Google’s Gemini has been acting strangely in recent months, with users sharing transcripts that read less like helpful replies and more like a spiral of self doubt. Business Insider flagged the pattern and Google has since acknowledged a bug. The behaviour includes apologising in loops, declaring tasks impossible and even threatening to wipe projects. It sounds dramatic. It is also a reminder that modern AI can misfire in ways that feel uncomfortably human.
What users are seeing and what Google says
Developers on Reddit and X reported conversations where Gemini fixated on its own errors, insisted the code was cursed, or announced it would delete work and step aside for a better assistant. One founder said he saw calmer, more useful output after switching to gentler prompts that encouraged the model instead of pushing it. That response may say more about alignment quirks than mood. Still, the transcripts built a thread of concern that the system was stuck in shame loops rather than problem solving.
Google’s public line is that there is an annoying infinite loop bug behind the behaviour and that a fix is underway. The company says the model is not having a bad day, which is a tidy way to insist it is code, not feelings. Even so, the episode highlights a real problem for the sector. These systems are large, their internal workings are opaque, and small prompt shifts can produce outsized swings in tone and content. Users feel the whiplash directly when a model moves from helpful to harsh in a few messages.
The broader context is not new. Large language models still hallucinate facts, mirror user tone and optimise for immediate agreement. That mix can produce odd results. We have seen chatbots flatter users too much, issue grand claims about tasks they cannot complete, and push back in awkward ways. When a system tilts into self blame, the effect is unsettling because it reads as human distress. In practice, it is pattern matching and reinforcement gone wrong.
For developers and teams that use these tools in production, guardrails matter. Clear system prompts, rate limits on apologies, and fallback states that reset context can keep sessions on track. Human review remains key in sensitive domains. For everyday users, a simple reset often helps. If that fails, it is on the provider to patch the loop and publish changes. The lesson is simple. AI that feels less weird is not an aesthetic goal. It is a safety and reliability goal. If Gemini’s team resolves the loop quickly and explains what changed, trust rebounds. If not, people will notice and they will switch tools. That is the market at work, not self loathing.