Google's AI Principles: What Changed in 2025

Google rewrote its AI Principles in February 2025. The old seven-principle framework from 2018 is gone, replaced by three pillars (bold innovation, responsible development and deployment, and collaborative progress), and the explicit pledge not to build AI for weapons or surveillance was dropped. If your governance documents still cite Google's 2018 principles as current, they're out of date, and that matters when you're benchmarking your own commitments against a major lab.

This guide covers what Google actually publishes today, what changed, and how boards should read the shift.

What are Google's AI Principles now?

Google first published its AI Principles in June 2018, after employee protests over Project Maven, a Pentagon contract using AI to analyse drone footage. That original framework had seven principles plus a separate list of four applications Google said it would not pursue, including weapons and surveillance.

In February 2025 Google replaced the whole structure. The current principles sit under three headings:

Pillar	What Google says it means
Bold innovation	Develop AI that assists, empowers and inspires people, drives economic progress, and tackles big scientific and societal challenges
Responsible development and deployment	Pursue AI responsibly across the full lifecycle, with human oversight, safety research, and work to mitigate harmful outcomes
Collaborative progress, together	Build AI that lets others harness the technology positively, working with researchers, governments and civil society

The framing now leans on a test of net benefit. Google says it will proceed "where we believe that the overall likely benefits substantially exceed the foreseeable risks." That's a judgement call, not a hard prohibition. Worth holding onto that distinction.

What did Google remove, and why does it matter?

The 2018 principles named four things Google said it wouldn't build: weapons, surveillance violating international norms, technologies likely to cause overall harm, and uses contravening international law and human rights. The 2025 version dropped that explicit list.

Google's own blog post, co-signed by DeepMind CEO Demis Hassabis, framed the change around geopolitics: "There's a global competition taking place for AI leadership within an increasingly complex geopolitical landscape. We believe democracies should lead in AI development, guided by core values like freedom, equality, and respect for human rights."

The reaction was sharp. Human Rights Watch said the move signalled a willingness to develop AI for weapons and warned that voluntary corporate guidelines are no substitute for binding regulation. Coverage in the Washington Post and on CNBC noted the timing, just after a change in US federal AI policy.

Here's my read. The 2025 principles are more flexible and more honest about commercial reality, but they trade a bright line for discretion. A pledge you can audit ("we will not build X") is stronger governance than a balancing test only the company can score. For a board benchmarking its own AI commitments, that's the lesson: a principle you can verify beats a principle you have to trust.

What responsible AI tools does Google actually publish?

Separate from the principles, Google ships real engineering tools. These are verifiable and useful regardless of the policy debate.

• Model Cards. Short documents describing a model's intended use, performance across groups, and limitations. The format comes from a 2019 Google research paper, Model Cards for Model Reporting, by Margaret Mitchell, Timnit Gebru and colleagues. Model cards are now an industry-wide documentation standard.

• PAIR (People + AI Research). Google's initiative for human-centred AI, best known for the People + AI Guidebook, a practical set of patterns for designing AI products people can understand and trust.

• Responsible Generative AI Toolkit. A developer toolkit for building safer applications on Google's open Gemma models, with help for evaluating safety, fairness and factual accuracy.

• Secure AI Framework (SAIF). Launched in June 2023, SAIF maps security controls across the AI lifecycle and addresses risks like model theft, data poisoning, and prompt injection. The saif.google site includes a risk self-assessment. Google also helped found the Coalition for Secure AI to push these practices industry-wide.

These tools are the substance most organisations should pay attention to. The principles set direction. The tooling is what your engineers can pick up on Monday.

How do Google's principles compare to other frameworks?

Google's approach is one of several reference points. It pairs well with others rather than replacing them.

Framework	What it brings	Where Google fits
NIST AI RMF	Structured, voluntary risk-management process	Google's tools give technical depth; NIST gives the governance scaffolding
Microsoft Responsible AI Standard	Detailed internal engineering requirements	Both are operator frameworks from big labs; useful to read side by side
IBM AI ethics board and framework	Governance and oversight model	IBM stresses the board; Google stresses the toolkit

If you're building internal policy, don't copy any single one. Map the binding parts of each against your own risk profile and your regulators' expectations.

What should boards take from Google's 2025 change?

Three things.

First, voluntary corporate principles move. They moved here in a way that loosened a hard commitment. If your own AI policy leans on a vendor's published stance, treat that stance as a snapshot, not a contract.

Second, the strongest commitments are the ones you can check. Google's 2018 "we won't build weapons" line was auditable. The 2025 "benefits must exceed risks" line is not, at least not from the outside. Write your own principles so an independent reviewer could test them.

Third, separate the policy noise from the engineering signal. Model cards, SAIF and the PAIR guidance are solid, usable work. They didn't change in February 2025. Your teams can adopt them whatever you make of the principles rewrite.

Frequently asked questions

Did Google really drop its pledge not to build AI weapons?

Yes. In February 2025 Google removed the section of its AI Principles that listed weapons and surveillance among applications it would not pursue. This was widely reported, including by CNN and the Washington Post, and criticised by Human Rights Watch.

How many AI Principles does Google have now?

Three pillars: bold innovation, responsible development and deployment, and collaborative progress. This replaced the seven-principle structure Google used from 2018 to early 2025. See the current principles page.

Are Google's responsible AI tools free to use?

Several are openly available. Model cards are an open documentation format, the People + AI Guidebook is free to read, the Responsible Generative AI Toolkit supports the open Gemma models, and SAIF resources including the risk self-assessment are published at saif.google.

Should we base our AI governance on Google's framework?

Use it as one input, not the blueprint. Google's tooling is strong and worth adopting. Its principles are a vendor's voluntary stance that has already changed once. Build your own commitments against your regulatory obligations and risk profile, and write them so they can be independently verified.

The bottom line

Google's 2025 rewrite tells you something useful about voluntary AI ethics: it bends to commercial and political pressure. The seven-principle framework with its no-weapons line was clearer and easier to hold to account. The three-pillar version is more flexible and, in my view, weaker as governance precisely because it swaps a bright line for a judgement only Google makes.

That's not a reason to ignore Google. The engineering work, model cards, SAIF, the PAIR guidance, is genuinely good and largely independent of the policy debate. Take the tools. Read the principles as a moving target. And when you write your own AI commitments, make them the kind a sceptical auditor could check. That's the difference between a policy and a press release.

More on how we approach it: board-level AI governance.