#ChatGPT as #Legal Briefwriting Tool

A theorem that is playing out today — and will likely continue for the foreseeable future:

(Machines + Humans) > ((Humans) OR (Machines)) [#Gestalt #Centaur]

An example: My newest adventure in #ChatGPT for #law:?

I asked #ChatGPT to craft #legal #arguments responding to OpenAI 's Motion to Dismiss the #CoPilot litigation.

https://docketalarm.com/cases/California_Northern_District_Court/4--22-cv-06823/DOE_1_et_al_v._GitHub_Inc._et_al/53/

FACTS:

• OpenAI and Microsoft ingested code from GitHub —?to train AI large-language models (e.g., Codex)
• That code was created by third party coders
• Nearly all of the code is covered by various licenses (e.g., MIT).
• The coders sued for breach of contract — among other claims — over alleged violations of license agreements.
• Microsoft filed a motion to dismiss
• Motions frequently contain Tables of Contents — which summarize the movants' arguments

What if we asked Microsoft/OpenAI to argue against itself? Asking GPT argue that OpenAI and Microsoft is liable for breach of contract — by training GPT itself?

STEP ONE: Use the Table of Contents to set a baseline?— then create counterarguments:

Prompt: Below is a table of contents from a Motion to Dismiss in Federal Court.?Please create a bullet-pointed list of counterarguments.

Output: A simple list of counterarguments (essentially negating arguments). But stay tuned...

STEP TWO: Flesh out each claim's elements.

Prompt: For each bullet point above, include sub-bullets for the elements of each claim.

Output:

STEP THREE: Flesh out exemplar facts.

Prompt: Now for each Level 2 sub-bullet (element), please provide Level 3 sub-sub-bullet examples of what could be potentially relevant facts, which show that Plaintiffs satisfied each element. Exclude facts relating to medical injury. Instead, focus on facts relating to commercial injuries and contractual injuries.

STEP FOUR: Now, to gather relevant facts (e.g., from clients, from the world) #ChatGPT will provide exemplar facts.

Prompt: For this factual claim — "OpenAI's actions were the direct cause of Plaintiffs' injuries" — provide factual examples of how a large-language model on training text would cause an author of that training text to lose money.

Output:

SUMMARY:

This took me less than one minute. How long would it have taken a low-level associate?

• If an associate charges $500/hour, would they have spent an hour? Maybe more ($600)?
• I probably wouldn't charge my 45 seconds of prompting
• But I would charge for fleshing out the facts — for now, at least: until the #LLM is able to synthesize facts reliably. (Maybe that's today, with the right input/prompting.) Then all I'm charging the client for is (1) validating caselaw and (2) editing/augmenting.

What are various ways that one could systematize the above?

• E.g., Incorporate into a "Litigation Workflow Pipeline."
• IF Motion to Dismiss THEN <Prompt_MTD_Response_Draft>.
• IF Complaint THEN <Prompt_Answer_Draft>