From the arXiv new submissions of the past 24 to 48 hours in cs.AI, cs.LG, and cs.CL, this digest selects four papers, each posing its own question regarding the trustworthiness of today’s AI systems: who actually delegates to AI, whether global ethical standards are applicable locally, whether multilingual RL training is fairly rewarded, and whether language models can reliably report on their own calculations. Curation was based on substance and thematic diversity: each paper provides a comprehensible method with robust numbers in the abstract, and no two papers address the same subfield.
Who really delegates to AI agents?
Hyeongjae Lee, Jihyang Cheon, and Lanu Kim present their study on the Agentic Adoption Index, introducing a new measure that assesses actual AI delegation rather than mere technical exposure, analyzing around 53,000 agent skill specifications, which they match with approximately 18,000 task descriptions from the US occupational database O*NET. The authors report that delegation is concentrated in professions that previous risk frameworks based solely on technical exposure hardly classified as at risk; peak values are found in medium-wage jobs requiring a bachelor’s degree, while both low- and high-skilled occupational groups lag behind. According to the authors, technical availability explains most of the variance, but not the lower adoption rates among particularly well-educated occupational groups – possible reasons include tasks that are difficult to specify in advance or professional discretion. An earlier investigation had already shown that AI-native startups structurally create fewer entry-level positions – the new study now provides a broader, cross-professional picture of where delegation to AI actually occurs rather than merely being theoretically possible. This matters because labor market policy debates have so far mostly relied on technical exposure rather than measured actual use.
Global AI ethics meets local practice
Ozioma C. Oguine, Munachimso B. Oguine, and co-authors investigate their study on global-local translation gaps in AI ethics, examining whether values such as fairness, transparency, and accountability are indeed universally applicable, as many AI ethics frameworks assume. The basis is interviews with 14 experts from ten countries. The authors report that locally, data protection is understood more as collective-relational rather than individual, transparency more as trust-building accountability rather than technical disclosure, and fairness more as access and representation equality rather than outcome equality – a translation gap between coded global frameworks and lived local practice, which the authors aim to address with more contextual ethical negotiation. An earlier study had already shown that Claude responds differently in terms of warmth or strictness depending on the language – the new work provides a structural explanation at the level of the underlying ethical frameworks. This matters because globally uniform AI ethics standards are increasingly becoming the basis for regulation and product decisions, even though their local interpretation apparently varies widely.
Multilingual RL training unfairly punishes some languages
Chenyu Zhou, Qiliang Jiang, and Xu Zhou demonstrate in their study on multilingual verifier bias in RL training that exact string matching in Reinforcement Learning with Verifiable Rewards (RLVR) – a training method that rewards models based on automatically verifiable answers – transforms format and style differences into language-dependent, incorrect punishment noise. Using a specially developed diagnostic suite and language-specific reward metrics for Japanese, English, and Chinese, they test three models (Qwen3-4B, Qwen3-8B, Llama-3.1-8B-Instruct) on the computational task benchmark MGSM. For Qwen3-8B, the false negative rate for Japanese answers is reported by the authors to be 64.2 percent compared to 12.2 percent for English and 7.3 percent for Chinese, even though the substantive solution is correct in each case; language-specific aggregation rules account for 55 to 78 percent of the selection gap, with over 95 percent of the necessary corrections requiring genuine cross-linguistic support rather than mere format adjustment. An earlier digest contribution had already shown that deceptive behavior occurs significantly more frequently in poorly covered languages – the new study shows that even the training method intended to correct such behavior treats languages unequally. This matters because RLVR is increasingly becoming the standard training method for reasoning models, whose quality would then depend on the randomness of the reward signal in the respective user language.
Language models do not recognize interventions in their own calculations
Emilio Ferrara examines in his study on Open-Weight Masked Introspection (OWMI) across eight open models and over 78,000 individual measurements whether language models can reliably report when targeted interventions in individual model components have altered their internal calculations – compared to sham runs and random disturbances as controls. According to Ferrara, the models achieve an average AUROC of only about 0.5007, statistically indistinguishable from random, with an effect size capped at 0.15 percentage points; one model achieves an AUROC of 0.647 with a trust-based signal. The necessary information is nonetheless present, as linear probes directly on the activations can recognize interventions with 75 to 95.8 percent accuracy, as well as specifically fine-tuned models that achieve nearly perfect results on unseen test cases. An earlier digest contribution had already shown that cognitive errors of language models can be read more reliably from internal state changes than from the visible output – the new study shows that the models themselves have not yet been able to articulate this internal information, even though it is objectively present. This matters because reliable self-reporting about their own calculations would be a prerequisite for many proposed AI safety procedures, which is currently not met.
All four papers are not independently peer-reviewed preprints – the reported numbers come from the experiments of the respective author teams and have not yet been externally replicated. This is particularly true for the ethics interview study, whose findings are based on only 14 experts, as well as for the introspection study, which relies on eight open models from a single author. Whether the patterns hold true for additional models, occupational fields, and independent replications remains to be seen.
