Learning path

AI Safety Research: Labs, Models, and the People Watching Them

AI safety isn't an abstract concern — it's a live research agenda being pursued (and debated) by specific labs, embodied in specific models, and tracked by specific people. This path moves from the organizations doing the work, to the models they've built with safety in mind, to the outside voices holding them accountable. No deep technical background required, though the later steps reward readers who want more detail.

Mixed level7 steps~42 min

7 steps

Begin →

Anthropic
Start here: Anthropic was founded explicitly around AI safety as a core mission, making it the clearest entry point into what safety-focused lab culture looks like.
Read →Beginner In-depth
OpenAI
The lab that first scaled large language models and whose internal safety debates shaped the field — understanding OpenAI's history gives essential context for why safety research exists at all.
Read →Beginner In-depth
Google DeepMind
Google DeepMind brings a distinct research tradition to safety, rounding out the picture of how the three major Western labs each approach the problem differently.
Read →Beginner In-depth
Claude
Claude is the model line where Anthropic's safety research — Constitutional AI, RLHF, interpretability — is most directly expressed in a deployed product.
Read →Beginner In-depth
Claude Opus 4.6
The current flagship release is where Anthropic's latest safety work lands in practice — a concrete case study in what the research actually produces.
Read →Beginner In-depth
GPT-5.5
GPT-5.5 is OpenAI's current flagship, useful here as a point of comparison for how a lab with a different safety posture ships its most capable model.
Read →Beginner In-depth
Zvi Mowshowitz
Zvi Mowshowitz is one of the most prominent outside voices tracking AI safety developments — reading this step shows how the research looks to a rigorous external critic.
Read →Beginner In-depth

Anthropic

OpenAI

Google DeepMind

Claude

Claude Opus 4.6

GPT-5.5

Zvi Mowshowitz