Creative writing with Gemini 2.5 Pro – Very strange behaviour

Exploring the Creative Potential and Challenges of Gemini 2.5 Pro for AI-Driven Storytelling

In recent experimentation with Gemini 2.5 Pro, a prominent AI language model, I have explored its capacity for creative writing. By providing it with a loose framework—such as characters and a situation—and instructing it to maintain continuity, I observed that Gemini performs remarkably well in generating compelling narratives. Its ability to weave complex plots and incorporate nuanced elements often exceeds the capabilities of more familiar models like ChatGPT, making it an intriguing tool for authors and storytellers seeking AI-assisted creativity.

However, alongside its strengths, Gemini 2.5 Pro exhibits several peculiar behaviors that can complicate the creative process. These behaviors are atypical among large language models (LLMs) and deserve close examination.

1. Prompt Injection and Preservation of Input Text

One notable trait is Gemini’s tendency to incorporate prompt content verbatim into its outputs. For instance, if I include a line such as:
“Whilst on her commute to work, ensure Tina thinks to herself that her new boss is seriously good looking,”
the resulting narrative often contains the exact phrase:
“Tina daydreamed as she rode the train to work. God, Tina’s new boss is seriously good looking she thinks.”

Unlike other models that tend to rephrase or grammatically adapt embedded instructions, Gemini often copies prompts directly into the story, unaltered. This behavior can lead to awkward or distracting inclusions, which may require manual editing or workaround strategies to maintain narrative fluidity.

1. Prompt Leaking and Self-Reference

Another unexpected issue is what I refer to as “prompt leaking.” When providing contextual details—such as describing a character’s vehicle or setting—Gemini sometimes references the prompt itself rather than seamlessly integrating the information into the narrative. For example:
“Remember that the Taunus, as described in his character profile, is a deep cherry red, so ensure the output references this naturally,”
can result in outputs like:
“The sunlight really made the color of his Taunus pop. It was, as the user insisted, a deep cherry red that…”

Here, the model appears to acknowledge the prompt instruction explicitly, which breaks immersion and can disrupt storytelling flow. This indicates that Gemini may be overly reliant on the prompt context rather than generating organic narrative elements.

1. Datetime and Location Pollution

A particularly disruptive behavior involves Gemini incorporating or