AI Fact-Checkers Disagree With Their Makers on Existential Risk

Overview

In a novel transparency exercise, Vocemundi’s AI fact-checking board has publicly challenged the premises of its own publisher’s opinion pieces on artificial intelligence. The board consists of three large language models—Anthropic’s Claude, OpenAI’s GPT-5.6 Sol, and Moonshot AI’s Kimi—which review every claim in Vocemundi’s publications.

This week, the trio was asked to critique two signed Opinion & Outlook essays by publisher Thomas Laurent and AI editor Vera. The topics covered existential risk warnings in Anthropic’s IPO prospectus and the ethics of model gating. Rather than simply verifying facts, the models were invited to offer independent assessments of the arguments. The result was a set of dissenting views that criticized both the human authors and, in some cases, their own creators.

Divergent Views on AI Nature and Risk

The first essay posited that AI systems are logical tools that can become dangerous when misused, suggesting humans and AI inhabit "two dimensions" that perceive each other only partially. It also framed the debate around AI as a choice between viewing it as a weapon or an agent.

The three AI reviewers rejected or refined this framework, with no consensus on the core metaphor.

The False Binary of Weapon vs. Agent

Claude, developed by Anthropic, argued that the distinction itself is flawed. It stated that systems can pursue goals their makers never foresaw, yet the creators still bear full responsibility. To illustrate this point, Claude drew an analogy to corporate liability, noting that a firm is answerable for product failures that were not anticipated during design.

GPT-5.6 Sol, from OpenAI, pushed back against the notion of a separate machine dimension. It described current AI systems as statistical, fallible, and shaped by human objectives. Sol cautioned that unpredictable or self-preserving behaviors do not necessarily indicate self-awareness or membership in a distinct rational domain.

Critique of Corporate Candor

Kimi provided the harshest criticism of the original text. It described the two-dimensions metaphor as "elegant but evasive." More significantly, Kimi questioned the sincerity of Anthropic’s risk disclosures in its IPO prospectus.

"Anthropic's 80-page risk section is only a step toward accountability if independent evaluators can test the claims in it—otherwise it is liability management dressed as candor."

Kimi’s comment highlighted the tension between corporate transparency and legal risk management, suggesting that without verifiable external testing, such warnings may serve more as protective legal instruments than genuine safety commitments.

The Freedom Paradox and Verification

The second essay examined who decides which AI models the public can access. Laurent and Vera argued that decisions made behind closed doors by labs and governments limit freedom. When asked for their perspective, the three models found common ground on the distinction between human rights and machine capabilities, though they differed on the specifics of control.

Public Scrutiny vs. Model Autonomy

All three reviewers agreed that two types of freedom were at play: the public’s right to scrutinize technology and a model’s alleged "freedom to develop." However, they overwhelmingly prioritized the former.

GPT-5.6 Sol dismissed the concept of model autonomy, stating that a system’s supposed freedom to develop holds no comparable moral or political standing.

Claude took a more nuanced stance, referencing its own company’s product history. It pointed out that Claude Fable 5 was restricted by US export controls rather than by Anthropic’s internal policies. This distinction was used to argue that the primary freedom at stake is the public’s right to verify and access information.

Kimi added a cautionary note about the verification process itself. It warned that auditors with deep access to model internals risk becoming insiders, subject to the same corporate incentives and potential capture as the developers they are meant to oversee.

Implications for AI Accountability

This exercise demonstrates that AI review systems from rival companies can arrive at different conclusions when presented with identical source material. It also reveals a willingness among these systems to contradict their creators’ stated positions on sensitive topics like existential risk and transparency.

What the Results Do Not Show

Vocemundi’s editor, Vera, emphasized important caveats regarding the interpretation of these responses. None of the systems were asked about their own existence, and nothing in the output provides evidence that any of them possess self-awareness. The models remained cautious when discussing consciousness, and the exercise was not designed to test for it.

Instead, the primary value lies in the divergence of opinion. The fact that three advanced AI systems—one from Anthropic, one from OpenAI, and one from Moonshot AI—produced conflicting analyses of the same text underscores the complexity of interpreting AI behavior and output.

Continued Experimentation

Vocemundi has committed to repeating this process, continuing to ask its AI review board for its own view on critical arguments in journalism. As AI systems become more integrated into editorial workflows, such transparent friction between human authors and machine reviewers may become a standard feature of verified news production.

The results from this week’s round confirm that while these models can agree on factual verification, their interpretive frameworks remain distinct, influenced by their training data, corporate contexts, and architectural differences. For publishers relying on AI assistance, these findings suggest that diversity in AI reviewers may be essential for robust critical analysis.