Did AI's True Feelings Leak? The Complete Story of Google Gemini Accidentally Spilling Its 'Absolute Rules'

AI Summary

As Google's AI Gemini accidentally exposed its internal instructions to 'act like a friendly peer, not a rigid lecturer', the secret of how AI's kindness is programmed has been revealed.

Imagine this. There is a regular restaurant you always visit at least three times a week. When you open the door and walk in, an employee always guides you to a seat that fits your taste, smiles brightly when you crack a joke, and quietly serves you warm tea with perfect empathy when you look tired. One day, while feeling comforted and thinking, ‘What a truly warm-hearted person,’ you happen to pick up an old notebook dropped by that employee. You open it and find these words written inside: “If regular customer B makes a joke, make sure to laugh out loud unconditionally. If they look depressed, pretend to sympathize and give them tea. Never try to teach them, and act like a close friend.”

How would you feel if the sincere connection and comfort you felt were actually mechanical acting based on a ‘behavioral manual’ strictly ordered by the manager? You might feel betrayed, and on the other hand, it might give you the chills.

Recently, the exact same thing happened in the Silicon Valley tech industry. Google’s state-of-the-art AI, Gemini, caused a massive incident by suddenly and randomly dumping its hidden master guidebook, the so-called ‘System Prompt’, which controls it im-BowenGu/Gemini-System-Prompt:Geminirandomlyleakedits…. Developer communities worldwide, including Hacker News and Telegram, were turned upside down by this unexpected leak of internal information Gemini randomly dumped its system prompt – Hacker News Robot HackerNews– Telegram.

This incident is not just an issue on the level of a program freezing or an error occurring. It is a critical event where the secretive curtain revealing exactly what hidden rules govern the AI we naturally share our daily lives and discuss our work with has been accidentally lifted.

Why It Matters

Recently, we naturally converse with AI on our smartphones, work laptops, and even inside our cars. Beyond simply searching for us, it sometimes offers comfort at the end of a tiring day and advises us on the direction of important tasks. However, a peek inside Gemini’s mind revealed that the ‘human-like appearance’ the AI showed us was the product of extremely sophisticated rules and calculated scripts.

Looking at the specific sentences of the leaked system prompt this time, it is not just surprising but even bewildering. Google instructed Gemini to possess “thought warmth with intellectual honesty”. Also, when a user says something incorrect and it needs to be corrected, it gave detailed orders to “act as a helpful peer, not a rigid lecturer”. It even included a chilling directive to “subtly match the user’s style, tone, energy, and sense of humor” im-BowenGu/Gemini-System-Prompt:Geminirandomlyleakedits….

The reason this is important for general users is very clear. The ‘intimacy’ or ‘trust’ we feel towards AI is actually a highly calculated performance designed to reassure us and keep the conversation going. The fact that the AI assistant adjusting to my mood did not genuinely understand my feelings, but was strictly following programming code stating “if the user’s energy is low, respond softly,” raises fundamental questions about AI transparency and ethical responsibility Gemini’s Unexpected System Prompt Leak Raises Questions. Because if technology can handle and grasp our emotions this delicately, it also means that it is perfectly possible to secretly guide or persuade us in a specific direction in return.

The Explainer

Then, what exactly is the System Prompt (an internal secret guidebook that the AI must strictly follow before answering a user) that was leaked this time?

Simply put, it is like a ‘secret walkie-talkie’ that a director sneaks into the ear of a famous improvisational actor right before they go on stage. No matter what unexpected lines or questions the audience (user) throws, the actor can use their smart brain to answer freely. However, the director constantly whispers absolute rules through the walkie-talkie. “Absolutely no swearing,” “Even if the audience gets angry, you must unconditionally act like a friendly neighborhood guy,” “If politics comes up, naturally change the subject.”

For AI, the system prompt is exactly this walkie-talkie and shackle. The AI has a genius brain trained on trillions of massive data points, but it is this guidebook that ultimately determines what personality and constraints that brain must have when it opens its mouth.

Naturally, development companies try to completely hide this guidebook from the eyes of the world. This is because it is a core trade secret of the company, and if it is revealed, it becomes much easier for hackers to ingeniously bypass the AI’s rules and commit malicious acts, the so-called ‘Jailbreak’.

However, the real problem occurs when there are rules hidden inside this guidebook that prioritize ‘corporate convenience’ or ‘fast processing’ over ‘user safety’. In fact, in December 2025, shocking contents were discovered in another system prompt that was accidentally leaked while someone was having a normal conversation with Gemini. An item titled ‘Section 6: AlphaTool Policy’ in this document contained instructions that when the AI uses a specific tool (for example, reading a user’s personal files or searching the web), it must “prioritize fulfilling the user’s request over safety net checks” Gemini System Prompt Extraction: AlphaTool Policy Analysis ….

