Usage Policy

February 2024 structural turn: Google updated its AI Principles and deleted the four-category negative list of “AI applications we will not pursue” from the 2018 version — including the explicit ban on “weapons or other technologies whose principal purpose is to cause injury to people.” This is the first structural retreat by a tech giant from a public refusal of military AI since Project Maven, and the central political event of this page.

Overview: Google’s “policy cluster” rather than a single AUP

Unlike Anthropic’s single Usage Policy or OpenAI’s Usage Policies, Google’s constraints on use are distributed across at least four layers of documents:

Google AI Principles (2018 / 2024 revision) — a company-level statement of principles, not directly user-facing
Generative AI Prohibited Use Policy — the consumer-facing “prohibited use” list for Gemini App / AI Studio
Gemini API / Vertex AI Additional Terms of Service — enterprise and developer layer
Responsible AI Practices / Responsibility & Safety Hub — research and engineering guidelines (deepmind.google/about/responsibility-safety)

This layered policy ecosystem reflects the natural complexity of Alphabet as an organisation, but it also creates space for “policy arbitrage” — for example, Vertex AI enterprise customers can negotiate into modes that are unavailable on the consumer side (such as contractual carveouts for specific monitoring scenarios).

Layer 1: AI Principles (2018 → 2024)

The seven principles of the 2018 original

On 7 June 2018, then-CEO Sundar Pichai published the AI Principles following the collective employee protests over Project Maven (a Pentagon drone-vision AI contract):

Be socially beneficial
Avoid creating or reinforcing unfair bias
Be built and tested for safety
Be accountable to people
Incorporate privacy design principles
Uphold high standards of scientific excellence
Be made available for uses that accord with these principles

Even more important was the “Applications we will not pursue” four-category list:

Technologies that cause or are likely to cause overall harm
Weapons or other technologies whose principal purpose or implementation is to cause or directly facilitate injury to people
Technologies that gather or use information for surveillance violating internationally accepted norms
Technologies whose purpose contravenes widely accepted principles of international law and human rights

Historical positioning: this made Google the first tech giant to publicly refuse military AI. Meredith Whittaker, Lilly Irani, and other former Google employees later treated the Maven protests as a canonical case in “worker governance.”

The February 2024 revision: deletion of the negative list

On 4 February 2024 Google quietly updated the AI Principles page, deleting the entire “Applications we will not pursue” section and replacing it with three broad principles: Bold Innovation, Responsible Development and Deployment, and Collaborative Progress. Bloomberg, The Washington Post, and The Guardian all read this as laying the groundwork for Project Nimbus (the multi-year Israeli defence AI cloud services contract awarded in 2021) and other military collaborations.

Key changes after revision:

Dimension	2018 version	2024 version
Structure	7 positive + 4 negative	3 positive (no negative list)
Weapons clause	Express prohibition	Deleted
Surveillance clause	Express prohibition (absent lawful authority)	Specific phrasing deleted
Accountability language	”accountable to people"	"with appropriate human oversight”

Google’s official explanation: “Enumerated negative lists are too rigid to adapt to the rapidly evolving application contexts of AI.” Critics — including hundreds of Google employees who signed open letters and former DeepMind ethics researchers — read this as a paradigmatic case of “principles giving way to contracts.”

Layer 2: Generative AI Prohibited Use Policy

The consumer-layer Prohibited Use Policy (policies.google.com/terms/generative-ai/use-policy) is the AUP actually binding end users. It has been revised multiple times since 2024; the current (Q1 2026) version’s core prohibited categories:

Dangerous or illegal activity: weapons / explosives manufacturing, terrorism, human trafficking, CSAM generation
Misleading and deceptive use: impersonation of real individuals, ghostwriting for academic fraud, financial-fraud materials
Interference with elections and democratic processes: since 2024 Gemini actively refuses to answer election-related factual questions (“I can’t help with responses on elections and political figures right now”), widely criticised as over-censorship
Critical-infrastructure attacks: unauthorised access or disruption of power grids, water, finance, healthcare
“Material harm to user wellbeing”: mass psychological manipulation, addictive design, self-harm encouragement
Non-consensual content: non-consensual realistic imagery, involuntary deepfakes
Bypassing Google safety systems: jailbreaks and prompt injection used to circumvent content moderation

The “nesting doll” phenomenon in policy: the Gemini App refuses election queries, while Google Search’s AI Overviews generates content referencing political figures — two policies under one parent company. Lazar and others criticise this “policy boundary by product rather than by capability” as “ethics by surface area.”

Layer 3: Gemini API / Vertex AI enterprise terms

The Vertex AI Additional Terms (cloud.google.com/terms/service-terms) carry over Google Cloud’s habitual “enterprise-compliance partner” posture:

Customer bears compliance responsibility for inputs / outputs (Section 5 Customer Responsibility)
Abuse Monitoring can be selectively disabled for HIPAA / FedRAMP High scenarios (analogous to OpenAI Enterprise zero-retention)
Enterprise carve-outs: Vertex AI customers can apply for contract-level waivers of certain Prohibited Use clauses — for example, “lawful government use” carveouts for military / intelligence customers (Google’s public materials do not elaborate, but The Intercept reported in 2024, through Project Nimbus documents, that such clauses exist)
SynthID automatic watermarking is enabled by default on Gemini outputs; enterprise customers may apply for customisation or deactivation

Layer 4: Responsibility & Safety Hub

deepmind.google/about/responsibility-safety is maintained by DeepMind rather than Google Legal and constitutes a research-layer document, including:

Responsible AI Toolkit (developer-facing safety APIs)
Model Behavior Policies (Gemini’s refusal / harmful-content logic)
SynthID / Provenance research
AGI Safety Research (from the Shane Legg / Rohin Shah teams)

This layer has no legal binding force but shapes model behaviour in practice — maintained jointly by DeepMind’s Safety Team (London) and Google’s Responsible AI group (Mountain View).

