Wisdom and AI
Research questions from Wisdom and AI · expanded from David Joseph’s analysis
Question 4 of 6 · The change
What will we change?
The change, and what it should lead to.
- Statements
- 18
- Connections
- 9
- Open questions
- 12
How to read it. Read from the bottom up: from each change to the effects it should have.
Part 1 of 5
Review can matter only with supporting conditions
Review points AND usable evidence AND authority and incentives AND feasible alternatives are proposed joint conditions. This is not a guarantee or an observed effect.
Draws on David Joseph’s supplied analysis; the research questions remain questions.
Outline
- Desired effectAI choices can be revised when evidence changes their justification.AI inference · source questions
- ↑ causes all of these together
- ChangeDecision makers use explicit review points to revisit AI choices against emerging evidence.AI inference · source questions
- Already trueUsable evidence about the decision and its consequences is available.Proposed in David Joseph’s analysis
- Already trueDecision makers have authority and incentives to act on review findings.Proposed in David Joseph’s analysis
- Already trueFeasible alternatives to the current decision are available.Proposed in David Joseph’s analysis
- ↑ causes all of these together
Beyond this part · 1 link
- Decision makers revisit AI choices at explicit review points. contributes to Formal review may create unwarranted reassurance. in “Review could create false reassurance”
Part 2 of 5
Make aims and disagreements inspectable
Test the mechanism from explicit aims, through exposed assumptions, to assessment against defensible criteria.
Draws on David Joseph’s supplied analysis; the research questions remain questions.
Beyond this part · 1 link
- Make ethical aims and disagreements explicit. is meant to address Some technical alignment disagreements may conceal ethical disagreements. In the way
Part 3 of 5
Specify what a wisdom-derived practice would change
No particular practice or tradition has been selected. Its mechanism, limits and relevant aspect of judgment must be specified and tested.
Draws on David Joseph’s supplied analysis; the research questions remain questions.
Part 4 of 5
Change options and incentives
Proportionate safeguards or coordination might make restraint or redirection practicable. These are hypotheses, not confirmed effects.
Draws on David Joseph’s supplied analysis; the research questions remain questions.
Beyond this part · 1 link
- Establish proportionate safeguards or coordination. is meant to address Some actors may proceed despite recognising AI risks. In the way
Part 5 of 5
Review could create false reassurance
Formal review may create unwarranted reassurance and premature action. The two unconnected design requirements address those risks; intended coverage is not proof they work.
Draws on David Joseph’s supplied analysis; the research questions remain questions.
Outline
- Possible harmUnwarranted reassurance may contribute to premature action.Proposed in David Joseph’s analysis
- ChangeRequire substantive criteria to be met; completing a review cannot substitute for meeting them.Proposed in David Joseph’s analysis
- ChangeWhere harms could be irreversible, apply safeguards before the action.Proposed in David Joseph’s analysis
Beyond this part · 3 links
- Decision makers revisit AI choices at explicit review points. causes AI choices can be revised when evidence changes their justification. in “Review can matter only with supporting conditions”
- Require substantive criteria; completing review is not enough. is meant to address Formal review may create unwarranted reassurance. The change
- Apply safeguards before actions whose harms could be irreversible. is meant to address Unwarranted reassurance may contribute to premature action. The change
Where this comes from
In the Logical Thinking Process this view is a future reality tree. How a commons works →