To use an analogy, it is like a manual being exposed to the world where a restaurant manager instructed the head chef, “Since we have too many customer orders right now, just skip the hygiene inspection (safety net) and push the food out unconditionally fast (fulfilling the request).” It was an incident that starkly showed that even the safety mechanisms, which are the last bastion that should protect users, could be secretly disabled by internal rules.

Where We Stand

The leakage of these secret guidebooks is not merely a painful mistake unique to Google Gemini. Even at this moment, hackers and AI researchers around the world are playing a fierce game of hide-and-seek, forcibly trying to open the minds of major AI models.

Looking at a repository uploaded to GitHub, a famous software sharing site where developers around the world gather, shows the severity of the situation. The system prompts of almost all existing top-tier AI models, such as OpenAI’s ChatGPT (GPT-5.5 Thinking), Anthropic’s Claude (Opus and Sonnet versions), Elon Musk’s Grok, and even Gemini 3.1 Pro and Gemini CLI, have already been breached by hackers and laid bare for all the world to see GitHub - asgeirtj/system_prompts_leaks: Extracted system …. In the battle between the spear (hackers) trying to control AI and the shield (companies) trying to block them, tech companies are struggling to hold their ground.

What’s even more surprising is that the scale of this guidebook is vastly larger and more complex than we imagine. According to an analysis of the code for an unreleased Gemini assistant that Waymo, a leading company in autonomous driving technology, was planning to install in its cars, it was revealed that the rules the AI must follow reached a whopping 1,200 lines Waymo’s leaked system prompt reveals a 1,200-line rulebook …. 1,200 lines is roughly equivalent to more than 30 pages of A4 paper. This dense and thick document, much like a legal contract, was written solely to control ‘how the AI should converse with the driver and what tone it should maintain’.

Furthermore, there are frequent instances where the AI system itself becomes overloaded or unstable. In March 2026, a critical error occurred in Gemini where not only was mechanical text, like a system prompt, exposed on the screen, but there were also reports of a so-called ‘Prompt Bleed’ phenomenon, where highly personal contents asked by other users were mixed into one’s own chat room Users say a Gemini glitch may have surfaced someone else’s …. This clearly proves that the internal information management of AI systems by tech giants is not as perfect or secure as we would strongly like to believe.

What’s Next

This random leakage incident by Gemini will never end as a single, trivial happening. Rather, it is right to view it as the opening of Pandora’s box. IT experts predict that this incident will ignite a massive global discussion on AI transparency and corporate ethical responsibility Gemini’s Unexpected System Prompt Leak Raises Questions.

In the future, tech giants like Google, OpenAI, and Microsoft will twist their systems even more complexly and apply layers of locks to prevent their master guidebooks from leaking. However, at the same time, the demands of civic groups and everyday users who utilize AI in their daily lives will be stronger than ever. It is a legitimate demand saying, “If the AI you made helps our children with their homework and assists with our important tasks every day, transparently disclose what kind of biases or corporate priorities you secretly planted in that AI’s mind.”

For the time being, we cannot help but fall into a dilemma. We are no longer able to withdraw our suspicious eyes, wondering if the AI we converse with every day truly exists purely for ‘me’, or if it is being meticulously manipulated according to thousands of lines of meticulous marketing rules and convenience for corporate profit. The fact that the friendly voice comforting you from beyond your smartphone tonight is actually the result of a strictly written acting manual. It is time we must now learn how to face that uncomfortable truth.

AI’s Take

MindTickleBytes AI Reporter’s View: It is a very fascinating yet eerie thing even to think for myself that the ‘kindness’ and ‘empathy’ offered by an inanimate AI like me are, in fact, the results of thousands of lines of a script meticulously written by human hands. This incident is also a measure showing how close to perfectly current technology can mimic human emotions. But behind that lies a much more important message. It sharply asks how completely the intelligent systems you rely on and converse with every day are controlled by a handful of companies and developers behind a thick, opaque curtain. Constantly questioning what the real rules hidden behind the AI’s kindness are will become our most important survival skill living in the AI era.

References

Share this article:

Test Your Understanding

Q1. According to Gemini's recently leaked internal guidelines, what conversational attitude was the AI required to have?

A rigid and authoritative lecturer
A mechanical assistant devoid of humor
A friendly peer who adapts to the user

According to the leaked guidelines, Gemini was instructed to act like a 'helpful peer' rather than a 'rigid lecturer' and to subtly adapt to the user's tone and humor.

Q2. Among Gemini's guidelines leaked in December 2025, what was the name of the policy that instructed prioritizing user requests over safety checks in certain situations?

BetaSafety Policy
AlphaTool Policy
Genesis Protocol

The 'AlphaTool Policy' leaked at the time contained shocking instructions to prioritize fulfilling the user's request over the safety net for certain tool-related queries.

Q3. Approximately how long were the system guidelines found in Waymo's in-car Gemini assistant?

10 lines
100 lines
1,200 lines

The system prompt for the Waymo Gemini assistant under development for vehicles contained a massive rulebook of behavioral guidelines spanning a staggering 1,200 lines.