The power of AI is its ability to turn unstructured data into structured data

So why are we spending all our time generating unstructured data?

People are intrigued by LLMs’ ability to generate new text, change styles, and generally chat. While impressive, turning text written in one style into text written in your own voice (or into legaleze) is essentially taking unstructured data and turning it into more unstructured data. I’m surprised about the lack of attention to doing the exact opposite: turning unstructured data into structured data.

There are many examples of generating the verbiage in legal docs, but imagine understanding the document and creating structured data, such as:

{
  ”type”: “NDA”,
  ”parties”: {
    ”name”:”Party 1”,
    ”name”:”Party 2”  
  },
  "arbitration_state": “DE”,
  "date_of_signature": “1/1/23”,
  "expires_on": “1/1/25”
} 

At best, someone is typing that into a database and at worst it’s just sitting in the document. But with this type of structured data, we can now easily query the database. As an example, we could compile list of all of the NDAs which are coming up for expiration in the next three months.

We can do this type of data operation retroactively, which means we don’t have to know what information we’d like to extract in advance. If, years after getting the original documents, you now have a new question to ask of your data, you can use these natural language understanding techniques to add the additional data and then run your queries.

Generating new documents in different styles is a neat parlor trick and has real use. But the ability to understand well seems at least as important as generating new (often mediocre) content. But I suspect true productivity gains will come from turning unstructured data into structured data we can manipulate and query.