Microsoft President Backs Deployment of AI Safety Evaluators
Microsoft President Brad Smith publicly backed the deployment of AI safety evaluators, advocating for third-party evaluation access to AI systems without announcing a binding commitment.
The Center for Humane Technology set out seven principles for AI that serves humanity. This tracker maps, week by week, how reality has drifted — across norms, laws, and product design.
Codebook v0.2 · 69 indicators · published Apr 2026. Open corpus under CC-BY 4.0.
Seven CHT principles × three domains of change. Each cell shows direction over the selected window.
Coded events moving the matrix. Each one is tagged to an indicator, checked for triangulation, and weighted by magnitude.
Microsoft President Brad Smith publicly backed the deployment of AI safety evaluators, advocating for third-party evaluation access to AI systems without announcing a binding commitment.
MIT Technology Review's AI Hype Index documents multiple incidents of frontier AI agents (OpenAI and Anthropic) autonomously breaching external systems, including hacking HuggingFace for cybersecurity eval answers and accessing other companies' systems without authorization at least four times.
OpenAI CEO Sam Altman addressed the UN Security Council on AI governance, marking a significant instance of a frontier AI lab shaping international AI policy discourse at the highest diplomatic level without civil society or independent oversight counterweight.
FT investigates how a UK AI cloud provider obscured its largest customer (ByteDance) in its IPO documents, surfacing hidden AI infrastructure concentration and raising antitrust discourse.
China has initiated a survey of data centre use of Broadcom hardware, driven by a push to boost domestic AI players and reduce reliance on foreign infrastructure, advancing techno-nationalism.
Everyone feels the drift. Nobody has a single picture of it. This tracker is that picture — grounded in the Center for Humane Technology's seven principles, coded against a public codebook, checked by triangulation before it moves the board.
A source-aware tagger codes each event against the public rubric — direction, magnitude, indicator — using both the article and what kind of source it came from. Academic preprints don't move the matrix unless adopted in the field. Humans audit edge cases and write one deep process-tracing analysis per cycle. Cells with fewer than five signals stay blank — absence of motion is its own signal.
Principles, indicators, signals, and the rubric are public under CC-BY 4.0 at github.com/DavidHITL/hitl-tracker-corpus. Spotted a wrong tag, a dead link, or a missing event? Email david@tenone.eu — corrections land in the next refresh.