Insight

Transparency is key to responsible AI

Building transparent AI models not only enables us to explain the data outcome in a responsible manner, it also helps us to overcome our fear of the unknown. In this quick-read, one of our in-house AI experts shares his thoughts on transparent AI.

Nowadays, artificial intelligence (AI) is becoming a capability of many businesses.

AI systems learn from their environment and computes outputs based on data in a manner that for some domains resemble the way a human processes information. A well-coded AI-model is a strong and valuable asset. However, if the data fed to the AI model is flawed or biased, it will corrupt the model and its predictions.

Therefore, businesses must learn to understand and take action against potential AI risks.

At Deloitte, we are strong advocates of data ethics and transparent AI. However, putting transparency into practice is no walk in the park: What information should be shared? In which situations? And with whom?

We had a virtual sit-down with Andreas Keller Leth Laursen, who manages our Analytics & Cognitive offering, to get his take on transparent AI.

It's all about the "why"

Deloitte's Trustworthy AI framework introduces six fundamental considerations that can assist in building a trustworthy AI business strategy. It is made to help businesses identify and prevent potential risks related to AI ethics. One of the six focus areas discusses how to make AI more transparent and explainable. 

According to Andreas Laursen, the most important question to ask when it comes to transparent AI is not what it does, but why.

An AI model is only as good as its data. So, instead of blindly accepting the data outcome, both companies and customers need to understand why the AI model has come up with its predictions. Why does it produce this particular decision? And how did the model reach its decision?


Credit scoring with AI as a supportive tool, not as human replacement

An example of an AI-powered use case is credit scoring. The AI model uses data to provide a personalized credit score based on information such as income, debt, credit history, age, gender, job situation, and other socio-demographic information.

"In this situation, the applicant should have the right to know how he or she is potentially being discriminated against. What information improves the applicant's chances of approval, and what information steers the model towards a credit rejection? The AI developer has to consider this requirement when developing the AI model," says Andreas Laursen.

"In building such a capability into the AI system," he continues, "you enhance the lender's decision-making process, you do not replace it. When the lender uses the AI model's outputs as inputs in his or her decision-making process, he or she gets a credit-scoring solution that combines the strengths of AI and humans. However, it remains the lender's responsibility to interpret and validate how the model determined the credit rating of the applicant before presenting a final decision."

Overcoming our fear of the unknown

With the increased focus on and awareness of GDPR, many of us now demand insight into how our data is used.

"Opaque AI models, whose predictions are impossible to inspect and explain, result in more reluctance to use the tool, increasing the risk of mistrust and skepticism " says Andreas.

Hopefully, the AI solutions of the near future will be a lot more transparent – in other words, open to inspection and with fully explainable data outcomes. 

According to Andreas Laursen, transparent AI will help us overcome our innate fear of the unknown, increasing consumer trust in the technology, and thus the appeal of using AI solutions:

"Transparent AI will – without a doubt – have a self-reinforcing effect. Once a user has been invited into the engine room to develop an understanding of the reasoning behind the predictions, he or she will become more comfortable with sharing personal data and will likely keep using the system," concludes Andreas.

Wondering how AI can help your business? Reach out to our expert using the contact information below.

Fandt du dette nyttigt?