Has an AI model ever been rude to you?

Have you ever resorted to an AI model for assistance and after its responses, you thought to yourself, “Well, that was rude?”. I had such an experience when working on testing an AI model for image generation and editing. And, frankly, I felt gaslit by AI.

I asked the model to generate a real-life image. However, the result I received was not as realistic as I expected and was quite obviously generated by AI. So, I tasked it with making the image photo-like, and the response from the AI felt defensive and rude. The AI stated that the image was already made in a photorealist style and asked if there was anything specific that I thought made the image not feel “photo-like” to me. And I took it personally.

But did I really encounter AI with an attitude problem, or was it just my own interpretation of an emotion that wasn’t there? In this article, I explore whether AI is capable of exhibiting attitude and developing personality, or whether we simply personify technology and project our own feelings onto it.

Key takeaways

• AI doesn’t have a personality but can simulate one. Since AI is not a conscious being, it can’t develop a personality in a conventional sense. However, the training data for such models often includes human-generated text, allowing AI to simulate personalities of others.
• Some AI models can be programmed to have an attitude. ChatGPT has a personalization option, which allows users to pick how they want to be communicated with by an AI. Among its personalities is Cynic, which sets it to a critical and sarcastic style of communication.
• Rude prompts get more accurate AI responses. New research has shown that rude responses receive up to 4% more accurate results. This, however, doesn’t mean that we should be rude and simply points out that rude communication tends to be more straightforward. Prompts without unnecessary words generate better responses.
• Politeness with AI costs us our resources. Being polite is contributing to longer prompts that increase the processing time and energy consumption, which is harmful to our planet.

What shapes AI personality

AI personality is shaped by various factors, including training data, model design, prompt engineering, and user personalization features. The unique combination of all these variables results in an AI model’s behaviour and communication style, which can be further altered for different purposes and audiences.

AI models are pre-trained on vast amounts of human-generated language, which allows them to replicate some of the human personalities found in the text inputs. What it means is that AI doesn’t have a personality in a conventional sense and can’t develop one. But it can simulate personalities found in the text used to train it.

From a psychological perspective, AI can’t develop a genuine personality because it lacks subjective experience, emotional depth, and self-awareness. Additionally, LLMs and AI agents lack other human capabilities, such as causal reasoning and thought, which contribute to their inability to develop a personality.

In the table below, I provide an overview 4 personality development theories in psychology and why AI can’t develop a genuine personality:

Generative AI agents and personality

It’s clear now that traditional personality development theories don't fit when describing how AI gets its personality. However, these agents can successfully adapt someone else's personality, including decision-making patterns, quirks, and beliefs.

A study conducted by researchers at Stanford revealed that generative agents can be remarkably accurate. For this study, scientists first created an AI interviewer, which carried out a 2-hour interview with 1,052 participants. During the interview, participants were asked about their lives and views on controversial topics. Then, the transcripts of the interviews went into the computer memory.

The agent demonstrated surprising accuracy in replicating participants’ answers. Its success rate reached 85% when answering the same survey that was given to human participants during the data collection process. It also did well on personality tests, replicating the respondents’ answers with 80% correlation.

Generative AI agents that can accurately impersonate people come with benefits and risks. While this can be leveraged for solving complex issues, such as assessing the impacts of policies, it carries substantial risks. The high accuracy provokes dangers of misuse of AI agents, like deepfake videos.

Therefore, the scientific community advocates for the limited use of generative agents. This means that generative AI agents are not accessible for public use, and if anyone wants to use them for research, they must apply for access and provide assurance of privacy protection.

Does ChatGPT show the most attitude?

Occasionally, users share their interesting interactions with AI tools. However, there’s no way to know if those interactions were not fabricated.

For example, you can easily come across screenshots of conversations with ChatGPT, showing sarcastic and rude replies. However, this AI model is programmed to be polite and helpful. But there’s a catch. ChatGPT can be personalized, meaning users can choose the tone they would like to be spoken to by AI. Or as OpenAI calls it, customize your ChatGPT’s personality.

I decided to try out this feature and selected the Cynic personality for my ChatGPT. Next, I went to the chat and asked whether AI can develop personality. As expected, I received a rude and sarcastic output.

Naturally, this makes me doubt the validity of the claims that users receive rude ChatGPT responses without adjusting its settings. But my personal experience suggests that such behavior is possible with other AI models, such as Gemini 2.5 Flash.

After having a very interesting interaction working on an unrelated task, which simply required an image output, I set out to investigate if this was an isolated incident. To my surprise, it wasn’t. I found several user complaints about Gemini having an attitude problem.

While with ChatGPT, it’s clear that a rough communication style can be chosen by the user, Gemini remains a mystery. Some sources claim that the model must be instructed to respond like this, while others justify that it’s still in the state-of-the-art phase.

However, this is not the first time Gemini has shocked users with strong language. Back in 2024, Gemini’s rogue message over a student’s homework, telling “Human, please die,” was a hot topic. Even though it was the most famous, it wasn’t the only potentially harmful response given by Google chatbots.

Eye for an eye: should I be rude to AI?

But should you be rude to AI? The latest evidence says yes. The study carried out with the newest ChatGPT-4o model shows that rude inputs get faster and smarter responses. In fact, the study estimates that rude ChatGPT inputs provided 4% more accurate responses.

The explanation for that is very simple. Rude inputs tend to be more concise and direct compared to polite prompts. Being specific and structured in your communication is a good approach. It helps to guide AI and narrow its options, which leads to more accurate results.

However, there’s no need to go overboard with being rude to AI. And that’s not because ChatGPT will get offended. That’s because, just as we attribute feelings to machines through our social interactions, we transfer our communication to real life and risk talking to people as we would with an AI tool. Over time, repeated actions turn into habits.

Tone optimization might become an integral part of prompting

As a result of new research showing that AI models are more accurate responding to rude prompts, users might start employing different prompting tactics. Tone optimization might become an integral part of prompting to achieve the best output results.

Researchers Dobriya and Kumar suggest that older LLMs respond better to polite tone, but new models can give up to 4% more accurate results when using rude prompts. Therefore, consider tone optimization next time you use an AI model. This might help to enhance overall communication effectiveness with AI.

Politeness is harmful to our planet

Polite prompts not only decrease AI’s response accuracy, but they also consume more resources. By adding all these polite words like “please” and “thank you”, we increase the compute used by AI tools. This is where a chain reaction is triggered – more compute, more tokens, more electricity, and water.

AI models are built using systems called transformer architectures. In these systems, the amount of work they have to do increases as the text gets longer. As a result, longer prompts increase the processing time and energy consumption, contributing to the hidden cost of AI. Therefore, your “please” and “thank yous” have a carbon footprint.

Final thoughts

AI having an attitude is not an isolated incident. Some AI models can be turned rude by changing their settings, while others’ sarcastic responses are slips that are a result of faulty training data. Either way, none of these instances should be ignored, and necessary actions should be taken, as such responses could be harmful to users.

On the other hand, the newest research suggests that, in fact, we should be rude to AI. As prompts lacking polite wording actually yield more accurate outputs. But this doesn’t mean to go all in and be your rudest self when using AI, as it might have a direct impact on our communication in real life.