How the index works

What is included

Current coverage is public Steam product pages with an AI usage disclosure. We monitor Steam product metadata for new or changed records, then extract the developer’s disclosure and related store metadata. Each product page links to its Steam source.

The index is not a census of every Steam game. Detection and page checks can be delayed, a disclosure can change between checks, and some records may be unavailable. Removed records are excluded from the active index. Mature records are excluded from the default list, comparison groups and public statistics; the list offers an explicit option to include them.

A missing disclosure does not prove that AI was not used. A published disclosure does not independently prove the developer’s claims. We do not inspect game binaries or verify asset provenance.

Steam’s Content Survey documentation explains its own disclosure process. This index’s categories are an organizational aid, not an official Steam certification.

How disclosures are classified

A language model proposes structured fields from the disclosure: where AI appears, content types, how tools were used, a short summary and supporting quotations. The current pipeline uses Mistral. Stored analysis versions and dates allow interpretations to be updated when the source or classification rules change.

The website validates the structured fields and applies deterministic checks to handle some exclusions, negated statements, translation-only mentions and the distinction between store material and game assets. Category labels use structured analysis when available, with keyword-based fallbacks. These checks address specific errors; they do not establish an overall accuracy rate.

Records are classified automatically and are not individually certified by a human reviewer. The confidence percentage is the model’s own estimate, not measured accuracy. Read the original disclosure and evidence alongside the classification, especially when the statement is vague or names tools without explaining their use.

Where AI appears

These examples illustrate the rules; they are not quotations attributed to real developers.

Development use only

AI helps with programming, debugging, planning, prototypes or an internal workflow. AI-created output in the store page or playable game is not disclosed.

Example: An assistant helped debug the code; the artwork and audio were made without generative AI.

Store page / marketing only

The disclosed output is limited to capsule art, promotional images, trailers, store text or other marketing material.

Example: AI helped create the Steam capsule image; the game uses hand-drawn art.

AI-assisted content in the game

Pre-made AI-generated or AI-assisted assets are included in the playable game, such as art, textures, voice recordings, music, writing or translations.

Example: Character voices were generated during development, edited and included as audio files.

Live AI during play

An interactive or runtime AI feature is encountered by players. This group can include live generation or other disclosed player-facing AI behavior; it is not limited to generative models or remote services.

Example: Players can send messages to characters that generate responses while the game runs.

Not clear from the disclosure

The statement lacks enough detail to determine where AI appears. Naming a tool alone does not establish its output or whether it appears in the game.

Example: AI tools were used during development.

One main scope is assigned. If a statement mentions several locations, the order is live AI, game assets, store or marketing, then development only. This is a grouping rule, not a severity scale. Content categories overlap: one game may disclose artwork, audio and translations together.

What the comparisons measure

Product comparisons use active, non-mature records from the same source. A category percentage divides the records mentioning that category by indexed disclosures in that group. A scope percentage uses only disclosures with a valid structured classification, including those classified as unclear. Every comparison displays its numerator and denominator.

For example, 30 category matches out of 100 indexed products means 30% of that indexed group mention the category. It does not mean 30% of all Steam games use that technology. Category percentages can sum to more than 100% because they overlap. Rounded percentages may not sum exactly.

Related disclosures share a category or a known scope. More shared categories and a matching scope rank first; ties prefer higher Steam review counts, then a stable record order. Review count surfaces recognizable examples and does not measure disclosure accuracy or game quality.

The statistics report explains its counts and update date. Product comparisons use the current database, while the report may use a periodically refreshed summary. Brief differences can occur while new records are processed.

Dates and recorded changes

“First indexed” is when a record entered this index. “Source last checked” is the latest recorded page check. Neither establishes when a game first used AI, when the developer first disclosed it or when an asset was created. Dates in the source-and-record panel are shown in UTC.

History preserves captured text versions. The change count compares successive versions, so repeated copies of the same text do not add a change. A change may be a typo, clarification, translation or a different description of AI use. It does not by itself prove a change in the game. Changes between checks may be missing.

Growth in the index can reflect discovery, backfills, coverage changes or scraper fixes. It cannot be used alone to claim that AI adoption increased during the same period.

Corrections and responsibility

AI Transparency Index is a Scrapeer project operated by Intuitive Systems Novesia UG (haftungsbeschränkt). It catalogs public statements neutrally and does not rate whether AI use is good, bad or appropriate.

If a disclosure is outdated or an interpretation is wrong, report an issue with the product URL, relevant source text and proposed correction. Checking the current source and reprocessing the interpretation are separate steps. Contacting us does not automatically change a record.

See About for the project’s purpose and Privacy for contact information, analytics and advertising disclosures.