AI Intelligence
Partially live — in active developmentMake AI outputs understandable, contestable and open to correction.
Understand the factors, evidence and limits behind AI outputs
Explainable AI helps people understand why a system produced a particular output, what information influenced it, and where uncertainty or limitations remain. In DISHA 4.0 HCOS™, explanations should be matched to the task and user. A useful explanation is not merely a technical feature-importance chart; it gives the person a meaningful way to inspect, question and correct the result.
Explore how DISHA can make intelligent outputs easier to inspect, question and improve.
Provenance tracking, confidence labels and audit trails are live; deeper decision narratives are in active development.
In depth
Explanation by audience
An individual may need to know why a learning resource was suggested. An analyst may need to inspect model drivers and validation. An administrator may need model version, data lineage and audit evidence. Provide layered explanations so the first view is understandable while technical detail remains available to authorized reviewers.
Show relevant factors
For a recommendation, identify the profile attributes or preferences that contributed, such as a stated goal, selected skill or preferred format. Distinguish actual inputs from inferred or missing information. Avoid implying causal relationships when the system only identifies statistical associations.
Evidence and uncertainty
Where a result is grounded in documents or structured data, link to the relevant evidence when permitted. Display confidence only when it is meaningful and calibrated. If an output is generated, approximate or based on incomplete information, state that clearly.
Correction and contestability
Users should be able to correct inaccurate profile data, change preferences, dismiss recommendations and report a problematic result. For consequential workflows, provide an appropriate human review or appeal route and preserve an audit trail of material changes.
Explanation quality assurance
Test whether users understand explanations, whether they reveal protected information, and whether they remain faithful to the model or rules. Explanations must not be fabricated after the fact or overstate what a model can reveal.
How it works
Capability domains
Frequently asked questions
Does an explanation prove the model is correct?
No. It helps interpret an output but does not replace validation, fairness testing or human judgment.
Can I challenge a result?
Users should have ways to correct input data, change preferences and request review where appropriate.
AI Intelligence
Experience it inside the live platform.
Open the module that carries this capability, or talk it through with the team.