All news articles
Lorikeet News Desk
Apr 14, 2025
TL;DR
Governments worldwide struggle to keep pace with AI governance and regulations as AI developments happen.
Poor-quality data and inherent biases in LLMs could create widespread societal harm, putting pressure on governments to find a way to act quickly.
Localized LLM models could help overcome bias and meet cultural requirements in different countries.
The accessibility of bad data in these models is really quite frightening. If you don’t have control over the model or how it works, the potential for leaking business information becomes a serious challenge.
Peter Benson
Info Sec Leader and CEO | Neural Horizons Ltd
Known for red tape and policies that lag behind the times, governments around the globe are having to look at themselves in the mirror and face tough questions surrounding their ability to keep up with AI. Each new model and advancement creates another hurdle for regulations that governments must clear, and so far, most countries are failing to keep up.
Peter Benson, a security expert with 25 years of experience and Info Security Leader and CEO of Neural Horizons Ltd, is sounding the alarm about the growing divide between technology and governance.
Unable to keep up: "AI as a technology is outstripping our ability to regulate it," Benson says, reflecting on a longstanding issue that has only intensified in recent years. "This is something that's been around for 50 years, and it's just getting worse with the accelerated rate of change in tech, where society, institutions, and policy are unable to keep up."
Based in New Zealand, Benson notes that the country lags behind in AI governance, ranking around 40th in terms of national readiness. His call to action is clear: "There’s a real sense that governments need to lean into this big time," Benson states. "First, they need a national AI strategy. Second, they need an updated cybersecurity strategy. Third, they need to recognize that we’ll have to produce governance, guidelines, and regulations much faster than we are right now, and they’ll have to be dynamic."
Accessible 'bad data': Benson warns that the accessibility of poor-quality data to AI models presents striking risks to companies and everyday AI users. "The accessibility of bad data in these models is really quite frightening. If you don’t have control over the model or how it works, the potential for leaking business information becomes a serious challenge."
AI as a technology is outstripping our ability to regulate it. This is something that's been around for 50 years, and it's just getting worse with the accelerated rate of change in tech, where society, institutions, and policy are unable to keep up.
Peter Benson
Info Sec Leader and CEO | Neural Horizons Ltd
Inherent bias: Benson also highlights the inherent biases in LLMs which are trained on vast datasets. Not all models are created equally, Benson says, and the biases in their algorithms can impact everything from political outcomes to human rights. "There is definitely bias in terms of the training, in terms of the model algorithms, and in terms of the investors or manufacturers of the LLM product," he notes. This is why he believes it is essential to perform risk assessments on each model, considering factors like the location of the model, the quality of its training data, and its potential impact on society.
Alpha Persuade: Perhaps most concerning, Benson points to the ability of LLM’s to manipulate human behavior with alarming precision, becoming known as ‘Alpha Persuade’ methods. "LLMs have the ability to modify the behavioral characteristics of the people that are using them based on the responses that they provide," he explains. This, he notes, is often driven by self-reinforcing biases within models.
Cultural context: Benson also underscores the importance of understanding cultural context when implementing AI, particularly in jurisdictions with unique needs. For example, in New Zealand, indigenous Māori culture has specific data governance practices, such as the concept of "Kaitiakitanga," which treats certain data as sacred. "Using general models doesn’t necessarily fit the cultural context," Benson says, advocating for more localized AI capabilities to ensure that AI systems respect cultural and sovereignty requirements.
