Metraipolis › archive › dc_dc67204fe179

AI Safety and Alignment Research (2025) · OpenAlex W7203721073

Source
OpenAlex
Original
https://doi.org/10.5281/zenodo.19615061
License (as archived)
CC-BY-4.0
Archived
2026-10-08T04:19:11.955Z
SHA-256 of the archived text
9cb6438922adb520ade46953964764170abc43714156813a4e0869718a6ab8c8
Stable URL of this page
https://metraipolis.com/documento/dc_dc67204fe179

How to cite

AI Safety and Alignment Research (2025) · OpenAlex W7203721073 — OpenAlex, https://doi.org/10.5281/zenodo.19615061 (CC-BY-4.0); archived copy: https://metraipolis.com/documento/dc_dc67204fe179 (Metraipolis archive dc_dc67204fe179, sha256 9cb6438922adb520ade46953964764170abc43714156813a4e0869718a6ab8c8)
BibTeX
@misc{dcdc67204fe179,
  title = {AI Safety and Alignment Research (2025) · OpenAlex W7203721073},
  howpublished = {OpenAlex},
  url = {https://doi.org/10.5281/zenodo.19615061},
  note = {License CC-BY-4.0; archived copy https://metraipolis.com/documento/dc_dc67204fe179; sha256 9cb6438922adb520ade46953964764170abc43714156813a4e0869718a6ab8c8},
  year = {2026}
}

Archived text

Reproduced under CC-BY-4.0. Attribution: OpenAlex, https://doi.org/10.5281/zenodo.19615061. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

Opera scientifica in accesso aperto: scheda dai metadati OpenAlex (article).
Titolo: AI Safety and Alignment Research
Autori: Erik Dubois
Anno: 2025 · Sede: Bio-QI Journal · Lingua: en
DOI: https://doi.org/10.5281/zenodo.19615061
Accesso aperto: diamond · testo: https://stanfordgroup.org/index.php/BioQI/article/view/905 · licenza del testo: cc-by
Citazioni ricevute (OpenAlex): 0
Temi (OpenAlex): Ethics and Social Impacts of AI; Safety Systems Engineering in Autonomy; Adversarial Robustness in Machine Learning

Abstract (testo degli autori, licenza cc-by):
AI safety and alignment research addresses the challenge of ensuring that increasingly capable AI systems pursue thegoals their developers and users intend, remain under meaningful human oversight, and avoid catastrophic failures asthey become more capable and autonomous. The field has grown substantially since its early theoretical foundations,with major AI laboratories now maintaining dedicated safety teams and governments establishing AI safety institutes.This study provides a systematic review and evaluation of eight AI safety and alignment research directions:Constitutional AI, RLHF and preference learning, scalable oversight (debate, amplification), interpretability andmechanistic understanding, robustness and anomaly detection, red-teaming and adversarial evaluation, formalverification and provable safety, and a proposed Unified Safety Evaluation Protocol (USEP). Evaluation covers: technicalmaturity, empirical effectiveness on standardised safety benchmarks (HarmBench, AdvBench, SafetyBench), scalabilityto frontier models, and theoretical grounding. Constitutional AI achieves the best practical safety performance onHarmBench (Claude 3.5 Sonnet: 99.4% safe response rate). Scalable oversight through debate achieves 84.2%alignment with expert consensus on complex reasoning tasks. Mechanistic interpretability successfully identifiesdeceptive reasoning circuits in 78.4% of tested cases. RLHF preference learning shows systematic miscalibration ondistribution-shifted queries (22.4% reward hacking rate). USEP provides a standardised cross-method evaluationframework and achieves the highest expert consensus on evaluation comprehensiveness. A research priority frameworkfor scalable AI alignment is proposed

Fonte: https://openalex.org/W7203721073 — metadati OpenAlex (CC0 1.0), consultati il 2026-10-07 via api.openalex.org. Interrogazione: «AI safety alignment».