How to ensure responsible genAI deployments through careful governance

Commercial context

Generative artificial intelligence (genAI) has quickly proved among the most readily deployed forms of AI, with countless mass market applications already in play across the communications industry - from customer care to network operations, genAI is helping to generate new insights, increase efficiency and boost revenues. But for that very reason its benefits must be accrued responsibly, with appropriate attention given to where it can go wrong.

GenAI can for instance, in some conditions, ‘hallucinate’ – meaning it generates content which is misleading, incorrect or offensive to users. Depending on the data it relies upon, it can also lose objectivity and demonstrate irrelevant or harmful biases. It can also be prone to the phenomenon of ‘jailbreaking’ – in which users are able to manipulate the responses of customer-facing AI chatbots with leading questions or statements, to trigger a potentially malign outcome. And, as new input values are gathered over the lifecycle of a genAI solution, ‘drift’ can occur in which the accuracy of predictions or responses can deteriorate as compared to performance in the training and testing phase, undermining credibility. How then can CSPs realize the productivity and revenue generation benefits of genAI while maintaining trust?

The solution

The Responsible AI - Phase II Catalyst is focused on enabling adoption of genAI at scale while maintaining trust among both internal and external users, by exploring the ethical and governance challenges CSPs face while attempting to use the technology in safe and reliable ways. To ensure genAI systems produce trustworthy, accurate, relevant and consistent outputs, the project is showing how to detect and prevent hallucination, while also addressing issues around regulated and protected data, the inclusion of objectionable or adversarial content. The focus use case of the Catalyst is customer-facing genAI chatbots, which by assessing relevant inputs from participating organizations, the project aims to show the potential vulnerabilities of such chatbots and how to protect against them.

The team is producing a white paper which examines these challenges and provides practical guidance to CSPs on emerging frameworks and solutions that will help them scale adoption of genAI to drive growth and improve operational efficiency. It will address key questions including the specific business benefits of genAI to CSPs in customer-facing scenarios, the risks associated with those use cases, and what a viable governance process for genAI therefore looks like, including which technical tools and approaches are required to support it.

The pillars on which such a governance process should rest are risk management, model training and data pipeline and model monitoring, which the report details. Proper risk management throughout the AI lifecycle requires adherence to governance, risk and compliance frameworks from industry bodies such as NIST, regulations such as the EU’s AI Act, and model providers such as AI Verify. Starting with these rules enables definition of business, technical and ethical guardrails needed to create appropriate dashboards within specific organizations to give engineers, deployers and other stakeholders the clear overview they need to ensure compliance across all elements of the given solution. Once those guardrails are in place, it’s also crucial to ensure that individual use cases are properly registered so they can be assessed against all the risk factors and compliance criteria from the relevant regulatory and compliance frameworks.

Flowing in part from the report is a demo video which sets out the risks where AI is not subject to effective governance - with examples of how hallucination can occur and the consequences of those, which can range from individual disgruntled customers to full-blown PR disasters and lawsuits. It also demonstrates the conditions under which jailbreaking can occur and its implications, and sets out how best to design out the potential for these scenarios. The video explains in detail how to manage model training and evaluation against test data, with pre- and post-production evaluation processes to assess against relevant AI metrics for hallucination, bias, and drift.

Applications and wider value

The project offers CSPs considerable commercial opportunities to improve sales, reduce customer churn and boost NPS by guarding against failures in genAI deployments. It also helps to protect CSPs against regulatory fines and reputational harm, lowers compliance costs and speeds time-to-market by offering step-by-step processes and interfaces to adhere to rather than relying on original research. More broadly however it makes a major contribution to the development of genAI across industries, helping organizations of all kinds draw on its benefits while minimizing risk to users and themselves.

“This Catalyst aims to ensure dependable, relevant, and consistent results from genAI,” explains Jindong Hou, Lead Research and Innovation Architect at Vodafone Group. “It addresses AI hallucination detection, objectionable content prevention, and monitors genAI models throughout their lifecycle. By enhancing telecom use cases like customer support and brand reputation, the solution provides CSPs with a standardized framework for evaluating customer-facing genAI, promoting safety, reliability, and transparency. This impact extends to software development practices as well.”

To learn more about how this Catalyst is helping to protect the future of genAI solutions in the telecommunications sector, please see the project space on the TM Forum website here.