Introducing ratemyNDA: Domain-Specific AI for Contracts

“Dreams are messages from the deep.”

— Frank Herbert, Dune

I. Intro

Welcome back to the Contract Rabbit Blog! Since our last post, we have been hard at work developing some exciting new AI products for contract lifecycle management (CLM). We are now thrilled to announce the launch of our first product: ratemyNDA (www.ratemynda.com), the first domain-specific language model for NDAs. In this post we will walk you through the product’s basic functionality (or you can watch our interview with Artificial Lawyer which includes a full product walkthrough here).

II. Product Design

Let’s start with an analogy. Imagine you’re in the grocery store, eyeing the avocado bin. You don’t consciously examine each fruit’s color, weight, and softness like a checklist. Instead, your brain runs a massive, subconscious comparison between the avocado in front of you and whatever it can remember about every avocado you’ve ever seen. Instant judgment. This – odd as it may sound – is a surprisingly good analogy for how lawyers read contracts. 

When lawyers read an NDA, they don’t just analyze what’s written. They compare what’s written to every other NDA they’ve seen, identifying patterns, spotting discrepancies, and recognizing whether certain language is normal or completely off the map. This ability to recognize these linguistic anomalies, these messages from the deep of the precedent document stack, is what makes legal review valuable and what gives lawyers confidence in their recommendations to clients. It’s not only about comprehension – it’s about comparison. And that is exactly what ratemyNDA is built to replicate.

III. Shortcomings of Existing Offerings

By contrast, when a large language model is “reading” an NDA, it is not comparing it to any other NDAs. ChatGPT is trained on zero signed NDAs. Zero. Don’t believe it? Just ask. In fact, someone (me) did:

It’s not a knock on LLMs; these tools were never designed to understand contracts so gaps like these in their training data shouldn’t surprise us. As a result, no LLM wrapper has the faintest idea of what is normal or not normal in an NDA, or even how to understand legal contracts holistically due to their domain-agnostic data architecture. This is where domain-specific language models come in handy and is key to why our agent handily outperforms chatGPT in the basic task of marking up an NDA, even when we provide chatGPT the exact training data it needs for the task (check out the video of this here).

How do these products make up for this gap in knowledge? Ask any LLM wrapper vendor what to do with your contracts and they’ll likely say: “Write a playbook.” “Playbooks”, or a list of instructions for how your company treats a certain kind of contract, are then fed into a large language model, which refers to the playbook when reviewing and editing a contract. The problem? This violates software’s first commandment: Don’t repeat yourself.

Playbooks are a sloppy and ineffective way to encode information that already exists. The best source of training data for AI contract review is simply the legacy contracts that a user has signed in the past; a human intermediary will only serve to corrupt the data. With ratemyNDA, users can easily leverage all of the accumulated wisdom in their legacy contracts by simply uploading any NDA they want the model to learn, which will then be added to the group of documents the AI is “thinking about” when it compares an incoming document against its training data. 

IV. Product Functionality

So what can ratemyNDA actually tell users about their NDAs? 

1. Analyze: The basic readout above compares an uploaded document against a default corpus of documents (this default corpus can either be one the user has uploaded, or a collection of thousands of publicly available NDAs that gets updated monthly). Each sentence of the uploaded document is ranked in the left-hand table in order of how common it is in the default corpus. The right-hand display shows the uploaded document with rare words indicated by brighter and brighter shades of red, thus allowing a user to pinpoint exactly where a document deviates from the corpus. 
2. Filter: While ratemyNDA comes with a default corpus of public NDAs, users may find that these documents aren’t all exactly relevant for reviewing their specific NDA. For instance, why would we want our AI to consider unilateral employment NDAs governed by Delaware law if we are negotiating a mutual and transactional NDA governed by California law? Had you done a “one shot” training on an LLM, you would be stuck with whatever preset instructions were included in your playbook. However, ratemyNDA offers users the flexibility to retrain their model in real time through its unique FILTER functionality. Essentially, users can type in exactly the subset of NDAs they want the model to think about (or they can simply auto-apply an AI-recommended filter) and the model will instantly rerun its analysis based on consideration of only the relevant subset of data that the user has indicated. As a result, its analysis will be tailored as precisely as possible to the user’s situation. 

3. Execute: After we’ve filtered for the right precedent documents for the model to rely on for its analysis, the next step is executing any necessary changes to the document. 

With a single click, users can optimize their NDA from their perspective and generate a revised version based only on the relevant precedent documents they have identified.

4. Audit: Finally, ratemyNDA outputs are auditable in a way no LLM wrapper’s markup of a document could ever be, since ratemyNDA’s changes are based on real contracts. Users can audit the change as in the image below, essentially “stepping through” their document and tracing any addition or deletion the model makes back to real documents signed by real people. What this feature essentially does is demonstrate to the user that ratemyNDA is not hallucinating when it makes these revisions and allowing them to instantly access the source for any changes it makes. This is not optional - it’s core to how lawyers work. You don’t just hope the language you’re using has been road tested. You have to know it.

V. Conclusion: The Future Is Narrow and Deep

The most powerful AIs in the future won’t be the ones that are trained on untold terabytes of data. They’ll be the ones that learn new data the fastest and can understand this data at the same level of precision as a human brain, and this is exactly what domain-specific models like ratemyNDA do. If Artificial General Intelligence (AGI) is to truly match and exceed the performance of a human brain on every task, it will require a layer of domain-specific language models interacting with LLMs to provide the ground truth for every field of knowledge. In this sense ratemyNDA, with its inch wide and mile deep knowledge of NDAs, represents a small but meaningful step towards AGI.