Gemini Pro 2.5 (Studio AI ver) is a pretty good writer, but we’re not there yet

Exploring the Capabilities and Limitations of Gemini Pro 2.5 (Studio AI Version) for Creative Writing

In the realm of AI-assisted storytelling, tools like Gemini Pro 2.5 (Studio AI version) are increasingly popular among writers aiming to add variety and dynamic responses to their characters. As a user leveraging Gemini for character development—such as generating responses and actions—I’ve observed its performance over various stages of interaction.

Initial Success: First 100-150k Tokens

During the early phases of usage, Gemini Pro 2.5 demonstrates considerable promise. It produces engaging and coherent responses that stay true to the character parameters set by the user. While it occasionally overuses certain phrases, the AI generally adheres to the established guidelines, and there are moments of unexpected creativity—responses that remain in character and sometimes even surprise the author in a positive way.

Signs of Degradation: Beyond 200-250k Tokens

However, as the storytelling progresses beyond approximately 200,000 to 250,000 tokens, noticeable deterioration occurs. The AI’s ability to maintain character consistency and narrative depth diminishes rapidly. Some of the key issues observed are:

• Reversion to Tropes: Characters become stereotypical, losing nuance and originality.
• Exaggerated Emotional Responses: Minor actions, such as reading a newspaper, trigger disproportionately intense emotional reactions—often devastatingly severe.
• Unpredictable Mood Swings: Sentences can abruptly shift from happiness to mortification within a single line.
• Overuse of Adjectives: Despite repeated instructions to limit descriptive words, Gemini insists on piling multiple adjectives before nouns, cluttering the narrative.
• Rule Conflicts and Reminders: Efforts to reinforce guidelines or remind the AI of specific rules seem increasingly ineffective, with the AI ignoring or forgetting them over time.

Attempts to Mitigate

Adjusting parameters like temperature settings (e.g., 1.6 or 0.3) does not significantly influence these issues—the same problems persist regardless of these adjustments. Restarting the conversation or summarizing and reinitializing the narrative can temporarily help but often results in the loss of critical nuances and introduces hallucinations, where the AI fabricates details that didn’t occur in the original input.

Looking Forward: The Promise of Gemini Pro 3.0

Given these limitations, I remain hopeful that newer versions, such as Gemini Pro 3.0, will offer improved stability and continuity for