Being polite to AI might cost – but it’s worth it

A few months ago, Currency urged users of ChatGPT, DeepSeek and other AI users to stick with the “Ag pleez” in their prompts. 

And with good reason: researchers from Japan’s prestigious Waseda University found that “politeness of prompts can significantly affect [the AI’s] performance”, as this is “thought to reflect human social behaviour”. Being rude to your machine could lead to greater bias and the wrong answers, they said. 

Only it seems all these extra lubricating phrases, like “Sorry, Siri” or “Dankie, DeepSeek”, are costing an arm and a leg, in the aggregate at least.

In recent days, Sam Altman, founder of OpenAI and Elon Musk’s nemesis, answered a question about how much all these unnecessary pleasantries add to the electricity costs of large-language models by saying “tens of millions of dollars”.

As Currency reported a few weeks back, even 100 words generated by an AI model consumes about half a litre of water. And while each query will likely only add fractions of a cent to the power costs, in the aggregate this makes a big difference to cash flow. 

This deference is not uniquely South African. A study by publisher Future of 1,000 people in December found that between 67% and 71% of people in the UK and US are polite to AI.

Moreover, another survey of 2,000 people in the US by market research organisation Talker Research found that 48% believed this wasn’t just a nicety; machines actually “deserve to be spoken to politely”. As it was, just 7% of the respondents said their approach was “sometimes impolite”, with people “admitting to swearing or being abrupt in manner”.

Of course, much of this politeness is based on a nebulous suspicion of some looming “AI karma”, with 39% of respondents in that Talker Research survey agreeing that “our past behaviour towards AI, Alexa, Siri and all things robot will be taken into account somehow”.

But as Altman points out – in an admission that many hadn’t thought of until now – this all comes with a cost. 

George Washington University physics professor Neil Johnson told the New York Times that this makes sense, since AI machines decipher requests in the same way that you might open a package from a retailer like Amazon – you have to wade through a mass of extraneous tissue paper to get to the content. And that tissue paper and cardboard costs money. 

For aspiring AI companies, for whom cash flow is at a premium, here is where the rubber hits the road. And the costs go beyond development; the cost to the climate is also immense. 

Microsoft, which is invested in OpenAI, said in May last year that its CO₂ emissions had risen nearly 30% in the preceding four years due to the investment in data centres needed to breathe life into AI. Google put its emissions at 50% higher over four years for the same reason.

“The computational power needed for sustaining AI’s growth is doubling roughly every 100 days,” said researchers at the World Economic Forum last year. A training model like OpenAI’s GPT3 is estimated to use about 1,300MWh of electricity – “roughly equivalent to the annual power consumption of 130 homes in the US”, they said. 

For OpenAI’s new model, GPT4, a typical query uses 10 times the energy of a typical Google search, according to Goldman Sachs.

More pleasantries, please?

And yet, South Africans are burning up the AI search engines. According to the Boston Consulting Group, South Africans were the 13^th most “enthusiastic” globally about AI, with 38% “excited”, 42% feeling “conflicted” and only 20% “concerned” about the tools. This put South Africa’s enthusiasm level higher than the US, UK or Australia – all countries that are far ahead in pioneering AI technology. 

Many of these queries will have been buttressed by pleasantries, with the occasional plea to the imminent AI overlords not to banish them to Gqeberha. 

Altman, however, believes the additional money spent on pleasantries to your AI chatbot is money “well spent” – a reflection of the still woolly idea that doing so provides some sort of more intangible utilitarian benefit.

But, as the Japan study suggests, there is a logic in this. Ben Wood, a chief analyst at CCS Insight agrees, telling TechRadar that being polite to AI performs a wider societal function.

“Personally, I think this is important from a societal perspective. If it becomes acceptable to be disrespectful to AI assistants, that behaviour will start to leech into interpersonal interactions, and that’s a slippery slope,” he said. 

_{Top image: Rawpixel/Currency collage.}