Google DeepMind

Lab Cope Score

HEAVY COPE

15/100

Data Transparency

No displacement metrics published

30/100

Framing Honesty

Productivity multiplier narrative

AGI Levels Defined

Level 1 Emerging → Level 5 Superhuman

Source

“Organizing Intelligence” — Research primer, May 2026, by Martin Gonzalez (Head of Organizational AI Research, Google DeepMind), foreword by Simon Bouton (Chief Experience Officer). Co-produced with Stanford HAI via the AI for Organizations Grand Challenge, launched late 2025.

Key Projects & Capabilities

Level 5

AlphaFold — Protein Structure

Superhuman narrow AI. Predicts protein structures with accuracy that took structural biologists decades to approach. Classified as Level 5 on DeepMind’s own AGI scale. The benchmark for what narrow superintelligence looks like in practice.

Level 5

AlphaZero — Strategy

Mastered chess, shogi, and go through pure self-play — then taught grandmasters entirely new strategies they had never conceived. Not just superhuman performance: genuine knowledge generation that flowed back to humans.

50 Years

WeatherNext — Forecasting Progress

15-day probabilistic weather forecasting. DeepMind claims WeatherNext represents 50 years of progress in meteorological science — delivered in a single model. Displaces decades of numerical weather prediction research.

Frontier LLM

Gemini

Google DeepMind’s frontier large language model. Positioned against GPT-4 and Claude across reasoning, multimodal, and agentic benchmarks.

Video

Veo — Video Generation

Text-to-video generation model. Competes with Sora and other video generation systems. No displacement analysis of video production workers published.

Embodied

Gemini Robotics 1.5

Generalist robots with advanced embodied reasoning and motion transfer. Moves learned in one physical context transfer to new tasks and environments. The physical-world equivalent of zero-shot generalisation.

Sim

Concordia — Organisational Simulation

Generative agent-based modelling platform for organisational simulation. Lets organisations model how AI agents interact within complex systems before deploying them. The displacement implications are left as an exercise for the reader.

Dialogue

AI Interpretability Research

Building a shared machine-human vocabulary. Future interpretability framed as interactive dialogue, not dashboards. Understand what the model “thinks” rather than just what it outputs.

The AGI Framework: 5-Level Scale

Heavy Cope Alert

DeepMind has defined a 5-level AGI scale — and positioned two of their own systems (AlphaFold, AlphaZero) at Level 5. As of early 2026, there is no consensus that Level 2 competent general AGI has been achieved. The framework creates the illusion of a managed transition while avoiding the displacement question entirely.

Level	Label	Description	Examples
1	Emerging	Early capabilities, narrow task performance	Early language models
2	Competent	Broad task competence approaching human level	No consensus achieved (early 2026)
3	Expert	Human expert-level performance across domains	—
4	Exceptional	Above human expert level in most domains	—
5	Superhuman	Far exceeds best human performance; generates new knowledge	AlphaFold (proteins), AlphaZero (strategy)

Note: Levels 1–5 apply to narrow AI systems. The same scale applied to general intelligence remains contested. AlphaFold and AlphaZero achieving Level 5 narrow AI does not imply Level 5 general AI is imminent — but DeepMind’s framing invites that conflation.

Democratic Deliberation & AI Mediation

A DeepMind AI outperformed human mediators at finding common ground on polarised political opinions.

This finding is buried in the organisational research primer rather than framed as a displacement of professional mediators, facilitators, and conflict-resolution specialists. Democratic deliberation is knowledge work. DeepMind just demonstrated their AI can do it better than humans. The word “displacement” does not appear in the surrounding analysis.

The Cope: “Extraordinary Science” Framing

The Kuhn Manoeuvre

The primer frames the current moment as a period of “extraordinary science” (invoking Kuhn) — where old rules break down and ideas scatter. This is accurate. The cope is positioning DeepMind at the research frontier of this rupture as a good thing for organisations without measuring what “old rules break down” means for the humans whose jobs were those rules.

Classic Deflection Patterns

Productivity Multiplier Narrative

Every AI capability is framed as multiplying human productivity rather than substituting human labour. AlphaFold doesn’t displace structural biologists — it “accelerates drug discovery.” WeatherNext doesn’t displace meteorologists — it “improves forecast accuracy.” The multiplier frame renders displacement invisible by design.

No Equivalent of OpenAI Signals or Anthropic Labour Research

DeepMind has published zero occupational exposure analysis, zero O*NET task mapping, zero employment effect studies. The absence is conspicuous. They produce systems (AlphaFold, WeatherNext, Concordia) that demonstrably replace expert human work — but publish no data on what that means for the humans doing that work.

Mission Statement Cope

“Build AI responsibly to benefit humanity, including organisations across every sector.” Benefit is undefined. Humanity is untracked. The mission statement substitutes for the labour market analysis that would actually tell you whether the benefit is real or redistributive.

Demis Hassabis

“AI will have probably ten times the impact the industrial revolution had, ten times faster.” The industrial revolution created mass unemployment before it created mass prosperity — over a century. Ten times faster means the disruption phase is proportionally shorter and sharper. DeepMind’s own CEO has quantified the displacement risk. Their research programme has not measured it.

DeepMind vs OpenAI vs Anthropic: Cope Comparison

Dimension	OpenAI	Anthropic	Google DeepMind
Labour data published?	Yes (Signals)	Yes (March 2026)	No
Displacement acknowledged?	Implicit	Minimised	Not measured
Occupational mapping	165 O*NET IWA codes	O*NET categories	None
Level 5 systems	None claimed	None claimed	AlphaFold, AlphaZero
Cope strategy	Rename displacement as adoption	Measure lagging indicator	Produce no displacement data
Cope score	72 • HEAVY	58 • MODERATE	78 • HEAVY