Scaling Game Content Production with LLMs and Data Augmentation

ubisoft · Media & Entertainment · 2023 · source Agentic: no · promotional bias: high · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: game writing data augmentation

Ubisoft needed to scale diverse NPC dialogue variations and fine-tuning data for internal paraphrasing models without forcing writers to manually generate tedious bark-tree variants.

Constraints:

• Narrative-driven games need many diverse bark-tree variations tied to character motivation.
• Internal paraphrasing models needed fine-tuning data that did not already exist.
• Writers needed a creator-first workflow where they curate and edit rather than accept low-quality replacement text.

Architecture

Ubisoft uses AI21's API and LLM outputs as data augmentation for internal writing models. Writers provide motivations or seed examples, AI21 generates candidate training data in the needed format, writers thumbs-up/edit or reject outputs, and the curated augmented dataset fine-tunes internal paraphrasing/NPC generation tools embedded into writer workflows.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  A["Writer motivation prompts"] --> B["AI21 generation API"]
  B --> C["Writer curation workflow"]
  C --> D["Fine-tuning dataset"]
  D --> E["Internal paraphrasing models"]
  E --> F["Writing tools"]

Components

Data flow

Writers define dialogue motivations and desired formats, AI21 suggests candidate variations, writers curate and edit the generated data, augmented datasets train internal paraphrasing and NPC-generation models, and those models are embedded into Ubisoft writing tools with pairwise comparisons and human feedback.

Why this architecture

Ubisoft uses AI21 for data augmentation rather than directly replacing writers: the model creates candidate training data, writers curate it, and internal tools remain writer-in-the-loop and fine-tuned to Ubisoft's creative needs.

Models used: AI21 vanilla model · Ubisoft internal paraphrasing models · NPC generator models

Scale

• Bark trees may need 5-10 ways to express a motivation.
  “painstakingly brainstorm 5-10 different ways to express motivation”article, stated
• AI21 let Ubisoft generate thousands of inputs for curation.
  “With AI21, Ubisoft could generate thousands of inputs and hand-pick the best ones to edit using their fine tuning datasets.”article, stated
• Some dialogue contexts need 10-15 variations.
  “there need to be 10-15 different ways for a townsperson to “get out of here””article, stated

Alternatives considered (roads not taken)

Manual creation of fine-tuning data

Rejected because: It would have taken too long and was tedious enough that Ubisoft was not doing it at all.

Simple paraphrasing

Rejected because: Bark-tree variation must pivot through character motivation rather than paraphrase one sentence.

Chatbot-style LLM in the game-writing process

Rejected because: Ubisoft preferred writer-in-the-loop data augmentation and fine-tuning over a lower-quality replacement.

Patterns

synthetic-data-generation (data-generation)

AI21 generates candidate examples in the correct form for Ubisoft fine-tuning datasets.

human-in-the-loop-gate (creative-governance)

Writers thumbs-up, edit, reject, and compare generated outputs before they become training data or tool behavior.

fine-tuning-adaptation (model-development)

Curated AI21-generated examples become fine-tuning data for Ubisoft internal paraphrasing models.

Tradeoffs

Creative workflow

Chose writer-in-the-loop generation and curation over fully automated dialogue generation.

Rationale: Writers can edit and approve generated data while focusing creative effort on polishing instead of blank-page generation.

Model approach

Chose AI21 vanilla model plus Ubisoft internal fine-tuning over heavy customization of the external LLM first.

Rationale: AI21's model fit data augmentation needs out of the box, while Ubisoft kept internal models for production writing tools.

Gotchas (what bit them)

Paraphrasing alone is insufficient for bark trees.

Root cause: NPC lines need variations that share motivation, not superficial sentence rewrites.

Resolution: Generate from the underlying motivation and let writers curate outputs.

Data augmentation costs can rise with repeated prompt experiments.

Root cause: Fine-tuning output quality required experimenting with multiple prompts.

Resolution: AI21 customized pricing to charge only for output tokens for Ubisoft's data-augmentation use.

Creative AI tools can become low-quality replacements if writers lose control.

Root cause: Game writing depends on human creative direction and tone.

Resolution: Keep the workflow writer-in-the-loop and use AI as data augmentation/inspiration.

Evaluation

Ubisoft evaluates the workflow through writer curation, pairwise output comparisons, human feedback, diversity/distinctiveness of outputs, and practical usefulness in writing tools.

Metrics: writer approval/edit decisions · pairwise output comparisons · training-data volume · output diversity · writer productivity

Outcomes

• Ubisoft could generate thousands of candidate inputs and curate the best for fine-tuning.
• The augmented dataset made outputs more diverse and distinctive.
• Work that was previously too tedious to do manually became a rapid scaling workflow for writers.

The case is an AI21 customer story and does not provide numeric production throughput beyond generated examples and variation counts.