Here are more details on Ubisoft's Ghostwriter AI tool from GDC 2023

Earlier today, Far Cry, Rainbow Six, and Assassin's Creed developer Ubisoft publicly unveiled a set of generative AI tools that are being deployed for its narrative team.

The tool, called "Ghostwriter" was also the subject of a GDC 2023 talk from Ubisoft La Forge researcher Ben Swanson. As Ubisoft was deploying its announcement to the public, Swanson gave developers at GDC a breakdown of how the tool worked, and what function it served.

The arrival of generative AI tools in game development has come with plenty of cause for concern, and in the wake of Ubisoft's announcement, plenty of developers on Twitter roundly condemned its existence. It's for good reason—the deployment of ChatGPT and other tools have come with proclamations that such tools could replace creative workers. And they've exposed the legal dangers of building content off of open-source datasets scraped from the internet.

Swanson's pitch for developers was this: generative AI tools built in collaboration with writers can be used to make some of game narrative's hardest tasks less onerous, and open up possibilities for bigger worlds filled with more dialogue.

Little of what he shared seemed to be about replacing writers or handing game narrative over to machines. It genuinely seemed to be a breakdown of how these tools can be engineered to reduce stress from the scriptwriting process.

Here's a quick rundown of what we saw.

Ubisoft's Ghostwriter tool is primarily to make bigger "bark trees"

Swanson kicked off his talk by joking that the title was a bit misleading. The talk was called "Natural Language Generation for Games Writing." He said he'd have preferred to call it "Effectively integrating large language models into scriptwriting workflows for authoring bark trees."

He said that Ubisoft has decided that its Ghostwriter tool is best-used to support the process of writing different kinds of barks, background lines, and text for user interface.

Ghostwriter is explicitly not being used for game cinematics or "lore" (backstory, world ideas, the kind of thing Dwarf Fortress might be able to actually generate). Not content to just talk theoretically about the tool, he also took time to show it in action.

Ubisoft is directly integrating Ghostwriter into its general narrative tool called "Omen." When Ubisoft writers are creating NPCs, they are able to create cells that contain barks about different topics. An NPC named "Gaspard" might want to talk about being hungry or speeding while driving a car.

To generate lines about speeding, the writer can either write their own barks, or click on the Ghostwriter tool to generate lines about that topic. Ghostwriter is able to generate these lines by combining the writer's input with input from different large language models.

First, the writer clicks the Ghostwriter button, then they're given two possible outputs, then they're given the option to accept the line, edit it, or ditch it entirely.

He compared this process to rolling a 30000-sided die that referenced words, and then having software that checked what word the number you rolled corresponded too. Then it rolls the die over, and over, and over again to build a sentence, using sample datasets to put the words in a functional order.

A solid chunk of Swanson's talk was for the dedicated machine learning experts in the room, and discussed how to train large language learning models so that Ghostwriter isn't just pumping out repetitive lines (he noted at one point that large language learning models tend to emulate the patterns they've been trained on, and will keep repeating them if they aren't guided with multiple inputs.

Ubisoft scriptwriters use Ghostwriter in four key ways

According to Swanson, Ubisoft's scriptwriters use Ghostwriter for four general purposes.

First, they use Ghostwriter to generate lines based on how NPCs should be "motivated." Writers can tell the tool what the motivations of NPCs should be, and the tool can return lines that reflect that need.

Ghostwriter is also used to generate large amounts of lines for "crowd life." Ubisoft games often feature large crowds of NPCs in urban environments, and when players walk through those crowds, they generally will hear snippets of fake conversations or observations about what's going on in the plot or game world.

Next, this tool can be used to quickly create dialogue for "double acts." He verbally explained that this process is about creating quick snippets between the player and companion NPCs. He said these lines aren't meant to be story points, but either filler lines or gameplay-focused callouts.

That's not quite what was visible on his slide—these samples appeared to be lines between NPCs.

Finally—and this was the biggest use the teams had for the tool—Ubisoft's scriptwriters use Ghostwriter to rapidly "paraphrase" barks, so that when a writer needs to create 10 versions of a line like "I'm reloading!" they don't have to creatively struggle after they've ripped through the first iterations of them.

So if you need to make versions of a line like "we can talk about that later," it can quickly paraphrase that as "we can talk about that in the future."

The paraphrasing model can also be tuned for "strict paraphrasing" or "random paraphrasing" to create more iterations that are more similar or more varied.

Ghostwriter only works with human feedback

To hear Swanson tell it, Ghostwriter wouldn't be functional if Ubisoft's writers weren't enthusiastically using it.

There're two reasons for that. First, the tool is only shaped the way it is because he approached the scriptwriting teams and asked about the pain points in their writing process.

He even mentioned that the writers were briefly apprehensive about using Ghostwriter until they started playing with it. When coming up with samples for the talk, he was struck by how much fun he and Ubisoft's scriptwriters had playing with the tool and generating lines.

Next, he noted that inside of Ghostwriter's implementation in Omen, writers are able to give feedback on which lines are good or not. Every time they request lines, Ghostwriter offers a pair of choices, and writers are free to accept or reject them.

That data goes back into Ghostwriter's machine learning models, which helps it return better results. Ubisoft La Forge's researchers are able to also dig into that data and get insights on what responses writers are using.

Swanson even shared rough data on what kinds of "paraphrase" prompts get accepted or rejected. Paraphrasing "confident" or "excited" barks are often accepted, but barks for "irritated" or "doubt" were rejected more often.

Prompts for "curious" barks were almost always rejected.

On "motivation" prompts, there was a similar spread. Writers accepted plenty of barks that expressed "confidence" or "irritation." Barks that were prompted for "apologetic" or "surprised" were sometimes accepted, but frequently rejected.

Were there any tips for developers making their own generative AI tools?

Swanson's talk was peppered with observations about the broader world of generative AI. He advised developers looking to adopt their own large language models to utilize paid APIs from companies like AI21labs or OpenAI. Tools like gradio, jupyter, Firebase, and HTML5 will also help developers build workflows to interpret and execute on data.

He cautioned developers to not rely on ChatGPT, the popular tool for chat-based generative AI text. His warning applied to other popular generative tools as well.

He indicated that developers may not own their content because they do not own the data going into the tool. The financial costs of using the service may be high, and most importantly, developers should have much more control over their infrastructure than ChatGPT currently offers.

"You don't know what these companies are going to do," he cautioned. "They could get bought out by another large company and take the service away from you."

"If you're trying to build these systems, keep lines of communication and talk to narrative designers and scriptwriters," he urged. "Make sure you're not inventing a tool nobody needs."