tools that make it super quick and easy to contact your lawmakers. It takes less than a minute to do so.
If you live in the US, you can use our tool to contact to your senator here: https://controlai.com/take-action/usa
If you live in the UK, you can use our tool to contact your MP here: https://controlai.com/take-action/uk
Join 1000+ citizens who have already taken action!
You can read the first two articles in this series here:
and here:
I keep flashing back to a cartoon I saw decades ago. It showed young Dr. Frankenstein writing "I shall not play in God's domain" repeatedly on a classroom chalkboard . . . and here we are . . .
Having grown up with the late Isaac Asimov's "Robot" stories, read the 'Colossus" books and seen the films, and experienced the irredeemably stupid AI applications currently in use, I can say that AI should never be released on the public in any form. It is a recipe for failure at all levels and all applications.