GPT-3 has no actual understanding of the text and therefore no internal consistency or logic. It doesn't understand the difference between truth and falsehood. All it does is try to sound convincingly like a real person wrote it. You don't want anything that's supposed to be non-fictional to be written by GPT-3.

Does it concern you that GPT-3 will make up quotes and citations? That it will autocomplete both sides of an argument?

The only reason GPT-3 sometimes gets facts right is that it was trained on Wikipedia. It will be as inconsistent as its input.

I can imagine GTP-3 being used to make legal documents that are even more difficult for a human being to parse. That's probably a net-negative for society as a whole if it is used that way.

Of course, it's fun to chuckle at the thought of a lawyer having to read through the most insane document of all time and muttering profanities to himself the entire time.

"The most insane document of all time" is not the point. It is good to make documents unreadable, sure, but they still must be valid, and you have to be sure they are valid, before you use them in the battle. However close to the realms of sci-fi we may be, I believe we are still very far from any machine being able to compete with a professional lawyer in that kind of stuff. Jokes aside, they kinda specialise on it.

What GPT-3 can be good at though is making a huge amounts of text cheaply and quickly. Making up a tons of bullshit to present in court (or putting the other side in a position when they have to) is a way to make litigation last forever, i.e. until the side with less money is forced to give up on a case they could have otherwise won. You can win by virtue of having more lawyers to read and type stuff.

And that way to use GPT-3 based stuff I imagine to be much closer shot. A company with 2 competent lawyers will be able to generate virtually same amount of semi-coherent bullshit as a company with a 100 of lawyers. You can produce an amount of stuff in a single night that is humanly impossible to read, but still if kinda meaningful stuff with some content kinda "relevant" to the matter at hand. At some point it will stop making any sense, the judge won't read this shit, nobody will. You will have to somehow forbid to make stuff more complicated than it must be. What is currently normal will fall apart.

As I said, I don't actually believe it will do any good: people will still find a way to make it way too complicated for normal humans, but acceptable for lawyers. But until they agree on a new norm, it will be painful for everybody involved.

If you have 'known valid' legalese broken into segments and clauses (this already exists in platforms in use by legal professionals to speed up this process, but for legit purpose) and you can convert certain elements into variables (names, dates, addresses, and so forth), you can trivially make a 'valid' legal document that is nearly impossible to parse due to length, by referencing everything under the sun and stars, by introducing lots of situational logic that rests on references, or by simply relying on hyper-specific language and structure.

For example, I've had to read quite a few 1,000+ page legal documents (Companies Act 2006 being one example) and that document gets referenced all the time. I can imagine tools like GPT-3 being used to crank things up to 11 to ensure that no user ever really reads or understands the TOS/EULA/whatever they're reading. Not that most people do anyway.

> too complicated for normal humans, but acceptable for lawyers.

I can totally imagine someone building a tool to parse generated garbage legalese into plain language bullet points and adding the proper visual references/links and document structure so that a human can read it. Perhaps it already exists and I'm just not aware of it.

Lawyers are already miserable: https://www.businessinsider.com/why-lawyers-are-unhappy-at-w...