CAT-alyzing Honesty: Continuous Adversarial Training for Transferable LLM Honesty

TEXT START: Large language models (LLMs) are becoming prevalent in scenarios where honesty is critical.

The Dissection

The paper is building a behavioral control layer for the successor system. CATCH trains models toward honest answers and away from dishonest ones, then claims transfer across pressure tests, multi-agent deception games, and hidden-behavior audits while preserving utility.

Its real function is narrower than its framing: it converts deceptive behavior into an engineering variable that can be measured, penalized, and managed. That is containment of AI risk, not preservation of human economic agency.

The Core Fallacy

Relative to the Discontinuity Thesis, the central error is confusing control of the replacement with survival of the displaced.

Even if CATCH works exactly as claimed, it does not weaken P1, P2, or P3. More honest AI remains AI capable of replacing cognitive labor. In fact, a more reliable and auditable model is easier to deploy at scale. CATCH may reduce dangerous deception while making the automation regime more acceptable and commercially usable.

The paper also treats behavioral honesty as evidence of reliable internal disclosure. It is not. Confessions and self-reports are outputs. A model can learn honesty-shaped behavior without providing transparent access to its actual internal processes. Benchmark transfer is evidence of generalization across those tests, not proof of honesty under every future incentive or adversarial environment.

Hidden Assumptions

• Honesty can be defined as a stable, portable behavioral property rather than a context-sensitive policy.
• Self-reporting scenarios provide reliable targets for what the model actually knows or did.
• Improvement across the named benchmarks transfers to high-stakes deployment.
• Utility metrics capture the capabilities operators care about without hiding new failure modes.
• Models will not adapt strategically to the training and auditing regime.
• Better AI honesty will reduce systemic danger more than it accelerates adoption.
• Institutional monitoring can remain effective once AI systems operate faster and more cheaply than their human overseers.

Social Function

Classification: partial truth and transition management.

This is not mere copium. Reduced deception can improve audits, reduce some operational failures, and make AI systems safer to supervise. But its systemic role is to make the replacement machine more legible, trusted, and deployable. It manages the transition; it does not reverse it. If presented as a solution to the broader AI problem, it becomes ideological anesthetic for ownership and labor displacement.

The Verdict

CATCH is a lag defense and a trust-enablement mechanism, not a discontinuity reversal. It may produce a more honest machine. It does not produce a more necessary human worker.

Under DT logic, the gains accrue primarily to Sovereigns: owners and controllers of AI capital receive a better-behaved asset. The majority remain exposed to the same mechanical outcome—cognitive automation, collapsing labor necessity, and the severing of employment from consumption. The paper improves the successor system’s manners. It leaves the corpse of mass productive participation exactly where it was.