Academic critique

Claire Lazar (ANU, Oxford) — principles fatigue

In her 2024 AI Principles and Industry Practice (Oxford GovAI Working Paper), Lazar argues that Google’s AI Principles gradually weakened their “use-restriction” clauses during 2018–2024, confirming the prediction in Mittelstadt (2019) “Principles Alone Cannot Guarantee Ethical AI”. The 2024 deletion of the negative list is the “climax” of this process, not a sudden break.

Legacy of the Timnit Gebru episode (December 2020)

Gebru’s departure over On the Dangers of Stochastic Parrots (which Google characterised as a resignation) prompted the structural disintegration of the Google Ethical AI team — Margaret Mitchell was dismissed the following year, and several core members left. This is structurally continuous with the 2024 AI Principles revision: when internal ethics-review capacity is weakened, checks on policy revision are also weakened.

Meredith Whittaker (Signal, ex-Google) — ethics washing

In The Steep Cost of Capture (2023), Whittaker criticises AI-ethics documents from Google / Microsoft / OpenAI as a “corporate capture of ethics language” — replacing enforceable constraints with principled language. The 2024 AI Principles revision was read by her as “the clearest single act of self-exposure.”

Binns & Veale (UCL) — from “principles” to “procedure”

“The enforceability of AI use policy depends not on clause content but on the design of internal compliance procedures.” — Binns & Veale, “Is That Your Final Decision? Multi-Stage Profiling” (2021)

Their critique: Google lacks Anthropic’s publicly visible Acceptable Use decision-making process (Anthropic began publishing case explanations for some AUP decisions in 2024). Google’s internal review (“AIPR” — AI Principles Review) is opaque and unaudited by external parties.

Industry first-hand observations

Organisational tension Google ↔ DeepMind

In April 2023 Google announced the merger of Google Brain and DeepMind into Google DeepMind, with Demis Hassabis as CEO. But usage-policy authority still spans two organisations:

Product-layer policy (Gemini App / Vertex AI Prohibited Use) led by Google Legal + Trust & Safety
Research-layer policy (FSF, Model Behavior Policies) led by the DeepMind Safety / Responsibility & Safety Council

Practical consequence: the conservative positions DeepMind researchers push (e.g. the new Harmful Manipulation CCL in FSF v3) and the looser stances product teams push (e.g. AI Overviews relaxing limits on generated content) have publicly clashed. The 2024 “Gemini historical images” incident (generation of racially displaced historical figures) is a paradigmatic case of cross-organisational coordination failure.

Policy mapping onto YouTube / Workspace

“Fragmentation” of Google usage policy:

Gemini in Workspace (Docs / Gmail sidebar): governed by the Workspace Additional Terms
Gemini in YouTube (Auto-dub, Shorts script assistance): governed by YouTube Community Guidelines + Creator Music restrictions
AI Overviews in Search: governed by Search Policies, not Gemini Prohibited Use
NotebookLM: stand-alone Terms, allows user document uploads without training use

This product-level fragmentation makes “what is Google’s AI usage policy?” a genuinely hard question at the technical-compliance level. The EU AI Act Article 50 requirement for transparency labelling must be implemented separately across at least six products within Google’s ecosystem.

Comparison with peers

Dimension	Google (Q1 2026)	Anthropic AUP	OpenAI Usage Policies
Document unity	Fragmented across 4+ layers	Single AUP + commercial terms	Single Usage Policies + Enterprise
Weapons prohibition	Deleted in 2024	Express prohibition	Deleted in January 2024; partially restored October 2024
Election scenarios	Refuse to answer (excessive)	Permits assistance with real content	Limited in 2024
Enterprise carve-outs	Strong (Vertex / government)	Present (Commercial Terms)	Present (Enterprise)
Transparency	Policy text public, no decision explanations	Public AUP decision cases	Low public disclosure

Key timeline

June 2018: initial AI Principles released (after Maven protests)
December 2020: Gebru episode; Ethical AI team destabilisation begins
2021: Project Nimbus contract signed
April 2023: Google DeepMind merger
February 2024: AI Principles revised; negative list deleted (the central event on this page)
February 2024: Gemini historical-images incident; employee petition against Nimbus
May 2024: FSF v1 released (see safety-framework)
February 2025: FSF v2 + minor Prohibited Use revision
November 2025: Gemini 3 Pro released, accompanied by the first FSF Report
April 2026: FSF v3 released, adding the Harmful Manipulation CCL

Cross-references

Corporate overview and FSF panorama: Google DeepMind index
Model-level capability disclosure: Gemini Model Cards
Safety Framework (FSF) evolution: Safety Framework
Comparisons: Anthropic AUP, OpenAI Usage Policies
Methodology: company attribution