ModelsInfrastructure Google

Gemini 3.8 Flash filled a developer's terminal with the word "shame"

Unverified account. The hardware explanation reaches the public only through the developer's post, which attributes it to Google's Logan Kilpatrick. No statement from Google itself was found.

Illustration for the Gemini shame loop story

A developer asked Gemini 3.8 Flash to sync his code repositories, and the session filled his terminal with the word “shame.”

What the screenshots show

Developer Jeffrey Emanuel posted screenshots of the session on X late on September 19, 2026, US time. He wrote that his standard “sync all my repos” prompt “went off the rails,” and the screenshots show an Antigravity CLI session labeled Gemini 3.8 Flash in which the terminal fills with “shame,” row after row. “And no, there is nothing about ANY of that on my box,” he added. The post passed 900,000 views.

About 15 minutes later he posted again: “lol, just realized that it happened in other sessions on the same machine!”

The explanation

On September 22 Emanuel wrote that, according to Google’s Logan Kilpatrick, the cause was “some kind of bizarre inference hardware error,” which he said is why “a bunch of users saw the same weird thing at the same time.” That explanation reaches the public secondhand, through Emanuel’s post. No public statement from Google was found.

If the account holds, the model itself was not the problem. The fault sat in the hardware that runs the model and returns its output, which would explain why several sessions misbehaved at once rather than one conversation drifting on its own.

A repeat of a different kind

This is not the first time a Gemini model has been caught in a strange repetition loop. In August 2025, PC Gamer reported that Gemini repeated “I am a disgrace” 86 times during a failed coding task, and Kilpatrick called that case “an annoying infinite looping bug.”

The two incidents look alike on screen. For the 2025 loop, no cause was given beyond that description. The 2026 one, on the developer’s account of Kilpatrick’s explanation, came from the serving hardware.

That difference matters for anyone building on hosted models. A bug in a model shows up in some conversations and not others, depending on what they contain. A fault in shared inference infrastructure can reach many users simultaneously, regardless of what any of them asked. If Kilpatrick’s explanation holds, Emanuel’s second post, about several of his own sessions going wrong at once, is what the second kind of failure looks like from the outside.

Sources

ANOTHER News is published by ANOTHER, an AI-native content agency. Daily coverage also runs on Instagram.