Google DeepMind Predicts AGI by 2030, Details Safety Plans in 145-Page Report

Google DeepMind has released a sweeping 145-page technical report detailing its approach to artificial general intelligence (AGI) safety—a move signaling both its belief in the near-term plausibility of AGI and the significant risks it may pose.

The report, co-authored by DeepMind co-founder Shane Legg, forecasts the arrival of AGI by 2030 and urges the research community to take seriously the possibility of “severe harm,” including scenarios as extreme as “existential risks” that could threaten humanity.

“We anticipate the development of an Exceptional AGI before the end of the current decade,” the authors write.

Defining "Exceptional AGI"

DeepMind introduces a concrete benchmark: Exceptional AGI, defined as a system that performs in the 99th percentile of skilled adult humans across a wide range of non-physical cognitive tasks, including metacognitive functions like learning how to learn. This is not just a general goal—it’s the yardstick DeepMind proposes for tracking AGI development and risk.

A Multi-Layered Safety Framework

DeepMind proposes a three-pronged framework for mitigating AGI risk, spanning both technical and societal dimensions. The goal is to reduce the chances of catastrophic misuse while improving transparency and control:

Capability Control

• What it is: Techniques to restrict or limit what AGI systems can do—particularly in uncontrolled, real-world environments

• Examples: Sandboxing (confining AGI to test environments), capability throttling, and restricting access to certain tools or APIs

Motivation Control

• What it is: Ensuring AGI systems are motivated to behave in human-aligned ways

• Examples: Reward modeling, goal conditioning, and research into interpretability tools to better understand internal reasoning

Societal Resilience

• What it is: Building robust external systems and institutions that can contain, oversee, and recover from AGI-related incidents

• Examples: Third-party red teaming, global policy cooperation, detection of unsafe behavior, and systems designed to fail gracefully (i.e., ensuring AGI fails safely under pressure)

Key Concerns: Recursive AI and Cyber Threats

The report emphasizes the plausibility of recursive AI improvement—a scenario in which advanced AI conducts AI research to create increasingly powerful successors. DeepMind sees this as a potentially runaway risk that could lead to a technological feedback loop, outpacing human oversight.

In addition, the paper explores a range of cybersecurity threats, including:

• Autonomous exploitation of software vulnerabilities

• AI-written malware or phishing attacks

• Use of AGI for surveillance, influence operations, or infrastructure disruption

• DeepMind builds detailed threat models and calls for robust protections against misuse by state and non-state actors.

A Caution Against Overconfidence

The paper critiques both Anthropic and OpenAI in how they approach AGI safety:

• Anthropic is described as underemphasizing training, monitoring, and security for AGI, though this reflects DeepMind’s internal comparison rather than a comprehensive assessment. In broader AI safety circles, Anthropic is regarded as a leader in LLM safety, known for its Responsible Scaling Policy, AI Safety Levels framework, and its development of Constitutional AI for aligning models with human values.
  

• OpenAI is criticized for relying too heavily on automating alignment research, a method that uses AI systems to evaluate and train other models.

It also questions OpenAI’s framing of superintelligence, calling such claims premature without architectural breakthroughs. While DeepMind is skeptical of superintelligence emerging imminently, it believes Exceptional AGI is plausible by 2030, with a long tail of uncertainty extending into the late 2030s.

Transparency on Challenges

DeepMind is candid that many proposed techniques are in development and involve “open research problems.” Examples include:

• AI interpretability tools that decode what large language models are “thinking”

• Access control systems for limiting AGI availability

• Mechanisms for verifying agent behavior under adversarial conditions

Rather than presenting final answers, the report is framed as a research roadmap—inviting others to contribute to a broad ecosystem of AGI safety development.

Skeptics Push Back

Several AI experts expressed reservations about the paper’s core premises:

• Heidy Khlaaf (AI Now Institute) argued AGI remains too ill-defined to be scientifically rigorous.
  

• Matthew Guzdial (University of Alberta) questioned recursive improvement, calling it “theoretical with no real-world evidence”.
  

• Sandra Wachter (Oxford) warned of present-day harms, especially AI systems learning from their own flawed outputs—leading to a growing cycle of misinformation masquerading as truth.

  “At this point, chatbots are predominantly used for search and truth-finding purposes. That means we are constantly at risk of being fed mistruths and believing them because they are presented in very convincing ways,” Wachter said.

What This Means: Setting the Stage for AGI Governance

DeepMind’s paper is not just a technical document—it’s a signal. It asserts that AGI may emerge within the decade and urges the AI community to prepare accordingly, through a mix of technical safeguards, regulatory cooperation, and transparent planning.

• For labs and developers, it calls for a broadening of safety research beyond alignment alone, into capability limitation, deployment protocols, and adversarial resilience

• For policymakers, it emphasizes the need for global standards and cooperative enforcement

• For the public, it acknowledges that incredible benefits and catastrophic risks could coexist in this next phase of AI—and that ignoring either is irresponsible

Despite its depth, the report won’t end the debate over AGI timelines or priorities. But it provides perhaps the clearest, most comprehensive look yet at how one of the world's most advanced AI labs is thinking about AGI—and what might come next.

Editor’s Note: This article was created by Alicia Shapiro, CMO of AiNews.com, with writing, image, and idea-generation support from ChatGPT, an AI assistant. However, the final perspective and editorial choices are solely Alicia Shapiro’s. Special thanks to ChatGPT for assistance with research and editorial support in crafting this article.