Entropy Regularization Buys RL Robustness That Certification Can't Credit
A July 2026 arXiv preprint proves entropy regularization in continuous-time RL lower-bounds robustness to joint perturbations, but ISO 26262 and IEC 61508 cannot credit it.
The archive · Page 3 of 5
Ethics, Policy & Safety
Where AI safety claims collide with reproducible measurement, where training-data harvesting collides with consent, and where deployment outruns the laws and norms meant to constrain it.
The archive
49–72 of 98 articles · Newest first
Fusion's ML Disruption Predictors Have No Shared Validation Standard
A new EAST paper makes real-time tokamak disruption predictors cheap, but ML safety systems lack shared cross-machine validation standards before commercial plants debut.
AI-Generated CSAM Risks Expose Filter-First Safety Gaps
A July 2026 ICML spotlight paper argues preventing AI-generated CSAM requires upstream design controls, because auditing, red teaming, and benchmarking cannot include it.
Treating AI Governance as Code Moves Compliance Into the Build Pipeline
The CANONIC preprint treats AI governance as a compiler check, lowering audit costs but confirming that structural admission cannot detect slop; it only produces an audit.
GitHub Issues Are Now Where GDPR and CCPA Compliance Gets Decided
An arXiv study of 32,820 GitHub issues finds developers negotiating GDPR and CCPA line by line, shifting privacy liability to maintainers and limiting automated scans.
DeepSeek V4 Peak-Load Pricing Breaks Continuous Access for API Users
DeepSeek V4's peak-hour pricing forces API teams to absorb cost volatility or abandon continuous access during Beijing business hours, while self-hosted teams maintain.
LARA Shifts Model Safety from Training to Decode-Time Constraints
LARA injects safety constraints into inference-time decoding via Lagrangian dualization, giving Best-of-N samplers formal guarantees that approach finetuning baselines while.
Stochastic Dominance Reveals Where RLHF Safety Filters Hide Tail Risk
Standard Safe RLHF optimizes for average harm reduction, leaving rare catastrophic outputs hidden in the tail. Stochastic dominance forces teams to bound the entire harm.
Uncertainty-Aware Reward Discounting Cuts Reward Hacking 93.6% in a Preprint
A three-day-old preprint cuts reward hacking 93.6% by down-weighting uncertain reward signals, but the result is unreplicated and may shift RLHF red-teaming if it holds.
Medical AI Liability Needs a Clinical Harness
A June 2026 preprint reframes medical AI governance around runtime-governed clinical skills, exposing who is liable when diagnosis, scheduling, and documentation chain fails.
Does More AI Regulation Actually Reduce Corporate Control?
A June 2026 preprint argues more AI regulation can reduce corporate control, pushing compliance outside engineering teams and eroding oversight of deployed models.
When an LLM Sets Your Price, Whose Long-Term Value Wins?
AIGP's pricing 'alignment' targets platform GMV and ROI over 14 days with no buyer-welfare term, leaving consumers to detect price discrimination from the tag alone.
Combining LLMs Doesn't Escape Shared Failures: A 67-Model Test
A 67-model test proves ensemble accuracy cannot exceed 1 minus β. Pairwise correlation underprices β by 2.5x, making multi-model agreement an unreliable governance check.
Task-Focused VLMs Suppress Hazards They Detect in Isolation, June 2026 Preprint Finds
A June 2026 preprint finds VLMs suppress hazard reports under task load while passing direct-probe evals, indicating safety scores overstate protection in actual deployment.
50 Years of Aviation Certification Expose a Structural Gap in AI Governance
A June 2026 arXiv paper argues AI governance lacks the epoch limits and proof surfaces that 50 years of aviation certification built in, breaking one-time approval stamps.
Do Reasoning Tokens Actually Make LLMs Safer? A New Paper Tests It
A June 2026 preprint finds refusal decisions are locked at a model's first token, undercutting the safety case for premium reasoning modes billed per thinking token.
Machine-Readable AI Usage Terms: Does ODRL's Permission Model Hold Up?
A June 2026 preprint grounds ODRL's permissions and prohibitions in the UFO-L legal ontology, naming who holds the power to declare a violation in vendor AI usage policies.
Who Audits the Safety Rules an LLM Agent Evolves for Itself?
AutoSpec grows LLM agent safety rules from annotated traces and hits 0.98 F1, but readable rules do not prove the rule set is complete. That is the open governance question.
When Vibe-Coded Software Is Safety-Critical, Who Verifies It?
A June 2026 preprint argues vibe-coded code cannot certify under aviation or automotive safety standards, shifting the audit object from prompt to verification artifact.
Can You Trust an AI Robustness Certificate? A Paper Says Verify It
A June 2026 preprint sharpens how neural-network robustness certificates are computed, but verifiers that issue them can return wrong verdicts and miss planted backdoors.
Can a Benchmark Catch When AI Discharge Summaries Drop Care Steps?
CareTransition-Audit scores 11 LLMs on whether AI discharge summaries keep every follow-up and medication step. The best reach moderate clinician agreement, kappa near 0.5.
Do LLM Personality Tests Measure Anything? A New Paper Says No
A June 2026 arXiv preprint finds 81 to 90 percent of LLM personality test variation stems from directional response bias, undermining persona and safety scores.
Community LoRA Mining Raises a Consent Gap for Style Generation
FreeStyle, a June 2026 preprint, mines community LoRA adapters as training data for image generation, shifting licensing burden onto contributors and platforms like Civitai.
When an LLM Narrates a Solver, the Explanation Drifts From the Math
A June 2026 arXiv paper isolates the narration gap in LLM-solver loops: prompt injection can invert a verified verdict at the prose stage, breaking reasoning-log audits.