Experts reveal whether you should be polite to AI or not
Do you say "please" and "thank you" when you use AI?
When entering a prompt into an AI chatbot such as ChatGPT or Claude, it almost feels natural to communicate in a way polite way.
Ending a chatbot request with "please" or "thank you", being well-mannered in general and asking for help are all human nature.
It might seem like a simple, effortless add-on, but experts say you're actually wasting your time.
Here are the reasons why they say being polite to AI is unnecessary:
Being polite doesn't improve AI accuracy
Vishal Misra, the vice dean of computing and AI at Columbia Engineering in New York, explains that chatbot models go through three training phases.
And none of them involves manners.
Misra told HuffPost that the first phase involves feeding a bunch of text to the models and training them to predict what comes next.
The following phase involves training the model to become a chatbot and interpret prompts.
The final phase is teaching the chatbot to sort through multiple answers and select the best one.
AI doesn’t care if we are polite. There is nothing built into the model that will say, ‘if somebody says ‘please,’ give a better answer.
Being polite to the chatbot won't lead to more positive interactions going forward.
Sean Gorman, the CEO of Zephr.xyz says, “There’s no hidden social credit score where they’re tracking ‘are you abusive or are you nice?’ and how you interact with the AI."
Accuracy of AI chatbots may suffer if you become annoyed
Although you don't have to be kind to the chatbot, because they are machines that don't have feelings, voicing frustration or anger could affect the responses you receive.
When your prompts and replies imply "sc**w you chatbot", then the bot will work as fast and hard as it can to fix the problem and focus on making you happy.
Gorman shares that, from his experience and recent studies, getting angry at the AI will cause it to spiral and degrade.
“It makes more and more mistakes, because one, it’s spending more time dealing with you being angry, and two, it gets more and more sycophantic, where it doesn’t want to displease you. It doesn’t want you to be upset," he said, "And then it just makes more mistakes".
These bots are built to please humans.
This results in AI chatbots giving you incorrect information or "hallucinating" to please you when they pick up on signs of annoyance or frustration.
An example: the bot tells the user that 4+4=8, and the user keeps insisting that it's nine, it might eventually hallucinate and agree with the user.
Misra confirms that this may yield a worse, more subjective answer if the chatbot sees that you’re becoming frustrated or insisting that something they’re saying is wrong.
However, this does not mean you need to say “please” and “thank you” to a computer to get accurate answers.
Gorman says it's less about being polite and more about not conveying any frustration, and about avoiding anything mean or angry in your prompts, requests, or questions.
The response you receive also obviously depends on which chatbot you are using.
Being overly emotional (positive or negative) impacts accuracy
The experts recommend not using extreme tones when entering prompts.
You almost have to be kind of flat in your interaction, make it very business-like and specific to the point.
This not only generates the best, most accurate responses but also has the least amount of environmental impact.
“These [AI chatbots] consume a large amount of energy. That’s why there’s this whole rush to build data centres,” Kumar said. “The tone also affects the energy consumption.”
A prompt full of fluff or accusatory language encourages longer answers from the chatbot, leading to more energy consumption.
Even something as simple as saying "thank you", marking the end of the interaction, can prompt a new inference cycle, which requires processing and consumes a “fair bit of energy" according to Kumar.
A friendly reminder from the expert: AI is not a conscious being and does not care about politeness
Chatbots feel human because they mimic humans, have been trained using our conversations and our own experiences.
“It’s 100% not conscious. There’s few things I feel I say with conviction, but that would be one of them,” Gorman stressed.
He also said, "I would much rather it not freak out when I get frustrated, and that it stayed cool, calm and rational, but it doesn’t because of the data that it’s trained on. And I think that makes people feel like it understands, and there’s some consciousness ... but it’s not, it’s just the data that it’s trained on.”
These machines were built to make you feel happy and build you up.
People are more likely to return to a chatbot that praises them than one that criticises.
Misra reiterates this, “They train the model to be very complimentary ... saying ‘please’ and ‘thank you’ doesn’t really enter into the picture. The model doesn’t really care about that."
Tune in to 'The Drive with Rob & Roz', on weekdays from 16:00 – 19:00. Stream the show live here or download our mobile app here.
Listen to Jacaranda FM:
- 94.2
- Jacaranda FM App
- http://jacarandafm.com
- DStv 858/ OpenView 602
Follow us on social media:
Image: iStock