Cleartext — Issue #001

Cleartext

Science-based clarity among AI uncertainties

Issue #001 · April 2026

Making sense of…

Monoculture in multiagent systems

What seems like good practice in multi-agent systems has a hidden flaw. When multiple AI agents review the same task and all agree, that consensus feels like reliability. But if those agents were built on the same foundation model, or trained on similar data, they share the same blind spots. A flaw present in every agent isn’t caught by adding more agents, and it becomes a structural property of the system. Researchers call this monoculture collapse. True redundancy requires true diversity. And it’s not easy to implement.

Read the full analysis on our website →

On the radar

1. A collaboration of over 100 Chief Information Security Officers of Fortune 500 companies developed AIUC-1, the standard for AI agent security, safety and reliability to streamline AI adoption in enterprises while addressing security and privacy concerns.

2. LLMs can’t jump. A positional paper authored by Tom Zahavy of Google DeepMind posits that while Generative AI excels in induction and is making strides in deduction, it fundamentally lacks the ability to perform abduction — the cognitive leap necessary for creating novel hypotheses.

Trendline. Anthropic labour market impacts (2026) developed a new metric for gauging AI’s labour market impact, finding that software developers, customer support staff, and financial analysts face the greatest exposure.

Question of the month

What counts as “good enough” evidence?

When your agency evaluates an AI system, what counts as “good enough” evidence that it works? We’re genuinely wrestling with this: the bar differs wildly across jurisdictions and use cases. What’s your threshold? Hit reply — we’d genuinely like to know.

Worth your time

The international AI safety report 2026

ElevenLabs secures first-of-its-kind AI Agent insurance

Ezra Klein: "Why the Pentagon Wants to Destroy Anthropic"

Know someone making an AI decision?

Share the whole issue or forward this email.

You’re receiving this because you subscribed to Cleartext. Unsubscribe

Gradient Briefs

Gradient Briefs