AI safety and alignment
AI safety research shows both persistent failures and measurable progress in reducing them. Better benchmark scores or a new model generation do not show safety in real use. Look at the specific model, its tools, permissions and deployment.
Research snapshot · updated Nov 12, 2026
Core thesis
The leading AI labs (Anthropic, OpenAI, DeepMind) acknowledge that aligning superintelligent AI is an unsolved problem. Anthropic's Constitutional AI is the most developed approach. Interpretability research (understanding what models actually do internally) is critical but early-stage. The EU AI Act mandates safety requirements. Geoffrey Hinton (Nobel 2024) left Google specifically to warn about AI risks. The tension: safety research slows capability development, creating competitive pressure to cut corners. This is the most important technology problem of the century.
State of the art (2026)
Safety research shows both persistent failures and measurable progress in reducing them. Anthropic's summer 2026 work on agentic misalignment tested how models behave in constructed scenarios. Anthropic describes these as evaluations, and they do not record real incidents. How often a model fails in a test does not tell us how often it would fail in use. Mechanistic interpretability, the study of what happens inside a model, was named one of MIT Technology Review's 10 Breakthrough Technologies for 2026. A better benchmark score or a newer model does not by itself show that a system is safe. That depends on the specific model, the tools and permissions it is given and where it is used.
Explore the deeper analysis in CanaryIQ
Explore what is covered below. These previews show the structure; log in to read the full analysis.
Signal stack
Evidence stacked leading → lagging
11 signals
Signal stack
Evidence stacked leading → lagging
Technology-native KPIs
Metrics that predict trajectory, tracked over time
3 tracked
Technology-native KPIs
Metrics that predict trajectory, tracked over time
Landscape map
Who builds what, and who depends on whom
63 players · 6 layers
Landscape map
Who builds what, and who depends on whom
Catalyst calendar
Dated events that will move the position
4 ahead
Catalyst calendar
Dated events that will move the position
Technology roadmap
Milestones on the path to maturity
8 milestones
Technology roadmap
Milestones on the path to maturity
Watchlists
Companies, people and papers, each with a condition for removing it
20 · 20
Watchlists
Companies, people and papers, each with a condition for removing it
Decision frameworks
What our view means for your job
Locked
Decision frameworks
What our view means for your job
Thesis changelog
When our view changed, and why
9 updates
Thesis changelog
When our view changed, and why
Change our mind
3 disconfirming conditions
You've read the verdict. The file is much deeper.
The full signal stack, technology-native KPIs tracked over time, the landscape of who depends on whom, the dated catalyst calendar, decision frameworks for every desk, live watchlists and a record of every time our view on AI safety and alignment has changed. All of it is in CanaryIQ.