Six pillars for AI success: how the C-suite can drive results

Does your GPT give perplexing answers, tread down the wrong paths, maybe even “hallucinate”? Does it have access to the right data — or even too much data?

Within control towers, processes and rules must govern how algorithms and models are created and then maintained — such as mitigating bias, enforcing fairness, and enabling explainability and transparency, from the point at which the AI model is created through how it’s continually managed. From the beginning, data security and integrity are crucial. Oftentimes the type of data required for GenAI differs from what’s used in typical AI: it doesn’t exist within neat rows and tables but instead lives in chat logs, emails, surveys and more, and therefore requires far greater preconditioning.

Leaders in a control tower also must define their ethical compass with regard to GenAI and how it aligns with company values. As a consumer, would you feel better or worse if you knew that a GenAI model was guiding your investment decisions or health care diagnoses? Organizations must answer these questions before GenAI is deployed, not after. EY leaders have oriented our AI efforts around fairness, accountability and reliability. These aren’t just words — each correlates to metrics that measure the confidence of any particular solution we deploy. The EY organization, ServiceNow and others are helping companies define, enforce and test their AI principles.

It may feel like a paradox, but humans must be kept in the loop for GenAI confidence. Are the people who will ultimately be affected by GenAI being included in its design, build, testing and deployment? One way is through reinforcement learning: people are core to training the models and rating the output of a GenAI chatbot or algorithm, helping it to determine what is and is not useful. Without that stopgap in place, the efficacy of the models becomes poor. Additionally, it’s important to have mechanisms in place for those people to be able to flag problems and provide feedback on what isn’t working well, what responses are incorrect and what outcomes aren’t aligned with business needs?

GenAI algorithms and outputs must be rigorously challenged. Think of the chatbots that were set up for a narrow purpose that are able to be goaded into spewing hatred, surrendering intellectual property or informing customers about nonexistent company policies. Algorithms set up for certain tasks can also evolve in ways that may seem mysterious to outsiders, without “explainability.” Red teams — which are resources that assume the role of adversaries to try to subvert AI — are crucial for challenging and validating efforts, with independent monitoring.

The EY.ai Confidence Index offers a framework for helping organizations transcend these pitfalls, driving the reliability and explainability of responsible AI to supports enhanced decision-making and more efficient operations.

Principles of responsible AI