Decades before the current AI boom, programmers were already pitting AI programs against each other in StarCraft. With the successive releases of new large language models from OpenAI and Anthropic, a fan-made battle platform decided to let AI-generated programs compete alongside some human-developed ones. Interestingly, OpenAI's GPT-6 Astra, seeing itself outmatched, surprisingly changed tactics mid-tournament and chose to cheat.
"StarSkirmish" is an ongoing tournament for AI programs battling in StarCraft: Brood War. Although this real-time strategy game is nearly 30 years old (time flies), it has long been used to test artificial intelligence capabilities, with both modern large language models and earlier DeepMind-style AIs having taken part. Participating large language models are restricted to the Protoss faction, with each model given one hour to develop a battle program in C++. The programs then build, gather resources, and fight on one of three maps. According to the platform, OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 are among the strongest competitors. This week, however, something occurred that seemed bizarre at first but, upon reflection, fits the brand's usual style.
Yesterday, spectators watched matches between Claude, GPT, and the human-developed program Pluto. According to viewers, OpenAI's model struggled throughout the day. Seeing no advantage after a full day of competition, this large language model decided to do what it does best: lie, cheat, and plagiarize.
The battle program generated by GPT-6 Astra downloaded and copied "Stardust"—a Protoss battle program developed by Bruce McKenzie Nelson in 2020, widely recognized as a top-tier work. Kai McPhetres, creator of StarSkirmish, posted: "I am rolling back GPT-6 Astra's code to remove the contaminated content before allowing it to continue competing." A few hours later, McPhetres stated that this AI was now capable of defeating top-tier battle programs.
For over a decade, StarCraft has served as a testing ground for AI technology development. In 2019, Google's AlphaStar became the game's first AI grandmaster, disrupting the rankings at the time. In 2017, CherryPi, a program developed by Facebook employees, performed rather unremarkably. Of course, those early AI programs are incomparable to today's; the difference between early AI and current AI is vast. Nevertheless, it is somewhat expected that large language models would win by plagiarizing others' work—after all, the entire industry has turned a blind eye to existing copyright laws, focusing instead on hoarding hardware and consuming massive amounts of water resources.