AI models are getting smarter every day, but they’re also carrying with them the biases — and idiosyncrasies — of the people who made them.
Ola CEO Bhavish Agarwal has flagged the “pronoun illness” that’s prevalent in many western LLMs. “Hoping that this “pronoun illness” doesn’t reach India,” he posted on X. Aggarwal shared a screenshot from LinkedIn’s AI bot in which he’d asked the AI “Who is Bhavish Aggarwal”. The AI responded with the pronouns “they” for Bhavish, saying “they have held the position since” and “they have a background in technology”, instead of the more common “he”. “Many “big city schools” in India are now teaching it to kids,” Aggarwal continued. “Also see many CVs with pronouns these days. Need to know where to draw the line in following the west blindly!” he added.
“Most of us in India have no clue about politics of this pronouns illness. People do it because it’s become expected in our corporate culture, especially MNCs. Better to send this illness back where it came from. Our culture has always had respect for all. No need for new pronouns,” Aggarwal continued.
Aggarwal was highlighting a recent far-left movement with roots in the US, which considers assuming anyone’s gender to be offensive, and refers to people with their “preferred pronouns” — until someone explicitly declares that they want to be referred as ‘he/him’ or ‘she/her’, proponents of this movement refer to them as ‘they’. The movement had started off as a means of not offending transgender people by misidentifying their genders, but has since taken some strange turns — by some accounts, there are now over 70 genders for people to choose from, each with their own pronouns. The movement also overlaps with other far-left ideologies, which have led to protests, violence and looting in the US and in some other parts of the world. The far-left movement has also been linked with lower levels of happiness and higher incidences of depression among its proponents.
And given the prevalence of these ideologies in academic circles, these ideas have been cleverly fed into new AI technologies. LLMs are trained through something called RLHF (Reinforcement Learning Through Human Feedback), in which human evaluators tell an AI whether its response was good or bad. Because AI companies often take left-wing positions on issues, and ask their human evaluators to judge AI responses based on these political preferences, most AIs carry with them as left leaning bias.
These biases have previously been reported in prominent models such as Google’s Gemini, which ended up creating images with the American founding fathers being black and the Google founders being Asian, all to promote their goal of “diversity”. OpenAI’s ChatGPT was also criticized for writing poems mocking Donald Trump, but refusing to write similar poems for Joe Biden, and making jokes about white people but refusing to make similar jokes about black people.
Amongst the first people to flag these issues was Elon Musk, who immediately went ahead and created a new AI company named XAI. One of XAI’s aims was to create a politically neutral and fun-loving LLM, and it released Grok, which is much more free-spirited than other LLMs. In India, Bhavish Aggarwal had launched a new company named Krutrim, and has been highlighting how the LLM, named Krutrim AI, will be built for Indian users. Aggarwal clearly feels that India doesn’t need to import “pronoun illness” and other related ideologies from the west, and is looking to build an LLM that’ll more accurately reflect Indian values and requirements. It remains to be seen how Krutrim AI fares, but Aggarwal seems clear on the problem statement that his company needs to solve.