-60% of CSPS Admit to Being Concerned About Potential Impacts of AI Bias-New Research Explores Four Types of AI Bias and Their Risks
Confidence in artificial intelligence (AI) and its ability to enhance network operations is high, but only if the issue of bias is tackled. Service providers (68%) are most concerned about the bias impact of ‘bad or incomplete data sets’, since effective AI requires clean, high quality, unbiased data. This is according to a new survey of communication service providers (CSPs) conducted by Heavy Reading on behalf of Accedian, the leader in performance analytics and end-user experience solutions. Conclusions from the survey highlight the intrinsic link between AI trust and AI confidence, as well as the opportunities for service providers if AI bias can be overcome.
AI is becoming an important part of service providers’ digital and network transformation efforts. According to the survey, AI is a top five priority for more than 40% of CSP respondents. What’s more, confidence in AI has increased over the past year among over half (55%) of service providers, with no apparent loss of confidence in AI. But despite these high levels of confidence, CSPs admit to concerns about AI bias; nearly 60% of service providers believe that AI bias is a concern, with nearly one-third viewing bias as a major concern. According to the report, types of AI bias include:
- Implementer bias, which occurs when those designing the AI system may have a specific goal in mind
- Training bias, occurring when AI training systems reflect the human biases of those who have trained the system
- Lack of expertise, which happens when AI is being used to make decisions in areas where it does not have sufficient experience
- Bad or incomplete data sets, where AI sees erroneous data or incomplete data
CSPs are confident they can manage bias and that it’s not likely to delay their AI progress. But the majority agree that AI bias, for example basing decisions on poor quality data, would affect network operations, most notably for `prediction of future events’, and `pattern detection and correlation’. Reaching the wrong conclusions due to not detecting AI bias could impact future network planning and a worst-case scenario of investing capital to upgrade capacity in the wrong areas. Likewise, failing to address AI bias in anomaly and pattern detection data can mean completely missing ‘undetected’ performance issues that are impacting customer experience and should in fact be prioritized.
“There’s absolutely no doubt that AI is becoming an integral part of network operations, but it’s early days and there’s still a long way to go if we are to see AI deliver its full potential,” said Richard Piasentin, Chief Strategy and Chief Marketing Officer at Accedian. “AI bias is and remains a huge topic of contention. Telecom networks generate significant amounts of data, which will only increase with the move to 5G. This data needs to be collected, cleaned and categorized effectively before it can be used to train an AI system. It’s important that the issues around AI bias are addressed and that CSPs have high quality data and the right analytics tools to unleash the potential of AI on their networks.”
Service providers were also asked where they saw the greatest opportunity for AI in service assurance. ‘Anomaly detection for operations, administration, maintenance and provisioning’ was ranked as the largest AI opportunity, closely followed by ‘predicting network faults’, ‘alert/alarm suppression and automated root-cause analysis’.
Service providers can address some of these opportunities using Accedian Skylight, which helps proactively ensure that networks meet increasingly stringent performance requirements and delivers precise, intelligent performance data to confidently automate service assurance.