Responsible AI
Loop Engineering by Fabio Correa, the Applied AI foundation volume on working with AI through framing, feedback, verification, memory, and human judgment.
Last updated 2026-10-25. This policy is rooted in Microsoft's Responsible AI Standard, the framework Fabio Correa worked within at Microsoft. The standard rests on six principles: fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability. Those principles name the values. The book's harness framework remains primary, putting the values to work through owners, controls, and records.
Commitments
Adopted for every system designed, built, or operated under this framework:
- Disclosure. People are told when they are interacting with an AI system and when AI output informs a decision that affects them. Carried by the dialog interface and the presentation layer.
- Recourse. A person affected by an AI-assisted decision can reach a human who can change the outcome. Carried by the fleet layer, through non-AI escalation and the human-in-the-loop posture.
- Opt-out. Participation is a choice, captured as consent, and a person can decline the AI path. Carried by the dialog interface and consent capture.
- Data transparency. The data a system is trained on and acts on is documented, so its behavior can be examined against its inputs. Carried by the memory layer and training-data documentation.
- Measured calibration. Confidence is presented with its evidence and limits, and measured against outcomes, so reliance is earned rather than assumed. Carried by the dialog interface (uncertainty signals), the evaluation layer (measured confidence against outcomes), and the prefrontal layer (visible evidence and limits).
- Meaningful human oversight. A human holds real authority to intervene, with adequate time and information, and appeal and escalation paths are defined. Carried by the materiality dial, which sets how much autonomy a task earns.
- Decision records. Every consequential decision leaves a record a reviewer can reconstruct: decision context, approvals, action log, and outcomes. Protocol traces alone do not meet this commitment. Carried by the defensibility bundle.
- Operator accountability. A named operator answers for each deployment and for the outcomes it produces. Carried by deployment context and fleet governance.
- Training-layer disclosure. Observed refusal and abstention patterns for the deployed use case are documented and monitored, and an escalation path exists when the system declines a legitimate request. Carried by the dialog and presentation layers and by deployment monitoring.
Foundations and further reading
The reading paths behind Chapter 16, organized by discipline:
- Psychology - cognitive load (Sweller 1988), trust calibration (Lee & See 2004), automation bias (Parasuraman & Riley 1997; Parasuraman & Manzey 2010), attention (Wickens et al.).
- Philosophy - moral agency (Floridi), consent in AI partnerships, the boundary between operator obligation and architecture obligation.
- Cognitive science - unified cognitive architectures (Anderson's ACT-R, Newell's SOAR), dual-process models (Kahneman).
- STS / sociology of work - algorithmic harm (Eubanks), accountability voids (Mittelstadt), Ironies of Automation (Bainbridge 1983).
Frameworks and regulation
This policy operationalizes the frameworks Chapter 16 draws on: the NIST AI Risk Management Framework 1.0 and its Generative AI Profile (AI 600-1), Microsoft's Responsible AI Standard v2, and the EU AI Act consolidated text at EUR-Lex.
EU AI Act status as of October 2026: Regulation (EU) 2026/1744 (the Digital Omnibus) is in force as of July 27, 2026, and it moved the application dates. Article 50 transparency duties apply from August 2, 2026 (providers of systems placed on the market before that date must comply with the Article 50(2) marking duty by December 2, 2026); high-risk obligations apply from December 2, 2027 for Annex III systems and from August 2, 2028 for Annex I systems.