The future of knowledge management: Talking to documents with generative AI
Data privacy
Depending on the particular use case, the issues of data privacy and regulatory compliance can be a potential inhibitor to deploying publicly accessible language models. The same concern applies to transmitting sensitive data over public APIs for vectorizing enterprise content. “Many organizations are bound by data privacy issues and can’t share their data,” Gupta commented. ChatGPT says it doesn’t expose any outside data, but many organizations are still wary. Obfuscation techniques, such as masking, can alleviate this issue and still enable users to safeguard their data for public models. Other anonymization techniques include adding or subtracting a random number from a person’s age, while still leaving medical patients, for example, in the same age range.
The bevy of open source models such as LLaMa or LoRA is an alternative that “you can run on your own machines,” Aasman said. “Then, when you create a vector, no one else is going to read your documents.” Another recent development calculated to assuage data privacy concerns while using foundation models involves private clouds that have ChatGPT installed in them. According to Aasman, hyper-scalers like Microsoft Azure offer these capabilities so “the government right now can use ChatGPT, and whatever happens, it will not be mixed with the ChatGPT the public uses.” Lastly, ChatGPT allows users to request the system not to remember anything from their documents. “You say you don’t want to store the results of [your] session, then ChatGPT will not use it for training,” Aasman clarified. “It’s a setting that you have to do with each prompt if you use the API.”
Re-envisioning KM
The fundamental pillars of KM—taxonomies, domain-specific data models, knowledge extraction, search, and text analytics—are as pertinent today as they ever were. Now, generative AI has rendered these constructs much more accessible to the enterprise. Its long-standing utility will be determined by surmounting models’ tendencies to generate contrived, inaccurate responses while embedding them into core KM processes.
“It’s an amazing time that we’re in right now—this is a computer trying to emulate a human response on a human interface,” Riewerts said. “Humans have one of the most complex interfaces alive, spoken languages. Behind the scenes, AI and ML are all statistical. It’s numerical based, and it’s computed in that way. There definitely is a balance between the two.”
One of the intriguing future capabilities of generative AI is the ability for humans to talk to their documents. It’s a melding of the mathematics of AI with the intricacies of the spoken word. This new approach enables the examination of a document to glean knowledge previously unavailable. For example, ask a hundred-page commissioned report to tell you its salient points or the extensive employee handbook to talk with you about certain policies. The potential for future KM is enormous.