GPT-4 Can’t Replace Striking TV Writers, but Studios Are Going to Try

As someone who is very close with (and supportive of!) people in the WGA— this is not what they’re demanding. They’re demanding that a studio cannot classify something AI-generated as “source material” in the same way a book, movie, play or another WGA script is. They want things generated by AI to be classified as “background material” (e.g. a Wikipedia article, newspaper clip, etc).

In a nutshell (but also, I’m not an expert), this gets down to how studios can pay or hire writers:

- A studio CAN hire a writer at a day-rate (think, as a one-off gig rather than stable job) to “touch up”/“rewrite”/etc source material such as other scripts. This is usually at a lower rate than if you hire the writer in full capacity. These jobs are usually short — maybe a couple weeks, could be just a single day.

- A studio already CANNOT hire a writer on a day-rate, give them a Wikipedia article, and claim that a writer making a story off of it is simply a “revision”. They must be hired in their full capacity, often with a writers room, to flesh it out. These jobs are usually longer, likely months.

The WGA’s demand is that ChatGPT and similarly generated content be placed into the second category.

A writer could still “use” ChatGPT themselves, and a studio could also provide writers with a ChatGPT-produced script, but the writers would need to be paid as if they were working “from scratch” (though, conceivably the studio still saves money in this context, as the room could be less weeks than otherwise). As the article points out, their fear is that a studio will begin “writing” unattributable AI-generated scripts, give them to writers, and demand they turn it into a full show or movie— on a day rate.

It seems like a very sane middle ground.

Personally, I think without special arrangements not in evidence, it will work fairly poorly. The RLHF tuning in particular seems to severely damage the stylistic flair and flexibility of ChatGPT-3/4. I don't mean just the really hilarious problems around BPEs and rhyming poetry (for a good laugh, ask ChatGPT to 'Write a non-rhyming poem'), but any kind of fiction comes out tending strongly towards saccharine, milquetoast, happy-ending, Hallmark Channel crap. I find it quite hard to coax interesting fiction out of the tuned models, and I don't see studios overnight learning how to prompt any better...

(If you want good fiction or poetry, davinci-002 is what I like to use, although it poses problems of its own.)

Imagine writing a stream of bullet-points with no care taken for proper sentence structure and have the LLM convert them to properly written paragraphs of what you were trying to convey. And be able to determine if you have conflicts between bullet-points for you to fix.

You may also be able to provide bullet-points on different scales/modes. Such as the overall setting of a story, specific dialogue between characters, character personalities, character views of each other, plot progression.

Suddenly anyone who can have coherent thoughts (which is not a low bar) and some imagination will be able to synthesize what is in their mind onto paper in well written form.

I bet GPT-4 and others can easily make an early cut movie. Nothing great, but just a collection of scenes that someone who actually has some taste can piece together into a completed work. That sort of work is obviously scary for the industry where you have writers rooms developing this sort of stuff. Imagine it going from all the animators and writers required to write the Simpsons every week into just an automatic script where Matt Groening does some quick editing before it goes out. That would mean a huge reduction of jobs especially at the early career level. On top of that, how would you even become another Matt Groening caliber producer in that environment? The entire context of how the senior talent got their talent has been made obsolete, so once they are gone there's no one to fill the shoes quite the same way.

First, the literal wording of this is also means Grammarly and translation applications cannot be used.

Second, the enforcement of this is very unclear. What happens if a film is halfway through production, and then gasp it turns out a writer rewrote the screenplay using AI? You can't use AI as source material, so arguably any use of that script or any work based off that script is "using AI as source material". Does the production just have to be cancelled? There's a problem here not only with proving that a writer DID use an AI, but proving that they DIDN'T and thus a production doesn't need to be burned to the ground.

Third, how far do you take this? AI even before ChatGPT and LLM's is already broadly prevalent in writing, so what exactly establishes the pedigree of something as clean from the influence of vile AI?

>It’s very easy to imagine a situation in which a studio uses AI to generate ideas or drafts, claims those ideas are ‘source material,’ and hires a writer to polish it up for a lower rate.

It's even easier to imagine a studio will simply continue hiring writers and buying drafts from writers using AIs. How can you possibly prove an idea came from an AI and not a person in the first place?

>can’t be used as source material or to train AI.

How many studios are actually going to train their own AIs? What stops a studio or contractor from using an AI trained on such source material? Wouldn't it be more prudent to ban the use of AIs which were trained on WGA materials?

This entire proposal seems to be based on the populist sentiment about creatives being the victims of both studios and AIs, so obviously we need to stop the studios from using AIs by banning AI usage in film productions. The problem is that the writers are the users of AIs and they're the only ones who know and can prove they used an AI in practice.

I'm sure it's good politics because it leads to headlines like "GPT-4 Can’t Replace Striking TV Writers, but Studios Are Going to Try".

It must be like a dream come true!

We’ll need to remember to sue first time they try.