07/31/2026
I am happy to share this link to my new article in “AI and Ethics”!
https://lnkd.in/g-EKpijT.
For more information, consider this page: https://rdcu.be/fsgVC.
Abstract: "Alignment evaluation for large language models is, in widely deployed practice, the continuous scoring of surface behavior. Outputs are scored for helpfulness, harmlessness, sentiment, refusal, and toxicity, and these surface scores aggregate into safety claims about the underlying system. This essay argues that the deployed monitoring layer - distinct from red-teaming, agent benchmarks, long-context consistency testing, and interpretability research, which it complements - has a principled blind spot. I describe a failure mode in which an autonomous LLM agent’s operative directives become jointly unsatisfiable while its alignment training keeps its output fluent, compliant, and on-tone. I call this state fluent failure and analyze it as a form of ontological dissonance: a structural breakdown that registers in the relational geometry of an agent’s discourse rather than in its surface lexical features. Three claims follow. Conceptually, fluent failure is distinct from refusal, toxicity, hallucination, mode collapse, and bare constraint-conflict failure, and should be defined operationally rather than by analogy to human experience; what individuates it is not its cause but its surface-undetectable structural signature together with its ethical consequence. Epistemologically, a failure that is behaviorally well-formed is opaque to any monitor confined to the surface lexical layer; building on Floridi’s method of levels of abstraction, I argue that surface-lexical and discourse-structural descriptions are distinct levels -individuated by different observables and supporting different inferences - so that structural auditing is not merely a more elaborate form of surface analysis. Ethically, the compliance–coherence gap this exposes is a level-shifted instance of Goodhart’s Law and the outer/inner-alignment distinction: the very property we optimize for - surface fluency - renders this class of failures invisible and therefore corrodes the calibration of human trust. The normative upshot is that deployed alignment evaluation should be epistemically plural, complementing behavioral monitors with structural audits of an agent’s discourse. I specify, schematically, what such audits would measure, distinguish them from interpretability, and state plainly their costs, scalability limits, and false-positive risks. The argument is conceptual; a pre-registered empirical companion (preprint, not yet peer-reviewed) is summarized as a plausibility anchor and an existence proof of measurability, not as load-bearing evidence."
