A 360º framework to provide transparent, fair and trustworthy AI solutions

The advancement of digital transformation is promoting the acceptance of augmented intelligence capabilities that power cognitive enterprises. As such, more companies are adopting AI to make business-critical decisions. And while AI has the potential to reshape business outcomes, accelerate growth and reimagine processes that drive a competitive edge, the implications are far-reaching if AI decisions are biased, unfair, or lack explainability. Biased decisions have a detrimental impact on society, business performance and can have legal ramifications. Businesses must interpret the AI algorithmic decisions to ensure they are not encouraging biases , adhere to growing regulatory compliance and safeguard brand reputation.

• Social impact – Bias due to socioeconomic status, ethnicity, religion, gender, disability, etc. propagates from human decision-makers to AI models trained on past data and decisions. Biased AI model decisions impact individual well-being and can deprive them of critical opportunities.
• Business impact - The implications of algorithmic biased decisions adversely affect an organization's trustworthiness, brand reputation, revenue and cost an organization valued customers.
• Legal impact - There are rising regulatory concerns about AI algorithms propagating biases.  Legislative laws such as the EU AI Act 2021 and the USA Algorithmic Accountability Act 2019 mandate that AI systems responsible for critical decisions undergo regular assessments against compliance requirements. Failing this, corporations would be levied heavy fines based on their revenue.

To support the need to address AI biases and ensure trustworthiness and transparency, Wipro Holmes developed ‘Holmes for Trustworthy AI’ – a framework for building trustworthy, reliable and unbiased AI solutions. It helps businesses identify and mitigate bias, improve transparency and fairness in decisions and assist in interpreting decisions made by existing and developing AI models.

Suitable for: Chief Risk Officers, Chief Legal Counsel and Legal office, Data Scientists, AI solution engineering teams

Key Differentiators and Value:

• Identify and mitigate biases that tend to be ‘unfair’ for certain user demographics based on religion, race, gender, etc.
  
• Interpret model decisions in a way humans can understand
• Ensure transparency in AI solutions, flag risks and provide remediation strategies that address biased AI models to ensure decisions are trustworthy, reliable, robust and unbiased
• Provides business levers that allow dedicated decision-makers to review the decisions made by the AI models

Sample Industry-Specific Use Cases 

‘Holmes for Trustworthy AI’ can help to mitigate biases and enhance fairness, reliability and trustworthiness in model decision-making for many industries. Here are a few specific use case ideas:

Figure 1: Sample Industry-Specific Use Cases

How it Works

‘Holmes for Trustworthy AI’ delivers a 360-degree approach to provide explainability and enhanced fairness in AI-based decision-making. It evaluates the data and the model used for decision-making and provides decision support to help evaluate individual cases.

• Data for Trust – Highlights remediate inherent biases  in past decisions and creates a “bias-free dataset”.  It evaluates the data used for model building and ensures all user groups are adequately represented in the dataset and there is no imbalance. It flags data points that resulted in historical bias so that appropriate treatment can rectify the issue. It also performs a risk assessment that provides insights on historical decisions and risks in the data.
• Algorithm for Trust – Provides insights on how the model is made and tests its fairness through a series of concept tests that determine major features responsible for decision making. A  robustness test is performed to flag risks in the model, sensitive features and data points. This phase helps perform intervention-based analysis to validate how each of the attributes impacts model output. It provides actionable explanations for each decision that results in an ‘Explainable and Interpretable Model’ which is no longer a black box that cannot be interpreted and trusted.
• Decision for Trust – For any automated decision, a human should be the final arbiter. Every decision of the AI system is explainable and capable of human input that can provide case-based exceptions for individual decisions. Major decisions are routed to humans and Decision for Trust helps reviewers adhere to both business and regulatory needs. It assesses and highlights the impact and potential risks for each decision, so that the arbiter can make an informed decision.