AI Usage at a Glance
Mar 21, 2018
OtherPractice documented: VACnet is Valve's deep learning system that watches how players move and aim in Counter-Strike matches and automatically flags those whose behavior looks statistically impossible for a human. Valve first revealed the system at a game developers conference in 2018. It processes roughly 150,000 matches per day and routes flagged cases to human reviewers rather than issuing bans directly.
Practice DocumentedView practice →Mar 25, 2018
OtherNew evidence: Valve has 1,700 CPUs working non-stop to bust CS:GO cheaters
Evidence AddedView practice →Aug 7, 2018
OtherPractice documented: Valve uses a machine learning model to assign every Counter-Strike player an invisible "trust score" based on their behavior across their entire Steam account history. Players with low scores are quietly matched against each other rather than being banned outright, which limits the impact of cheaters and toxic players on the wider community. The system has been running since November 2017.
Practice DocumentedView practice →Nov 13, 2018
OtherNew evidence: Valve using machine learning and deep learning to catch cheaters on CS:GO (794 words)
Evidence AddedView practice →Mar 15, 2019
ModerationPractice documented: Steam runs an automated system that watches all game review activity across the platform in near real time, looking for sudden unusual spikes that might indicate a coordinated attack — sometimes called a "review bomb." When the system spots something strange, it flags the game for a human team at Valve to investigate. Valve announced and began enforcing this system in March 2019.
Practice DocumentedView practice →Jul 10, 2019
RecommendationPractice documented: Steam uses a machine learning system to recommend games to players based on what they and millions of other users have actually played — not on genre labels or tags. The system launched as an experiment in July 2019 and now powers multiple features across the Steam store, including the front page, the Discovery Queue, and the "Play Next" shelf.
Practice DocumentedView practice →Nov 19, 2019
OtherPractice documented: Valve holds multiple patents describing machine learning applications for its hardware products and multiplayer networking. These cover predicting player input to reduce lag in online games, tracking eye gaze direction in VR headsets, and determining the physical position of a VR headset using sensor data. The patents confirm Valve's machine learning R&D extends well beyond anti-cheat.
Practice DocumentedView practice →Aug 28, 2020
ModerationPractice documented: Every post, comment, screenshot, guide, and Workshop item submitted to Steam passes through an automated content check system before it goes public. If the system is uncertain about a submission, it holds the content with the message "awaiting analysis by our automated content check system" while a human review takes place. This applies to all user-generated content, not just reported items.
Practice DocumentedView practice →Jun 7, 2021
ModerationNew evidence: What's up with "This comment is awaiting analysis by our automated content..." Permanently
Evidence AddedView practice →Apr 9, 2023
RecommendationNew evidence: Valve breaks down how Steam visibility works: “Algorithms react to player interest quickly and automatically”
Evidence AddedView practice →Jan 9, 2024
Creative GenPractice documented: Steam requires every game developer who uses generative AI in their product to label it publicly on the store page, so players know before they buy. Valve introduced this policy in January 2024 and updated it again in January 2026 to clarify which AI uses trigger disclosure and which do not.
Practice DocumentedView practice →Oct 2, 2024
ProductivityPractice documented: A Valve engineer publicly described using ChatGPT like an advanced search engine to find a specific type of matching algorithm he needed for the game Deadlock. He found the Hungarian algorithm this way and it was then built into Deadlock's hero selection matchmaking system. The engineer shared this on social media in October 2024.
Practice DocumentedView practice →Oct 3, 2024
ProductivityNew evidence: Deadlock's new matchmaking algorithm was recommended to Valve by AI: "I found it using ChatGPT"
Evidence AddedView practice →Jul 18, 2025
ModerationNew evidence: AI-generated reviews on Steam are becoming a problem | Copyleaks
Evidence AddedView practice →Sep 15, 2025
OtherPractice documented: VAC Live is Valve's upgraded anti-cheat system for Counter-Strike 2, running machine learning analysis during live matches — not after the fact — so it can detect cheating in real time and cancel an affected match immediately. It replaced the earlier VACnet system when CS2 launched in 2023.
Practice DocumentedView practice →Jan 16, 2026
Creative GenNew evidence: Steam updates AI disclosure form to specify that it's focused on AI-generated content that is 'consumed by players,' not efficiency tools used behind the scenes
Evidence AddedView practice →Valve holds multiple patents describing machine learning applications for its hardware products and multiplayer networking. These cover predicting player input to reduce lag in online games, tracking eye gaze direction in VR headsets, and determining the physical position of a VR headset using sensor data. The patents confirm Valve's machine learning R&D extends well beyond anti-cheat.
