Generative AI (GenAI) has grown explosively, and so has public concern over the technology's safety, accuracy and fairness. Questions like: "Is generative AI coming for my job?", "Is the technology increasing the spread of misinformation online?" and "Are biases in the training data putting me at a disadvantage?" are top of mind and in news headlines.
All of this has exacerbated public demands for insights into GenAI. In fact, a survey conducted by TELUS International showed that 71% of respondents agree it's important for companies to be transparent with consumers about how they are using generative AI. This is integral to building trust in the technology.
But building that trust is no small task, in part because the inner workings of GenAI technology can be a bit of a mystery. Large language models (LLMs) — the forces behind generative AI — are trained on unimaginably large datasets. GPT-3, for example, is estimated to have been trained on 45 terabytes of text data, or the equivalent of about one million feet of bookshelf space, according to McKinsey & Company. While the size of these models contribute to their capabilities, it's also caused their inner workings to become increasingly opaque. Comprehending and retracing how LLMs output the results they do can be difficult and sometimes impossible, even to the data scientists who build them. Known as the "black box" AI phenomenon, this is what "explainable AI" aims to address. Employing an explainable AI framework allows humans to "look under the hood" of the models to better understand how they arrive at the output they do.
The concept of AI explainability dates back to research published in the 1980s; however, the phrase "explainable AI" is said to have been coined by the Defense Advanced Research Projects Agency (DARPA). The agency, an arm of the United States Department of Defense, conducts research that assists the government in making decisions on technological investments used for national security. DARPA's explainable AI program, launched in 2017, aims to create machine learning techniques that produce more explainable models, as well as offer insights into how emerging AI technology can be understood, trusted and managed.
While DARPA may have initiated the concept of explainable AI, many other organizations have followed suit, acknowledging the benefits of greater clarity in artificial intelligence design and implementation. Organizations today of all sizes and across all industries are integrating explainable AI principles into their business practices.
Explainable AI principles
The key to an explainable AI framework is trust, and there are several ways to foster it. Consider the following four principles that serve as the foundation of explainable AI:
1. Maintain transparency
Open communication about how generative AI can be, or is being used, is imperative to explainable AI. This can be accomplished through accurate policy and procedure documentation.
One example of a formal, written communication is an acceptable usage policy (AUP), which outlines how generative AI can, and should, be used in relation to the values of the organization. This could include its practices regarding data privacy and security, how the business monitors GenAI usage, how the risk of bias and discrimination in the organization's generative AI systems is mitigated and how the technology can and cannot be leveraged. For example, Salesforce published an AUP to offer clear guidelines as to how its generative AI products may not be used, including for weapons development, biometric identification, medical or legal advice and more.
Clear communication and documentation can also help mitigate fears employees may have about the adoption of the technology within your organization. "If the composition of the employee workforce will be affected [with GenAI implementation], the communications plan should also detail out the available transition path to really exciting emerging new jobs that your company will be supporting," Steve Nemzer, TELUS International's director of AI growth and innovation, said in the webinar, Building trust in generative AI. The plan should demonstrate the upside of the technology and how it can lead to significant benefits. For example, a customer service agent could use GenAI technology to create knowledge-base articles in real time for faster problem solving and operational efficiency.

Building trust in generative AI
Brands are eager to reap the benefits of generative AI (GenAI) while limiting potential risks. Join TELUS International’s Steve Nemzer, director of AI growth and innovation, as he shares best practices for leveraging GenAI without compromising your organization’s goodwill.
Watch the video
2. Know your model's training parameters
The parameters of the GenAI model should be made clear. This includes, but is not limited to, the training that went into it, how it was created, the data that was used and where it came from, what the model excels at and its benefits.
Training your customized GenAI model using enterprise or industry-specific data can help you achieve this goal. This is because the scope of this data will be strategically focused, which helps to reign in the parameters. It also helps to ensure your model's output will be more precise for your business.
You'll also want to integrate content policy guardrails to ensure the model avoids outputting unwanted content. For example, guardrails can be implemented to restrict the model from outputting content related to industries that are outside of your focus area. So, if your sector is finance, you can restrict your model from giving you output related to healthcare or real estate.
"The result is, hopefully, a high-performing customized GenAI model that understands your business and can power agent assist tools and self service tools and generate personalized content for customers," said Nemzer.
3. Meet regulatory requirements
As GenAI use cases become increasingly common in regulated industries, it's essential to demonstrate adherence to specific policies and statutes. Include provisions in your company's generative AI AUP to address compliance with regulatory requirements.
It's important to note that these regulations continue to evolve, as the proliferation of generative AI use cases has exacerbated the call for intervention. "The White House, the European Union, governments around the world, all are workshopping new directives for responsible AI development," said Nemzer.
For example, the European Union (EU) has proposed the Artificial Intelligence Act, which if passed, would assign AI to one of three risk categories. First, applications categorized as presenting unacceptable risk, such as government-run social scoring based on personal characteristics or social behavior, would be banned. Second, applications deemed to be high-risk would be subject to legal requirements. For example, AI-powered tools used to scan career or financial applications to rank applicants. And third, AI applications that are deemed as not posing any risk would be left unregulated.
The EU's Artificial Intelligence Act is just the first of many regulations in the works. The Biden-Harris administration in the U.S., for example, has proposed the Executive Order on the Safe, Secure and Trustworthy Development and Use of Artificial Intelligence. It aims to govern the development and use of safe and responsible AI. As outlined in the order, "The rapid speed at which AI capabilities are advancing compels the United States to lead in this moment for the sake of our security, economy and society."
Furthermore, last year the Biden-Harris administration launched the artificial intelligence cyber challenge to help protect U.S. software deemed to be of particular importance, such as code that helps run the internet and critical infrastructure. The competition includes collaboration with prominent AI companies and will award almost $20 million in prizes. "This is basically just the start," said Nemzer. "People are talking about billions of dollars in prizes."
4. Document decision-making results
Decision-making documentation helps provide a plausible explanation for a model's output as opposed to simply taking the content produced by the model at face value. It enables developers and users to trace output back to a specific measure, like a parameter set during the model's training or fine-tuning, or a guardrail implemented as a safety control.
Having documentation of how the model got to where it is provides an audit trail should things get off-track. "The most important piece is to maintain an understanding of how the models are arriving at the outputs they generate," said Nemzer. "We don't want to get to a point where the technical teams and the data scientists don't understand how the model works."
Advance your generative AI initiatives
As your organization adopts generative AI, building trust in the technology is integral to your continued success. Gain more valuable insights on maintaining transparency with customers and employees, including best practices for integrating critical bias-mitigation initiatives, in our new webinar, Building trust in generative AI.