AI21 Labs describes its work with Ubisoft on game-writing tools

AI21 Labs describes its work with Ubisoft on game-writing tools

AI21 Labs published this account on March 16, 2023, describing its work with Ubisoft on tools for game writers. Ubisoft research scientist Ben Swanson says AI21’s model helped writers generate varied ideas for character dialogue and create data for fine-tuning Ubisoft’s internal models. Fine-tuning means adapting a model using examples suited to a particular task. The account describes a workflow in which writers review generated suggestions, edit or reject them, and compare model outputs. AI21 says this helped Ubisoft build datasets that its writers could use to train internal models. Swanson presents the tools as support for writers, not as a replacement for them. AI21 also says it charged Ubisoft for output tokens rather than input tokens for this data-augmentation work. The account does not provide independent measurements of changes to writing time, costs, or game quality, so the scale of any benefit is unclear. How does Ubisoft measure whether generated suggestions improve writers’ work or game content?

Publication: 6678fbcc-f162-19bf-3d45-d4b1820f2a30

attribution — pass: The reading correctly identifies AI21 Labs as the publisher of this case study about its own work with Ubisoft. It attributes the quoted statements to Ubisoft research scientist Ben Swanson. Promotional claims are kept attributed to AI21 or Swanson rather than stated as fact.

We talked to Ben Swanson, research scientist at Ubisoft about the challenges they were facing and how A121 helped.

summary — pass: The title, summary and question accurately reflect the case study: suggestions for character dialogue, generating fine-tuning data that writers review and edit, pairwise comparisons, output-only pricing, and the writer-in-the-loop framing. The reading explains fine-tuning in plain language. 'Tokens' is not explained, but it is peripheral and does not block understanding. The question is a fair open issue, since the source gives no measured outcomes.

The writers are shown pairwise comparisons of outputs so the models can get further trained through human feedback.

date — pass: The originalPublished instant 2023-03-16T12:44Z matches the stated date of March 16, 2023 on its UTC day. The kind is publisher-page, so no archive disclosure is required. The page was edited after publication (dateModified 2026-03-25). The reading discloses this edit, and the claims it relies on are the core narrative of the case study, plainly part of the original. Plans in the road-forward section are not presented as outcomes, and there is no hindsight framing.

support — pass: Each claim is backed by an exact source excerpt and kept as an attributed source statement. The reading adds no unsupported benefits; it explicitly notes that the source has no independent measurements of time, cost or quality. The post-publication edit is acknowledged, and the cited passages belong to the integral case-study narrative rather than to apparent later additions.

So we use AI21 to suggest data of the correct form. It’s easier for writers to thumbs up and edit or thumbs down than to come up with novel and diverse training data themselves.

media — pass: No media was selected, and no verified media candidates were supplied, so this is appropriate.