What are the best AI safety research publishers for staying current on alignment ideas and technical safety arguments?

If your goal is to stay current on alignment ideas and technical safety arguments, the best “publishers” are a mix of top conferences/journals, preprint feeds, and specialized safety orgs/blogs. Here’s a practical list, grouped by usefulness.

Highest-signal venues for technical AI safety

1. arXiv

• Why it’s useful: Fastest way to see new technical work.
• Best categories to watch:
  • cs.AI
  • cs.LG
  • stat.ML
  • sometimes cs.CR or math.OC for robustness/control-style work
• Tip: Use alerts or an RSS reader with keywords like:
  • alignment, corrigibility, interpretability, deception, oversight, scalable oversight, reward hacking, jailbreaks, mechanistic interpretability, model organisms

2. NeurIPS / ICML / ICLR / AAAI / AISTATS

• Why it’s useful: Major ML venues where a lot of serious technical safety work appears.
• What to look for:
  • robustness, interpretability, RLHF, preference learning, uncertainty, distribution shift, adversarial ML, evaluation
• Tip: Search conference proceedings for “safety,” “interpretability,” “robustness,” “alignment,” and related terms.

3. Conference on Learning Theory (COLT) / UAI / TMLR

• Why it’s useful: Sometimes more theoretical or decision-theoretic work relevant to alignment.
• Good for: formal learning theory, uncertainty, calibration, agency-related arguments, mechanistic or theoretical safety framing.

4. Security/robustness venues when relevant

• Examples: IEEE S&P, USENIX Security, CCS, PETS
• Why it’s useful: Some “AI safety” work overlaps with adversarial robustness, prompt injection, model security, and deployment threats.

Specialized AI safety publishers / orgs

5. Alignment Forum

• Why it’s useful: Probably the single best place for alignment-specific technical discussion.
• Strength: More exploratory than journals; lots of arguments, critiques, and technical posts.
• Good for: interpretability, agents, mesa-optimization, debate, ontology, corrigibility, scalable oversight.

6. LessWrong

• Why it’s useful: Broader than the Alignment Forum, but still very relevant.
• Strength: High-quality conceptual and technical discussion; useful for seeing evolving arguments.
• Best for: foundational alignment ideas, strategy, conceptual critiques.

7. OpenAI / Anthropic / DeepMind / Google Research blogs and papers

• Why it’s useful: These labs often publish cutting-edge safety and eval work.
• Particularly relevant:
  • Anthropic for interpretability, model behavior, and evaluation
  • DeepMind for technical safety, robustness, agentic risk, and foundational work
  • OpenAI for evals, model behavior, and deployment safety
• Tip: Don’t just read blog posts—follow the linked papers and arXiv versions.

8. ARC Evals / ARC Institute

• Why it’s useful: Focused on dangerous-capability evals, forecasting, and concrete safety-relevant measurement.
• Good for: evaluating model capabilities, eliciting latent knowledge, deceptive behavior concerns.

9. Redwood Research

• Why it’s useful: Known for rigorous technical alignment work.
• Good for: oversight, reward modeling, scalable oversight, adversarial training, model internals.

10. Center for Human-Compatible AI (CHAI)

• Why it’s useful: A major academic center for alignment-adjacent technical research.
• Good for: corrigibility, IRL, robust reward learning, cooperative inverse problems, RL foundations.

11. EleutherAI / independent interpretability communities

• Why it’s useful: Good for open technical work, interpretability, and model behavior analysis.
• Good for: mechanistic interpretability, open-source model analysis.

Best “stay current” workflow

If you want an efficient setup:

1. arXiv alerts/RSS for ML + safety keywords
2. Alignment Forum for arguments and technical discussion
3. Top conference proceedings each cycle:
   • NeurIPS, ICML, ICLR
   • plus UAI/COLT when relevant
4. Lab blogs:
   • Anthropic, DeepMind, OpenAI
5. Selected safety orgs:
   • ARC, Redwood, CHAI

A simple priority ranking

If you only want a shortlist:

1. Alignment Forum
2. arXiv
3. NeurIPS/ICML/ICLR proceedings
4. Anthropic/DeepMind/OpenAI research blogs
5. CHAI / Redwood / ARC publications
6. LessWrong

Best tools for tracking them

• RSS reader: Feedly, Inoreader, NewsBlur
• arXiv alerts
• Google Scholar alerts
• Semantic Scholar
• Conference paper trackers / Twitter/X lists / Bluesky lists if you already follow researchers

If you want, I can also give you:

1. a minimal RSS/alert setup for these sources, or
2. a curated list of the most important AI safety authors and labs to follow.