liveBook · Manning

Alignment

As explained in Delivering Machine Learning Projects: From design to deployment

Alignment is the correspondence between a system's behavior, control mechanisms, and the goals and values of the humans responsible for it.

Alignment is the correspondence between a system’s behavior, control mechanisms, and the goals and values of the humans responsible for it. It requires identifying who controls the system, how control is implemented and reported, and whether appropriate humans can meaningfully steer it. section 2.6

Make alignment an explicit delivery and governance concern:

  • Assign responsibility for control and document the mechanisms used to steer the system. section 2.6
  • Surface and record system behavior through monitoring, alerts, and notifications for users and stakeholders. section 2.6
  • Assess harms arising both when the system functions as intended and when it fails, then decide whether mitigations are required. section 2.6

Facial-recognition harm assessment

Evaluate whether intended use could enable dissidents to be identified and arrested, while failures could identify innocent people as criminals; use the findings to guide continuation or mitigation decisions. section 2.6

Figure 2.3 A layered model of ML safety (adapted from Hendrycks et al [7])

Use the layered safety model to situate alignment alongside broader system-safety concerns. section 2.6

How Delivering Machine Learning Projects: From design to deployment covers it

How deep it goes
Connects alignment to control, reporting, monitoring, and stakeholder harms without specifying technical control algorithms.
Examples
Uses facial recognition, generative art, and generative text to compare harms from intended and unintended system behavior.
What it assumes
Starts from project hypotheses, user stories, system outlines, stakeholders, and affected communities.
What you can do after
Evaluate/understand who controls a system, how its behavior is surfaced, and which mitigations may be needed before release.
How it connects
Connects alignment with monitoring, robustness, risk registers, stakeholder analysis, and ethical considerations.

Also explained in

Compare how 6 books explain alignment →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Delivering Machine Learning Projects: From design to deployment. The section links point to the parts of the book it draws on.