Generative AI and extractive AI are both useful approaches to artificial intelligence, but there are important distinctions between the two. In this Insight, we will focus on these differences.
Generative AI specializes in creating new content in response to user prompts. Generative models will generally respond very rapidly, basing answers on patterns learned from a wide corpus of training data.
In contrast, extractive AI models base their responses on information extracted from user-provided data with more focus on response accuracy. Extractive AI models often utilize a Large Language Model (LLM) that is trained on a wide array of natural language data. However, the response itself is drawn from a far narrower knowledge base (e.g. a PDF, spreadsheet, or insurance form provided by the user).
It can be easy to confuse extractive and generative AI because they often use a similar user interface (e.g., both may require uploading a document and providing natural language prompts). Generative AI products for consumers have become very popular in recent years, leading many to favor the generative approach. However, this is problematic when extractive AI is better suited to the task at hand. Understanding the distinction between the two is key to implementing effective enterprise AI solutions.
Neither approach is bad. Indeed, both generative and extractive models are incredibly powerful and have the potential to transform technology, the economy, and our society as a whole. However, issues arise when people do not understand the distinction between the two and try to use one approach to solve problems that fit better into the other paradigm.
We see this most often when users apply generative AI solutions to extractive AI use cases. Let’s start with a use case that is best suited for generative AI.
Generative AI Use Case: “Please write a haiku about life
insurance underwriting.”
When given this prompt on a particular day in February, ChatGPT (a generative language model) generated the following response:
“Risk measured in time,
Underwriting’s careful art,
Life’s policy penned.”
Note, as stated above, this is a response that was generated “on a particular day in February.” Another day or time could have generated a completely different response. There is no presumption or expectation that generative AI must be consistent.
Neither approach is bad...
Now let’s look at a use case that is better suited for extractive AI.
Extractive AI Use Case: “The Language Model will be provided a series of structured documents, unstructured documents and x-rays. A series of insights are to be extracted and those extractions must support future activities in the underwriting process. Therefore, the data points extracted must be correct or else the upstream processes will be inefficient at best. Make sure to identify inconsistencies in the documents. If fields are missing, report as null. Making up answers is totally unacceptable. Oh did I mention, some of the “structured documents” on occasion people will have written all over in hard to discreet handwriting? Not always but sometimes.”
A response for this hypothetical prompt cannot be given because it would depend on the data provided to the extractive model. For an insurer to get comfortable with the above use case, a lot of testing must be done. Before implementing extractive AI solutions, insurers will need to practice prompt engineering, find an optimal level of human involvement in AI workflows, and ensure that security risks are minimized.
...however, issues arise when people do not understand the distinction between the two.
It is easy to see how proponents of extractive AI can get frustrated when the use cases are confused.
Consider the following complaints that are common for
AI solutions:
“Our company isn’t comfortable with AI.
So many mistakes.”
“AI will not work in our processes. We can’t have the answer change every time we ask a question.”
“My kids play with these tools–now I am supposed to use them as part of my enterprise solution?”
“Answers are incredibly fast. Makes me uncomfortable and I am not surprised the answer is often wrong.”
When encountering these sentiments, it is best to go back and examine your use case. Ask yourself, “Is this use case best suited for a generative or extractive approach?” Then ask “Which solution am I using for this use case?”
When generative AI is applied to a use case where extractive AI is a better fit, suboptimal results will lead to bad conclusions. It can lead to organizations deciding that AI doesn’t work when in reality, it is the approach to AI that isn’t working. Real business value can be gained by examining and correcting a company’s approach to AI.
The difference between generative AI and extractive AI is straightforward. Yet, Lazarus AI still sees these approaches confused on a daily basis. This confusion prevents companies from getting real value out of AI and creates misconceptions about the viability of enterprise AI solutions. Both approaches are powerful and both have uses in the modern corporation. Misusing one of these approaches is like misusing any other tool: it can lead to endless frustrations and inefficiencies.
Lazarus is an AI technology company that develops Large Language Models (LLMs) and associated solutions for industries such as insurance. The team at Lazarus is available to discuss all your AI needs regardless of use case or industry. Contact john@lazarusai.com for any feedback.