SteamGPT Leaks from Steam Client: Here’s What Valve is Building with Artificial Intelligence
In the files of the Steam client updated on April 7, 2026, strings of code have appeared suggesting that Valve is working on something called SteamGPT. The details were uncovered by Gabe Follower, a content creator specializing in datamining on Valve, who posted the relevant fragments on X. The strings were already visible to anyone keeping an eye on the SteamTracking project on GitHub, which automatically tracks changes made with each client update.
It seems that Valve is working on a SteamGPT feature that will apparently deal with Steam support issues and is somehow connected to Trust Score and CS2 anti-cheat? 
From what is understood, this is not a consumer function: SteamGPT seems clearly intended for internal use, resembling an AI system meant to assist Valve employees in managing two distinct yet connected operational areas. The first is customer support. The second, more interesting, pertains to an approach to anti-cheat mechanisms similar to Counter-Strike 2’s Trust Score.
What the Code Says
The fragments extracted from the client refer to a task queue architecture that includes task creation, response management, labeling jobs, test results, and data for fine-tuning a model. The code suggests that it is something more complex than a simple generic chatbot linked to Steam’s FAQ; instead, it appears to have the characteristics of a structured classification and data analysis system, complete with a continuous training pipeline.
A specific module called SteamGPTSummary collects and synthesizes account-level data: profile information, Steam Guard status, security history, country of registration, VAC status, associated phone number details, fraud flags, and playtime. The apparent goal is to provide employees with a readable and immediate summary of an account’s “health,” rather than forcing them to manually sift through separate logs and databases.
There are also references to Trust_GetTrustScoreInternal, CSbot, and player_evaluation. The latter variable is particularly significant: player evaluation is an anti-cheat mechanism that assesses the player’s in-game behavior and does not just search for cheat software installed on the system. This means that SteamGPT could have direct access to the behavioral pipeline already used by VACnet in CS2. Among the strings, there is also SteamGPTRenderFarm, which suggests operation on a dedicated computational infrastructure rather than something running in the background on servers already occupied by other tasks.
The CS2 Node and the Trust Factor
The Counter-Strike 2 community has been complaining for years about the quality of anti-cheat with VAC, which is notoriously slow in detecting new cheats, and VACnet (the active ML system) that has never fully met expectations. The integration of SteamGPT with Trust Factor Matchmaking could change the paradigm: instead of a system that detects only suspicious software, there would be an AI layer that correlates multiple signals such as in-game behavior, account history, purchase patterns, and device data to build a more articulated risk profile. And, probably, at a “Steam system” level rather than a single title level.
However, the leaked code does not show SteamGPT autonomously issuing bans nor replacing VAC as the primary system, indicating that, at least in the current stage of development, the AI functions as a decision-support tool for human reviewers without making decisions on its own. A prudent choice, given the risks of false positives on particularly skilled legitimate players.
There would also be an evidence log linked to specific matches, suggesting that the system may attach contextual evidence (fragments of gameplay sessions, recorded suspicious events) to the reports processed by the AI to compile structured documentation for those who must make the final decision.
Valve and AI: A Quiet Entrance
Valve has so far been absent from the AI marketing that has characterized almost every other major tech company in the last two years. No announcements, no keynote on artificial intelligence, no press releases with “AI-powered” in the title. SteamGPT, however, seems to represent the first AI tool developed internally by Valve as a company.
Last year, Valve President Gabe Newell stated: “AI is essentially going to be a cheat code for people who want to take advantage of it.” This statement lent itself to ambivalent interpretations but, in retrospect, sounds like a description of what Valve is building: an internal tool that leverages AI before malicious actors do. Meanwhile, the platform already hosts over 110,000 games with generative AI content, amounting to about 7% of the entire Steam catalog, indicating that Valve has already come to terms with this technology on the editorial front.
No Confirmation, Many Questions
Valve has not confirmed the existence of SteamGPT, and as is often the case in these matters, the code found could belong to prototypes never intended for release. It is not uncommon for Valve to accumulate experimental work in client files without ever bringing it to production. However, the discovery is detailed enough to indicate at least a serious level of development: a pipeline architecture, a data aggregation module, references to dedicated computational infrastructure, and mechanisms akin to existing systems like VAC and Trust Score are not the kind of things built in one’s spare time or as simple proof-of-concept.
The questions that remain open primarily concern privacy and human oversight, given that SteamGPTSummary aggregates a significant amount of sensitive data per user: how many people at Valve will have access to these summaries? How often will AI-based decisions be reviewed by a human operator? And most importantly: if the system issues an erroneous judgment on an account, how easy will it be for the user to contest it? These are questions Valve will need to be ready to answer if and when SteamGPT emerges from the lab.