Successful adoption of enterprise language models will need to overcome significant challenges to effectively, responsibly and securely train them.
At a time when the hype surrounding Open AI’s ChatGPT has prompted 45% of executives to increase their investments in artificial intelligence (AI)*, the new expert.ai research “Large Language Models: Opportunity, Risk, and Paths Forward” reveals that more than one-third of enterprises (37.1%) are already planning to train and customize language models to meet their business needs.
A significant majority of enterprises (78.5%) realize that the efforts required to effectively train a usable and accurate enterprise-specific language model is a significant undertaking which will require dedicated resources and budget. Almost three-quarters of enterprises surveyed have budget or are discussing adding budget to support large language model (LLM) adoption.
“Enterprise specific language models with a human-centered approach are part of the future”, says Marco Varone, founder and CTO of expert.ai. “Business natural language use cases always require some degree of domain-specific training applied to existing proprietary or open-source LLMs. Specific enterprise models can be smaller, more efficient, faster and less resource-hungry while still maintaining high performance. Having subject-matter experts monitoring and refining data and inputs throughout the process ensures accuracy, transparency and accountability.”
The study shows that while only a few surveyed favor a LLM training moratorium (21.2%), the majority of AI professionals and practitioners (70.6%) point to the need for commercial and malicious use in AI regulations. Top adoption challenges include: data privacy and security (73.1%), accuracy and quality for production model deployment (51.2%) and knowledgeable resources on how to build and train LLMs (40.7%).
For companies that are prioritizing AI transparency and responsibility, generative AI and LLMs may bring real risks for their ESG objectives and performance, with truthfulness (69.8%), bias (67.3%) and leaks of proprietary data (62.6%) as the top concerns.
Regardless of the direction an organization chooses, basic AI data governance principles still apply to generative AI and LLMs. While 24.0% of survey respondents indicate that further restrictions need to be put in place to test LLMs and provide clear communication of applied policies, 38.8% feel some additional degree of freedom should be encouraged, and 34.3% cite that current principles are adequate.
*Gartner Poll Finds 45% of Executives Say ChatGPT Has Prompted an Increase in AI Investment – May 3, 2023, link
About the survey study
The expert.ai “Large Language Models: Opportunity, Risk, and Paths Forward” report summarizes survey results from 300+ business, technical and academic natural language AI experts from around the world. The interviews were conducted online by expert.ai in April 2023 with the goal to explore potential opportunities and risks associated with generative AI and provide recommendations for a path forward that enterprises can use in their development and deployment of LLMs.