OpenAI published details on 16 August 2017 about its Dota 2 bot's performance limitations following a demonstration at The International 2017 tournament [source]. The system, which had defeated professional players in controlled 1v1 matches days earlier, exhibited significant behavioural constraints that were not initially disclosed.

The bot operated under a restricted ruleset that excluded most of Dota 2's hero roster, item combinations, and core gameplay mechanics. OpenAI confirmed the agent could only play Shadow Fiend in mid-lane scenarios with a subset of items available. Summons, invisibility, illusions, and other standard game elements were disabled. The model also required limitations on courier usage and ward vision.

According to the post, the system trained for approximately two weeks of real-time gameplay using self-play reinforcement learning. OpenAI stated the bot learned "entirely from self-play" without human demonstrations or hardcoded strategies. However, the reward function and game-state representation were manually engineered by researchers.

The disclosure followed public questions about the scope of the bot's capabilities after its exhibition matches. Professional players and community observers had noted the constrained environment differed substantially from competitive Dota 2. OpenAI acknowledged these limitations were necessary for the current system architecture.

The post indicated OpenAI planned to expand the bot's capabilities toward 5v5 matches with fewer restrictions. No timeline was provided for when the system might operate under standard competitive rulesets. The company framed the work as progress toward general-purpose AI systems that can handle complex strategic environments.

The constraints represented a significant gap between the demonstrated capabilities and the initial framing of the achievement as defeating professional Dota 2 players.