The principles apply to individual decisions and institutional controls.
02 — THE NINE RESPONSIBLE AI PRINCIPLES
A framework for evaluating AI systems.
These principles are written for the teams that put AI systems into the world: deployers, integrators and fine-tuners. Each principle names a failure domain, the commitment that addresses it and the controls that implement it. Where the obligation differs for organisations that train models, the principle says so.
PRINCIPLE 01 — COMMITMENT
Human Augmentation by Oversight
We commit to assess the consequences of incorrect outputs and automated actions and to design systems with human oversight to ensure aligned and safe outcomes, enabling Human Augmentation by Oversight.
FAILURE MODES
Automation of decisions the AI Act treats as high risk or prohibits
We commit to introduce processes proportionate to risk that mitigate unintended discrimination across groups in our production AI systems, enabling Fairness by Calibration.
We commit to develop tools and processes to enhance the transparency and explainability of AI systems where reasonable, enabling Explainability by Justification.
FAILURE MODES
Unnecessarily convoluted agentic graphs
Stated reasoning diverging from actual behaviour
Lack of deterministic flows where systematic execution required
PRACTICAL CONTROLS
Interpretability and attribution tooling
Transparency artifacts such as model and system cards
We commit to develop the infrastructure that captures the lineage of our AI systems' operations, so historical steps can be reconstructed and reproduced, enabling Provenance by Reproducibility.
FAILURE MODES
Unpinned model versions
Gaps in data governance throughout system
Historical steps that cannot be reconstructed
PRACTICAL CONTROLS
Version pinning for models, prompts and configuration
Determinism in atomic steps through seed/input metadata
We commit to identify, document and systematically support the reskilling of our domain experts and mitigate the impact on workers affected by automation, enabling Reskilling by Design.
We commit to evaluate our AI systems against the domain-specific consequences of their behaviour, and to ensure robust monitoring capabilities are in place for as long as the system operates, enabling Evaluation by Observability.
FAILURE MODES
Benchmark contamination and overfitting
Lack of telemetry, tracing, logging and general observability
We commit to build and communicate privacy processes that protect the data of stakeholders interacting with our AI systems directly and indirectly, enabling Trust by Privacy.
FAILURE MODES
Personal data extractable from a model by prompting
Sensitive data flowing to third parties the user never agreed to
One user's data surfacing in another user's session
PRACTICAL CONTROLS
Privacy impact assessment before connecting data to a model
Memory segregation across users with no shared session context
End-to-end encryption of personal data in transit and at rest
We commit to develop the processes and infrastructure required to secure our AI systems across their lifecycle, and to treat the safety of their actions as part of that boundary, enabling Safety by Security.
FAILURE MODES
Agents granted standing permissions beyond their task
Models, data and tools adopted without vetting or integrity checks
Agent credentials shared across users, tasks and environments
PRACTICAL CONTROLS
Threat modelling against the MLSecOps Top 10 and OWASP agentic attack vectors
Trust boundaries separating model instructions from external content
We commit to evaluate whether our AI systems pursue what was intended rather than what was measured, and to test for misgeneralisation, deception and misuse where reasonable, ensuring Human Alignment by Intent.
FAILURE MODES
No definition of what the system must never pursue
High risk systems that cannot be interrupted once running
Objectives never revisited as the deployment context changes
PRACTICAL CONTROLS
Periodic objective reviews as deployment conditions change
Post-deployment audits of behaviour against the stated intent
Monitoring for divergence between the metric and intended outcome