US Patent 11,717,748 B2 (filed November 2019, granted August 2023) describes using a trained machine learning model to predict what a player is about to do with their game controller, allowing the system to generate game responses proactively and compensate for the delay caused by network latency — improving the experience in multiplayer games. US Patent 12,321,514 (granted July 2025) covers using machine learning to determine the direction a VR headset user is looking, based on data from quadrant photodetectors. A third application (20250189783, filed February 2025) describes machine learning-based position tracking for VR headset components using angle-sensitive light detectors. Together these patents indicate Valve is developing or has developed machine learning systems for latency compensation, eye tracking, and spatial tracking in its hardware product line.
Valve holds multiple patents describing machine learning applications for its hardware products and multiplayer networking. These cover predicting player input to reduce lag in online games, tracking eye gaze direction in VR headsets, and determining the physical position of a VR headset using sensor data. The patents confirm Valve's machine learning R&D extends well beyond anti-cheat.
Valve uses a machine learning model to assign every Counter-Strike player an invisible "trust score" based on their behavior across their entire Steam account history. Players with low scores are quietly matched against each other rather than being banned outright, which limits the impact of cheaters and toxic players on the wider community. The system has been running since November 2017.
VAC Live is Valve's upgraded anti-cheat system for Counter-Strike 2, running machine learning analysis during live matches — not after the fact — so it can detect cheating in real time and cancel an affected match immediately. It replaced the earlier VACnet system when CS2 launched in 2023.
VACnet is Valve's deep learning system that watches how players move and aim in Counter-Strike matches and automatically flags those whose behavior looks statistically impossible for a human. Valve first revealed the system at a game developers conference in 2018. It processes roughly 150,000 matches per day and routes flagged cases to human reviewers rather than issuing bans directly.
Steam operates a text filtering system in Steam Chat and in games like Counter-Strike 2 and Dota 2, automatically blocking or replacing profanity and slurs. Players can adjust how much filtering is applied to their own experience. The system shipped as Steam Labs Experiment 011 before being integrated into the main platform.
Every post, comment, screenshot, guide, and Workshop item submitted to Steam passes through an automated content check system before it goes public. If the system is uncertain about a submission, it holds the content with the message "awaiting analysis by our automated content check system" while a human review takes place. This applies to all user-generated content, not just reported items.
Steam runs an automated system that watches all game review activity across the platform in near real time, looking for sudden unusual spikes that might indicate a coordinated attack — sometimes called a "review bomb." When the system spots something strange, it flags the game for a human team at Valve to investigate. Valve announced and began enforcing this system in March 2019.
Have evidence about Valve Corporation's AI practices? Submit a report.
Submit a report →AI Trace is free and nonprofit. Support our work
A Valve engineer publicly described using ChatGPT like an advanced search engine to find a specific type of matching algorithm he needed for the game Deadlock. He found the Hungarian algorithm this way and it was then built into Deadlock's hero selection matchmaking system. The engineer shared this on social media in October 2024.
Valve engineer Fletcher Dunn posted publicly that he used ChatGPT to search for a bipartite matching algorithm — a type of problem-solving method that assigns players or items to slots based on preference scores. He described the value of the tool as helping him find technical concepts he did not already know the name for, using it as a way to search by description rather than by keyword. The Hungarian algorithm he discovered through this process was implemented in Deadlock's matchmaking within days. Dunn described this as an "internal only" use of AI as a research tool, consistent with Valve's January 2026 policy update that explicitly excludes AI-powered development tools from Steam's disclosure requirements. This is the only confirmed case of a Valve engineer publicly disclosing the use of a large language model for internal development work.
Steam operates a text filtering system in Steam Chat and in games like Counter-Strike 2 and Dota 2, automatically blocking or replacing profanity and slurs. Players can adjust how much filtering is applied to their own experience. The system shipped as Steam Labs Experiment 011 before being integrated into the main platform.
The Steam Text Filtering system uses curated word lists as its primary mechanism, with players able to set their own filtering sensitivity. Game developers can integrate the same filtering into their titles via the Steamworks API. Valve stated that default word lists would continue to be refined based on user data, indicating ongoing data-driven maintenance rather than a purely static rule set. While the system's core mechanism is list-based rather than a machine learning model, its continuous refinement from usage data and its role in automated content filtering make it relevant to this ledger. The system supports multiple languages.
