AI’s Knowledge and Judgment
Asks under what conditions AI uses knowledge, and how it judges when things are uncertain and when goals conflict.
What is needed for several AIs to use the same words with the same meaning?
Even the same sentence means different things depending on the situation it is applied to. If it is not settled whether ‘delivery date’ on a purchase request means the shipping date or the day installation on site is completed, several AIs use the same word while judging by different standards. This does not mean that every participant has to hold all the same information, but it should be possible to check what has been settled and what is unconfirmed, and where the latest evidence is.
What comes in
- Requests and goalsWhose problem, and what problem
- Documents and rulesSource · version · scope of application
- New observationsChanges · exceptions · conflicts
Common context
- Goals and terms
- Evidence checked and its sources
- Conditions to be kept
- Unconfirmed items and current state
Roles that follow
- Material review roleChecks the evidence needed
- Condition review roleCompares rules and exceptions
- Integrated judgment roleJudges by connecting the evidence
From the integrated judgment role back to the common context — changes and new evidence are reflected
The common context is the goals, terms, evidence, constraints and current state to be referred to together. When materials are updated, the affected judgments are reviewed again, and conflicts and unconfirmed states are also recorded.
Read more — Using the same words with the same meaning
The first task in using knowledge is to make the subject and the conditions clear. Even when the same sentence is read, its meaning differs depending on the situation it is applied to. The expression ‘delivery date’ written on a purchase request may mean the shipping date, or it may mean the date installation on site is completed. If this difference is not settled, several AIs use the same word while judging by different standards. The performance of finding relevant documents and whether those documents support the current judgment can be evaluated separately. The common context means the goals, terms, evidence, constraints and current state that several participants refer to in order to carry out a task together. This does not mean that every participant has to hold all the same information. Even if each role deals with different information, it should be possible to check which facts have been settled and what is unconfirmed, and where the latest evidence is.
In a purchase review system, the specifications of the requested item, the version of the rules to be applied, the evaluation criteria and the unconfirmed items can be kept in the common context. When a new quotation comes in, it is marked that the relevant facts have been updated, and judgments that relied on the earlier materials are reviewed again. At this point, a way of recording the state of conflict is needed rather than quietly overwriting two values from different sources. This is because further judgment is needed to decide which value to use.
How should judgments be made when the evidence is uncertain?
A judgment contains both the facts observed and the inferences drawn from those facts. That AI gives a natural explanation and that the actual judgment was made as that explanation says are different claims, and if errors went down on the day a review step was added, it cannot be concluded straight away that the step was the cause of the reduction. Uncertainty arises from several causes — materials may be missing, materials may conflict with one another, or the goal itself may be ambiguous — and if these states are expressed only as a single low confidence score, it is hard to tell which action is needed.
The explanations AI gives are used as material for review, but it should not be said on that basis alone that the whole internal computation has been checked. Stating the scope that can be observed is the starting point of research that can be checked again.
What to do according to the kind of uncertainty
| Kind of uncertainty | Example | Check or response needed |
|---|---|---|
| Missing materials | No specifications on the request | Request additional materials |
| Conflicting materials | The values in the quotation and the rules differ | Check the source and timing, record the state of conflict |
| Mismatched conditions of application | Another organization’s rules were applied | Check the scope of application and when the rules came into force |
| Ambiguous goals | Priorities have not been set | Ask the decision-maker for a judgment |
| Unconfirmed result | Execution was requested but completion could not be confirmed | Judge after re-checking the external state |
Instead of combining them into a single confidence score, response actions are distinguished by kind.
Read more — Judging when things are uncertain
A judgment includes both the facts observed and the inferences derived from those facts. That a product name is written on a quotation is an observable fact. The conclusion that the product suits the purpose of the request needs further judgment about specifications and conditions of use. That AI presents a natural explanation and that the actual judgment process followed that explanation are different claims. If errors went down on the day a review step was added, it cannot be concluded straight away that the step was the cause of the reduction in errors.
Along with the accuracy of conclusions, the Scholars’ Council for AI can study under what conditions additional checks, deferral and re-examination are useful. Uncertainty arises from several causes. Materials may be missing, there may be conflicts between materials, and the goal itself may be ambiguous. If these states are expressed only as a single low confidence score, it is hard to distinguish the actions needed. Appropriate judgment also includes the ability to defer a decision. If deferral is frequent, work is delayed; if there is too little deferral, decisions with weak evidence may be carried out. Which side is favorable depends on the impact of the task and the cost of checking.
When goals conflict, which comes first?
In actual work, AI often has several goals to follow: a purchase request should be processed quickly while meeting the specifications, reducing unnecessary costs and completing the necessary reviews. When evaluation criteria are set, it is useful to separate the results to be achieved from the conditions that must be kept. If all items are combined into one score, an important constraint can be hidden by high scores on other items.
Results to achieve
Reducing processing time can be a goal for improvement.
Conditions that must be kept
Whether the requested specifications are met can be a condition that has to be checked separately.
Read more — Criteria when goals conflict
When goals conflict, it should be decided what to put first. In actual work, AI is often given more than one goal to follow. A purchase request should be processed quickly while meeting the specifications, reducing unnecessary costs and completing the necessary reviews. When evaluation criteria are set, it is useful to distinguish the results to be achieved from the conditions that must be kept. If all items are combined into a single score, an important constraint can be hidden by high scores on other items. Values should lead from abstract declarations to concrete judgments.
The criteria for judgment of the Scholars’ Council for AI should be open to revision. When the setting of use changes or new counterexamples appear, the appropriateness of the criteria is reviewed again. If the reasons for and effects of changes are recorded, it can be traced by which criteria past results were evaluated.
See the principle of evaluating results and failures together