Generative artificial intelligence (AI) and large language models have enormous potential that has led to the rushed rollout of products and systems that have not been fully thought-through. The hype is such that many are jumping onto the Generative AI bandwagon without first considering the downstream implications or waiting for offerings that come with better performance and usefulness to mature.
“Generative AI brings risks as well as benefits,” confirms Theo Hill, senior director of product management at Smarsh. “Archives will have to handle the increased volumes of content that these models will create, as well as identify new types of risks in that data. New techniques are needed to determine those risks and we are focused on how technology can be used to better identify risks against a landscape of radically increased data generation.”
Finding the data points that suggest compliance risks exist becomes increasingly challenging as data volumes and the reliance on AI increase, introducing more inaccuracies. “Generative AI can create egregious mistakes but, because those mistakes are nicely formatted and confidently created, they’re believable,” says Hill. “That means your teams have to check everything which can negate the purpose of using AI in the first place.”
The risks don’t stop there. “Generative AI has been trained on general internet information which contain bias. These biases can be amplified as AI is increasingly utilised,” he adds. “Biases can lead to unfair targeting of certain groups, which is a serious issue in and of itself. However, an often-overlooked issue in supervision and surveillance is that bias toward risk detection to these targeted groups can create blind-spots elsewhere in your monitored population — thus degrading your surveillance effectiveness rather than improving it.”
Businesses therefore need to carefully assess the ways in which they will use Generative AI and Large Language Models (LLM). “Some models are built by scraping the internet and this is facing legal challenges from owners of copyright content so the models may have to be updated over and over as these cases are resolved over time,” explains Hill.
