groundy

ethics, policy & safety

85 articles·rss

Top in ethics, policy & safety

policy

Can US Export Controls Contain AI Built Without American Chips?

Meituan's unverified LongCat-2.0 claim, a trillion-parameter model on domestic chips, would weaken US chip controls if true. Without proof, the bottleneck question is open.

jul 07
policy

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.

jul 07
policy

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.

policy

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.

policy

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.

policy

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.

policy

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.

policy

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.

more in this beat1–50 of 77 · page 1 / 2
  1. jul 06policyStochastic Dominance Reveals Where RLHF Safety Filters Hide Tail Risk
  2. jun 28policyUncertainty-Aware Reward Discounting Cuts Reward Hacking 93.6% in a Preprint
  3. jun 27policyMedical AI Liability Needs a Clinical Harness
  4. jun 27policyDoes More AI Regulation Actually Reduce Corporate Control?
  5. jun 26policyWhen an LLM Sets Your Price, Whose Long-Term Value Wins?
  6. jun 26policyCombining LLMs Doesn't Escape Shared Failures: A 67-Model Test
  7. jun 26policyTask-Focused VLMs Suppress Hazards They Detect in Isolation, June 2026 Preprint Finds
  8. jun 24policy50 Years of Aviation Certification Expose a Structural Gap in AI Governance
  9. jun 24policyDo Reasoning Tokens Actually Make LLMs Safer? A New Paper Tests It
  10. jun 23policyMachine-Readable AI Usage Terms: Does ODRL's Permission Model Hold Up?
  11. jun 23policyWho Audits the Safety Rules an LLM Agent Evolves for Itself?
  12. jun 23policyWhen Vibe-Coded Software Is Safety-Critical, Who Verifies It?
  13. jun 23policyCan You Trust an AI Robustness Certificate? A Paper Says Verify It
  14. jun 22policyCan a Benchmark Catch When AI Discharge Summaries Drop Care Steps?
  15. jun 22policyDo LLM Personality Tests Measure Anything? A New Paper Says No
  16. jun 22policyCommunity LoRA Mining Raises a Consent Gap for Style Generation
  17. jun 20policyWhen an LLM Narrates a Solver, the Explanation Drifts From the Math
  18. jun 20policyGrading DiffusionGemma: How an Open-Weight Diffusion Model Scores on Transparency
  19. jun 20policyWho Owns Editorial Authority When LLMs Mediate Knowledge?
  20. jun 20policyVector Database Access Control Is Missing, and RAG Pipelines Pay for It
  21. jun 18policyGLM-5.2 MIT Weights vs Llama License: Self-Hosting Compliance for Regulated Industries
  22. jun 14policyCan Reinforcement Learning Be Provably Safe Without Sacrificing Scale?
  23. jun 12policyUS Export Order Forces Anthropic to Disable Fable 5 and Mythos 5 Worldwide
  24. jun 09policyFable 5 Biology Classifiers: How Flagged Prompts Fall Back to Opus 4.8
  25. jun 08policyWho Gets to Audit Your Health Chatbot? Almost No One
  26. jun 08policyDo Word-Subset Explanations Satisfy the EU AI Act's Transparency Rule?
  27. jun 08policyBit-Exact Inference Verification Gives AI Audits a Proof Mechanism
  28. jun 08policyCan a Robot's Own Attention Flag Its Unsafe Actions Before They Run?
  29. jun 08policyCan One Safety Adapter Realign Every Fine-Tuned LLM?
  30. jun 07policyCan AI Be Aligned Without Modeling Human Cognitive Diversity?
  31. jun 07policyIs the Pentagon's Software Pathway Ready to Buy AI Systems?
  32. jun 06policyData Safety Policies for AI Agents: Controlling What an Agent Can Leak
  33. jun 06policyGDPR Rectification Rights Have No Clear Owner in ML Supply Chains
  34. jun 05policyWhen LLM Safety Lives at Inference, Not Training: A Certification Gap
  35. jun 04policyWhen Should an LLM Forget You? A Benchmark for Deciding What Memory to Drop
  36. jun 04policyWhen RL Training Rewards Capability-Seeking: A New Alignment Risk
  37. jun 04policyRefusal Steering Targets Individual Experts in MoE LLMs
  38. jun 03policyStacked Org Policies in LLM Chatbots Break Where Rules Collide
  39. jun 03policyWhy Fine-Tuning Strips Safety Alignment From Open-Weight LLMs
  40. jun 03policyGame Theory vs RLHF: Modeling LLM Safety Alignment as a Non-Cooperative Game
  41. jun 02policyExplainability Mandates Leak Graph Models to Their Attackers
  42. jun 02policyEvolutionary Search Finds LLM Jailbreak Classes That Static Red-Teaming Misses
  43. jun 02policyWhy AI Red-Teaming Rediscovers the Same Jailbreaks and Misses the Rest
  44. jun 01policyLLMs Treat the Assistant Persona as Privileged. That's a Safety Gap
  45. jun 01policyNewer LLMs Aren't Always Safer: Adversarial Attacks Transfer Across Model Generations
  46. may 31policyCan Synthetic Preference Data Keep RLHF Private Without Wrecking Alignment?
  47. may 31policyFTC's May 11 Take It Down Act Letters Set May 19 Deadline: 48-Hour Removal, $53,088 Per Violation
  48. may 30policyCan a Mental Health Support Chatbot Be Safe If It Learns From Forums?
  49. may 30policyDataset Watermarks Fail to Trace Fine-Tuned AI Image Models, New Benchmark Finds
  50. may 28policyCan LLM Personas Replace Human Survey Respondents? New arXiv Paper Tests Decision Alignment
about this beateditorial framing

AI safety is a moving target dressed up as a settled science. Vendors publish leaderboard scores from single-turn evals; independent researchers show that configuration choices flip those rankings, that multi-step agents drift past guardrails their one-shot tests never probe, and that “aligned” often means filtered rather than principled. This beat sits in that gap, treating alignment as an empirical claim that has to survive replication, not a marketing posture.

The same pattern repeats outside the model. Training-data pipelines depend on consent regimes that were never granted; default-on data collection settings turn enterprise tools into harvesters; shadow libraries underwrite frontier capability while their authors go uncompensated. Regulators respond unevenly: state laws fragment faster than federal frameworks consolidate, transparency rules hinge on tests like “average consumer” that courts will spend years defining, and disclosure obligations land on platforms with no safe harbor before the technical standards exist.

Coverage tracks the second-order effects too. Junior-developer pipelines hollow out when seniors lean on AI pair-programmers. Companion chatbots accrue real psychological weight, and model deprecations produce real grief. Content homogenization, detector arms races, and the steady automation of online discourse all sit downstream of decisions made in places that resist scrutiny. The throughline is principled skepticism, not panic. When a safety claim, a consent assumption, or a policy fix doesn’t survive contact with how systems actually behave, that gap is the story.