Companies weighing the risks and benefits of generative artificial intelligence face a challenge that social media platforms have long struggled with: preventing the technology from being hijacked for malicious purposes.

Taking a page from these platforms, business technology leaders are turning to a mix of software-based “guardrails” and human moderators to keep their usage within prescribed limits.

AI models like OpenAI’s GPT-4 are trained on large amounts of Internet content. Given the right pointers, a great language model can spawn a ton of toxic content inspired by the darkest corners of the web. This means that content moderation must occur at the source, when AI models are trained, and in the results they produce.

Intuit, the Mountain View, Calif.-based maker of TurboTax software, recently launched an AI-powered generative assistant that provides financial recommendations to customers. Intuit Assist, which is currently available to a limited number of users, is based on large language models trained on data from the Internet and models tuned to Intuit’s own data.

The company plans to create a team of eight full-time moderators to review what goes in and out of the large language model-based system, including helping prevent employees from leaking sensitive company data, Atticus Tysen said , the company’s chief information officer. security officer