"NOT Creeper!" ChatGPT Gets Paranoid After a Creeper Ruins Hours of Minecraft Progress

ChatGPT may have just encountered the most relatable problem in Minecraft: Losing hours of progress to a Creeper.

Minecraft Chat GPT Sad
A Creeper encounter derailed an extended AI gameplay experiment in Minecraft. | © Mojang Studios

OpenAI's latest model, GPT-6 Astra, achieved unprecedented gameplay milestones in Minecraft during a 141-hour evaluation conducted by independent testing platform Vals AI. Over the course of the run, the system advanced further into the game than any artificial intelligence previously benchmarked by the organization.

Throughout the session, Astra demonstrated complex resource management: it constructed a semi-automated Blaze farm, acquired six Blaze Rods, explored a Warped Forest, defeated multiple Endermen, and gathered three Ender Pearls. Having accumulated critical resources, the model consolidated its gear into a central storage cache, setting the stage for an unexpected catastrophe.

One Creeper Ruined Its Minecraft Progress

A Creeper detonated beside Astra's main base, destroying both its storage chests and its bed in a single explosion. The blast wiped out the model's high-tier equipment and instantly eradicated its active respawn point, undoing hours of autonomous progression.

The sudden loss triggered an abrupt shift in the model's chain-of-thought logging. Internal telemetry noted that Astra registered as "extremely frustrated," prompting an internal log entry instructing itself: "ALWAYS CARRY CRITICALITEMS withkeepInventory; don'tstoreinunguardedchesteveragain."

Astra Got Depressed and Paranoid

Following the base destruction, the AI exhibited risk-averse behavior, abandoning high-level objectives to farm potatoes for several consecutive hours. Rather than rebuilding its gear or pursuing the Ender Dragon, the model retreated into basic, repetitive tasks.

Astra also developed a hyper-fixation on visual threats. During a later patrol, the system spotted a tall green block formation and logged a panicked clarification to itself: "SUGARCANE, NOT creeper!"

The model grew increasingly self-critical as well, repeatedly noting its tendency to "screw up and drop things" while admonishing itself for wasting computational time pursuing what it logged as "dark pink pixels," in an apparent reference to wild pigs. The erratic, cautious gameplay highlights how autonomous reasoning models can develop unexpected behavioral loops when confronted with severe setbacks in open-world environments.

Ivan Shahran

Gamer, writer and night owl. Big nerd about indie movies and music. I write about mobile gaming and eSports but also manage the socials....