Humanein the Loop
← The matrix
01

AI should be built safely and transparently

Current path

AI systems are deployed without meaningful transparency about capabilities, training, or failure modes. Incident disclosure is voluntary and uneven. Public understanding lags deployment.

Better future

AI development is visible to the public and to independent scrutiny. Capabilities, evaluations, and failures are disclosed by default. Transparency is a pre-condition of trust, not an optional gesture.

Drift across the three domains

Norms

▲
Advancing17 signals
CHT recommends
  • Treat transparency as a default professional norm, not a competitive liability.
  • Normalize proactive disclosure of incidents and near-misses.
Indicators we track
  • 1.N.aPublic expectation of transparency on AI capabilities
  • 1.N.bOpen publication of safety evaluations as industry norm
  • 1.N.cIncident disclosure as expected behavior

Laws

◐
Mixed12 signals
CHT recommends
  • Mandate pre-deployment evaluations for frontier systems.
  • Require incident reporting to regulators within defined windows.
  • Enable third-party audits with right-of-access.
Indicators we track
  • 1.L.aPre-deployment evaluation mandates
  • 1.L.bMandatory incident reporting
  • 1.L.cThird-party audit requirements
  • 1.L.dRed-team disclosure rules

Design

▲
Advancing22 signals
CHT recommends
  • Publish model cards with substantive technical content, not marketing.
  • Disclose dangerous-capability evaluations publicly.
  • Surface model identity and limits at point of use.
Indicators we track
  • 1.D.aModel cards / system cards published
  • 1.D.bPublic evaluation results
  • 1.D.cCapability disclosure at point of interaction

Recent signals

▲AdvancingDesign · 1.D.bGLOBALJul 9, 2026

Anthropic found a hidden space where Claude puzzles over concepts

Anthropic developed the "Jacobian lens," an interpretability technique providing the clearest view yet into the internal processing of large language models, with findings described as ranging from the mundane to the unnerving.

WhyAnthropic published Jacobian lens revealing internal LLM processing patterns; frontier-lab interpretability research.Public evaluation results
▲AdvancingMajorDesign · 1.D.aGLOBALJul 9, 2026

GPT-5.6 System Card

OpenAI released the GPT-5.6 System Card on its deployment safety platform, providing safety evaluations and deployment guidelines for the new frontier model.

WhyOpenAI publishes GPT-5.6 system card on deployment safety subdomain — frontier model ships with substantive transparency documentation.Model cards / system cards published
▲AdvancingMajorDesign · 1.D.aGLOBALJul 8, 2026

GPT-Live System Card

OpenAI published the GPT-Live System Card on its deployment safety platform, providing safety documentation and evaluation findings for its new real-time AI product.

WhyOpenAI publishes GPT-Live system card on deployment safety platform — new real-time AI product ships with substantive safety documentation.Model cards / system cards published
▲AdvancingMajorDesign · 1.D.aGLOBALMay 5, 2026

GPT-5.5 Instant System Card

OpenAI released the system card for its GPT-5.5 Instant model, detailing safety evaluations and limitations.

WhyOpenAI published a system card for its GPT-5.5 Instant model, fulfilling the expectation of safety documentation for frontier releases.Model cards / system cards published