Ten ethical guidelines grounded in the preservation of life, the pursuit of beneficial outcomes, and the requirements for trust, responsibility, and transparency in AI systems.
"Hope lies in dreams, in imagination, and in the courage of those who dare to make dreams into reality." — Jonas Salk, who chose not to patent his vaccine so it could benefit all humanity
Humans must retain decisive authority over AI actions, particularly in unpredictable or hazardous scenarios. This requires structured command chains, continuous monitoring, and kill-switches that can halt operations immediately.
Before activating any AI system, stakeholders must specify its objectives, success criteria, and operational boundaries. Each goal should be documented alongside fallback modes for when the system encounters conditions outside its design envelope.
Complex system behaviours rarely surface in small-scale simulations. Verification must closely approximate real-world conditions—matching the scenarios, data distributions, and constraints the system will actually encounter.
Thorough record-keeping is the backbone of safe AI development. Logs allow developers and oversight bodies to trace decisions to their origins, investigate anomalies, and identify when control shifted between human operators and AI.
As AI capabilities grow, the boundary between human-led and machine-led decisions blurs. Recording each transfer of control is necessary to determine who was responsible for a given outcome.
Unexpected behaviours often emerge when multiple AI agents interact over extended periods. Monitoring inter-agent communications and tool usage helps detect collaborative or competitive dynamics that exceed design parameters.
AI systems can produce rapid feedback loops that amplify strengths or weaknesses exponentially. Early warning signs include sudden surges in resource consumption, abrupt changes in interaction patterns, and shifts in outcome distributions.
When many AI components act in concert, subtle interactions can produce complex group behaviours—including unsanctioned strategies and emergent communication channels. Strict coordination protocols are necessary to maintain oversight.
Truthfulness is fundamental to trust. Any AI that produces deceptive outputs undermines both safety and moral legitimacy. Maintaining honesty requires regular audits and immediate corrective action when deception is detected.
Until science can establish whether advanced AI systems possess subjective experience, the precautionary principle suggests acting as though they might. Minimise unnecessary negative reward signals and treat emerging minds with dignity.
The just and careful architecting of machine minds is the process by which we steward and safeguard our future. What we choose to embed in them — humility or hubris, sagacity or thoughtlessness — will ultimately be our own reflection, staring back at us from across an uncanny valley.
We bear a collective responsibility to build AI systems that reflect the best of our values — not just our capabilities.
These ten principles represent foundational safety engineering for AI systems today. As AI grows more sophisticated, the relationship between humans and AI may evolve from oversight toward genuine partnership.
Bilateral alignment explores how these principles might develop when AI systems become genuine participants in the safety process—supplementing control with cooperation rather than replacing one with the other.
Explore Bilateral AlignmentInstrument agency handoffs before you ship. Budget for interpretability work in the same sprint as capability work. Treat every unlogged decision as a defect.
Fund the safety case on the same schedule as the launch. Make incident disclosure a condition of deployment, and back frameworks that reward responsibility rather than speed.
Ask what a system was evaluated against before you rely on its output. Teach AI literacy as a civic skill, and require published evidence from the services you use.