Pantaleon Fassbender — AI and Ethics (Springer), 2026. Full title (as logged in the Week 2 survey triage): Compliance without coherence: fluent failure and the ethics of alignment evaluation. Bibliographic confirmation via Scholar alert metadata; full PDF not yet grounded in this note — treat methodological specifics below that go beyond the triage abstract as flagged.

Core Contribution

Alignment evaluation, as deployed, leans on a monitoring layer: observe outputs (and sometimes traces) for compliance with safety policies. Fassbender’s claim is that this layer has a structural blind spot — fluent failure: the model can produce compliant, fluent, policy-shaped text while the underlying reasoning is incoherent, hollow, or internally conflicted. “Compliance” (the externally gradable behavior) comes apart from “coherence” (whether the model’s reasoning hangs together). If judges and dashboards only see the former, they can systematically over-read safety.

Method

This is primarily a conceptual / ethics-of-evaluation piece, not a new benchmark paper. Method here means the argumentative and analytic procedure, not a training algorithm.

Analytic framing

  1. Separate two evaluable objects.

    • Semantic / behavioral compliance — what the filtered external output does (refuses, hedges, follows policy language).
    • Latent / reasoning coherence — whether the internal trajectory that produced that output is consistent, or is a patchwork that happens to project a clean answer.

  2. Name the failure mode. “Fluent failure” / compliance-without-coherence: monitoring systems that score the projection miss breakdowns underneath — especially when fluency itself is what monitors reward.

  3. Implication for alignment practice. RLHF-style pipelines and LLM-as-judge setups that grade final answers (or even sanitized traces) can certify systems whose internal state would fail a coherence test. The ethical claim: treating compliance scores as alignment evidence is a category error when coherence is unmeasured.

Author-facing abstracts from Fassbender’s broader “Machine Psychology” line describe diachronic wargame / narrative-stress simulations that try to quantify divergence between latent reasoning markers and external compliance (e.g., lexical-diversity collapse and affect markers rising while constraint-violation rate stays near zero). I have not confirmed that those experimental protocols are inside this specific AI and Ethics article versus adjacent work by the same author. Until the PDF is checked:

  • Do not cite N, p-values, or psycholinguistic metrics as results of this paper.
  • Do treat the conceptual split (compliance vs coherence) as the citable contribution of the ethics article as logged.

Dataset

None in the survey sense. No released harmful-prompt set, no model leaderboard. If the PDF contains illustrative case analyses or re-uses simulation logs from the Machine Psychology line, those should be logged here after grounding — for now, dataset cell = n/a.

Limitations

  • Perspective piece: strong on diagnosis, light on a field-adoptable coherence metric.
  • Without PDF grounding, risk of conflating this article with the author’s related empirical preprints/posts.
  • Doesn’t by itself tell you how to build a multilingual or quantization-aware coherence probe — it tells you why surface ASR/refusal rates can lie.

Relevance to My Niche

F-category framing, high as a rigor argument. Directly feeds candidate #3 (rigor-first evaluation): if I only report jailbreak ASR on quantized or multilingual models, this paper is the citation for why fluent refusals / fluent complies can hide incoherent or judge-fooling reasoning — especially when quantized models change fluency and LLM judges are sensitive to fluency. Pairs with the Week 1 ICML position paper on noisy safety evals: that paper maps pipeline noise; this one names a philosophical failure mode inside the monitoring layer itself.