Concerns over exposing confidential data is one of the issues preventing organizations from integrating commercial large language models (LLMs) into their workflow, a survey published on Wednesday revealed.
In the survey, which involved 150 respondents from 29 countries, Predibase found that less than a quarter of organizations are comfortable using commercial LLMs like ChatGPT. Instead, organizations favor open-source LLM models that can be customized.
“At a high level, data privacy concerns top the list. In our discussions, nearly 40% of companies voiced concerns about sharing proprietary or sensitive data with LLM vendors,” Predibase’s report said.
Besides privacy concerns, organizations also mentioned difficulties with customization and fine-tuning, the cost of training AI models, latency, and hallucinations among the general challenges preventing them from adopting open-source LLMs.
Since OpenAI’s ChatGPT was released in November 2022, several top organizations, including Samsung and Apple, have banned or restricted the use of LLMs in the workplace over fears that employees could share proprietary data with AI chatbots. In the case of Samsung, the ban was announced after some engineers at the company inadvertently shared confidential data with ChatGPT.
What are Businesses Looking for in LLMs?
While some organizations are still on the fence about LLMs, 44% of the respondents said their organizations are currently experimenting with LLMs, and about 14% said their organizations are using it to aid their work processes in some capacity.
Among the organizations using LLMs, 35% said they use the technology for generative tasks like summarizing and creating content, while 33% said they were using it to extract information and put unstructured data into order. The respondents also mentioned using LLMs to improve their products, gauge customer satisfaction, get answers to questions, and generate recommendations.
“Based on our survey, it’s clear: teams are looking for ways to customize and deploy open-source LLMs without giving up access to their proprietary data to a commercial vendor,” the report said.
While the respondents expressed a desire to customize AI for specific tasks, only 11% said their organizations have done this and are satisfied with the results.
The Challenges That Lie Ahead
LLMs present promising solutions, but organizations must overcome hurdles to customize them for their specific needs.
“The real challenge lies not just in building these capabilities but in ensuring they deliver secure, consistent, and ethical results,” the Chief Analyst of AI platforms, analytics, and data management at Omdia Bradley Shimmin, noted.
If you decide to use commercial LLMs in your organization, we recommend creating strict policies around sharing sensitive and intellectual property data with the bot. Read our article on the privacy risks of AI to understand why it’s important to avoid leaking private information to AI chatbots like ChatGPT.
Predibase’s respondents include executives, developers, data scientists, machine learning (ML) engineers, and product managers from small and large organizations. The survey was conducted between May and July 2023.
For more AI-related industry news, follow us on X (Twitter), Threads, and Mastodon!
