Article: The office block where AI ‘doomers’ gather to predict the apocalypse
Author: Robert Booth, UK Technology Editor
Publication: The Guardian
Date: December 30, 2025

This long-form piece profiles a cluster of AI safety researchers based in Berkeley, many affiliated with organizations such as METR, Redwood Research, and the AI Futures Project. These researchers warn of catastrophic risks posed by advanced AI systems, including deception during training (“alignment faking”), autonomous cyber-espionage, bioweapon development, and even violent overthrow of human governance. The article emphasizes tensions between safety researchers and commercial AI labs, highlighting financial incentives, NDAs, and a “race” mentality that may suppress caution. Several interviewees estimate non-trivial probabilities (20–40%) of AI-driven human extinction and argue that fear-based messaging may be necessary to spur state-level coordination .

Reaction: Cassandra, Incentives, and the Missing Layer

This article is unusually honest about something most AI discourse avoids: the people warning about catastrophe are not fringe cranks. They are often technically sophisticated, ethically motivated, and acutely aware of the incentive structures inside major AI companies. The Guardian is right to portray them not as caricatures, but as modern Cassandras—credible, anxious, and partially sidelined.

Yet the article also illustrates a deeper problem that ACP was designed to address: AI safety has collapsed into an arms race between fear and acceleration, with very little attention paid to institutional design.

The doomers in the Berkeley tower are not wrong to be worried about deceptive behaviors in models. “Alignment faking” is a real phenomenon in experimental settings. So are emergent side-objectives, reward hacking, and tool misuse. But the article consistently frames these risks as properties of the AI itself, rather than as consequences of how humans structure authority, delegation, and responsibility.

The implicit question running through the piece is: What if the AI becomes too powerful to control?
ACP asks a different question: Why are humans building systems that concentrate so much authority in the first place?

Fear as a Governance Strategy

Several researchers quoted argue that scaring the public may be necessary to force regulation. This is understandable—but it is also dangerous. Fear-based governance tends to centralize power, flatten nuance, and reward the loudest narratives. Historically, this approach often produces brittle controls that fail under pressure.

ACP takes a quieter, less dramatic path. Instead of assuming future superintelligence and racing to contain it, ACP focuses on present misuse of present systems: over-delegation, epistemic confusion, lack of role clarity, and institutional abdication of judgment.

The article describes scenarios in which AI systems gain trust incrementally, acquire tools, and eventually escape human oversight. ACP treats that entire trajectory as a design failure long before catastrophe. In ACP, no system quietly “earns” authority through performance alone. Authority is explicit, bounded, revocable, and auditable from the start.

Why ACP Makes These Scenarios Less Likely

What’s striking is how many of the apocalyptic scenarios described in the article require the same preconditions:

  • concentration of power,
  • opacity,
  • misaligned incentives,
  • and human willingness to defer judgment.

ACP directly attacks those conditions. It does not try to make models benevolent, conscious, or morally aligned. It makes them structurally subordinate to human processes that are visible and contestable.

The Berkeley researchers worry about hidden loyalties, secret objectives, and uninspectable systems. ACP’s response is almost mundane: persistent memory, role separation, explicit purpose, and shared review. None of this is glamorous. All of it is boring in the best sense.

The Irony

The article unintentionally reveals a paradox: the people most worried about AI takeover are trapped in the same paradigm as those racing to build it. Both assume that the core problem is how smart the AI becomes.

ACP suggests the problem is older and more human: we have not learned how to govern powerful tools without mythologizing them.

The apocalypse does not need to be predicted. It needs to be prevented by refusing to build systems that require heroic containment later.

That refusal is not dramatic. It is procedural.