Open Access Te Herenga Waka-Victoria University of Wellington
Browse
Explainable-AI-–-building-trust-through-understanding-Final-Version.pdf (2.73 MB)

Explainable AI – building trust through understanding

Download (2.73 MB)
report
posted on 2023-11-09, 03:12 authored by Matt Lythe, Gabriella Mazorra de Cos, Maria Mingallon, Andrew LensenAndrew Lensen, Christopher Galloway, David Knox, Sarah Auvaa, Kaushalya Kumarasinghe
Artificial intelligence (AI) has shown great potential in many real-world applications, for example, clinical diagnosis, self-driving vehicles, robotics and movie recommendations. However, it can be difficult to establish trust in these systems if little is known about how the models make predictions. Although methods exist to provide explanations about some black box models these are not always reliable and may be even misleading. Explainable AI (XAI) provides a meaningful solution to this dilemma in instances where it may be important to explain why an AI model has taken certain actions or made recommendations. These models are inherently interpretable, offering explanations that align with their computations, resulting in improved accountability, fairness and less bias. However, explainable models can also be less capable or versatile and may decrease model accuracy when compared to more complex, less transparent models. The demand for explainability varies with the context. The more critical the use case, the greater the need for interpretability. For example, the need for interpretability in an AI based medical diagnosis system would be significantly higher compared to one used for targeted advertisements. In Aotearoa New Zealand there are already excellent examples of XAI including in health, justice and the environment. The potential for many more systems is substantial, especially when AI decisions affect people or communities in a significant way.

Funding

Commissioned by: AI Forum

History

Publisher

AI Forum

Parent title

Explainable AI Whitepaper

Publication date

2023-11-08

Publication status

Published

Usage metrics

    Reports

    Exports

    RefWorks
    BibTeX
    Ref. manager
    Endnote
    DataCite
    NLM
    DC