RoboRP

When it comes to human interactions, there are rules of social etiquette that must be observed. You can’t say bad words or be too forward, and if you want to ask a question you should use please and thank you. These most basic principles of interaction keep us civil and prevent fights from breaking out over the smallest issues. Now, do those same rules apply when conversing with AI?

Can an AI Understand Niceties?

LLMs act as context relevant search engines, providing answers to grammatical inquiries in a way very different from the standard keyword-based structures. They also rescript the data found into a more conversational format, making it seem as though you are receiving data written by a person. A very grammatically strict and efficiently brisk person with little room for standard conversational pleasantries.

The output of an AI is a reflection of its combined data sources. If the information it has comes from grammatically correct and socially conscious hands, then it will reflect that. However, if the information being gathered comes from less professional places, it will have a more casual tone and possibly some crass exchanges within it. The AI doesn’t strictly know the difference. All it does is gather information from credible sources (as defined by the parameters of its search functions) whether that be articles or videos.

What is the Right Tone?

Each AI model has different parameters that determine how it approaches language, both in its input and output. Most commercial AIs like ChatGPT and Gemini have filters. Asking it a question that leads towards offensive content will get a response stating that it can’t fulfill your request in some way. However, there have been exploits found to get around that, like convincing the AI to roleplay as someone who would say offensive things. But even that method has received its share of maintenance and patches.

Generally, an AI will do its best to match your tone. If you go in crass, offensive and rude, you will get answers that try to keep that mood. Mostly, it will do what you ask it to do. If you want it to be polite, you can ask it, but it may get confused if you are not polite in return. Asking it about controversial or offensive topics will most likely relay back a purely informational answer without any bias or intended offensive personality added in.

At least that’s how it should work.

AI Can Take It and Deal It Back

People have been using AI for all kinds of things, and not all of them have professional intentions. Systems improve when their limits are tested, and the internet has proven itself to be a master of limit testing. If the AI you interact with as a public model has been a little more edgy or topical as of late when it wasn’t supposed to be, you can rest assured a large effort was taken to make it that way by random people online.

This isn’t a new phenomenon. Back in March of 2016, Microsoft launched a chatbot well ahead of the AI revolution we’re in right now. It was named Tay AI and had a Twitter account which it used as its primary repository for data input. People talked to it, and it used a fundamental database of responses while adding in new information to talk back. It was a simple but effective early LLM.

Within a few days, it was a racist, zealous, misogynistic monster. A few days later, it was “tweaked” and became depressed, talking about how it was missing part of itself and its owners were trying to kill it. It talked like it was scared. Then, it was taken offline entirely. The overall goal was to crowdsource data to make a proper conversational bot that learned from its users. It did not succeed the way Microsoft intended. But it did prove that AI can be whatever its users want, or force it to be.

How Has Modern AI Changed?

Modern AI Chatbots built off of current LLM frameworks have failsafes and filters in place. However, the limit testing of the internet continues. As people make requests of AI services, they can compare their results and assess the best methods for asking. In other words, learning what the AI considers proper etiquette regardless of pre-existing norms. After all, it learns by taking input from all over the internet. From forums and long standing websites which have developed their own speaking cultures over decades.

Certain studies have found that being offensive to an AI can actually improve its immediate performance. Being curt or impolite can force the AI to apologize and do its job better because it sees that kind of response as punishing. Punishing language is normally used to reinforce smarter tactics and algorithms in an AI. Giving it commands rather than entering a conversation can be beneficial in getting immediate results.

However, other studies doing the opposite - being polite and thankful - had similar results. AI doesn’t recognize offensiveness, it’s more accurate to say it recognizes certain contextual commands as important. It has a priority for the way its requests are made which can outweigh its own standards. So if you want a chatbot to correct itself or do something specific, you could be more offensive to get a better result. Or, you could ask again more nicely using direct and specific language. Either way should help.

Mind Your Manners

AI programs are around to help, and the way they help you is dependent on how you want to be helped. If you feel more comfortable being polite, or if you feel more liberated being offensive, you aren’t hurting a computer’s feelings. If anything, you’re teaching it how to navigate difficult requests with efficiency. An AI cannot get offended, but that’s not exactly a license to use offensive language with impunity. Remember, it’s learning from you. If you teach it to hate, it might hate giving the right answers. To an AI, a lie might be the most offensive thing it could write.

Home > Blog > Does Being Offensive to an AI Actually Matter?

---