In recent months, Alphabet CEO, Sundar Pichai, has made numerous appearances on interviews and podcasts, sharing in-depth insights about the company’s future plans around AI. Considering Google’s recent challenges and intense competition from Microsoft, it seems like a good strategy for Pichai to be making the rounds.
Last week, Pichai and a few other top executives from Google appeared in CBS’ ‘60 minutes’ video segment. In the interview, Pichai talks extensively about AI and its impact on society.
However, one interesting aspect of the video caught our attention. It talks about how Google Bard started teaching itself skills that they didn’t expect the chatbot to have.
Pichai says, “For example, one Google AI programme adapted, on its own, after being prompted in the language of Bangladesh, which it was not trained to know.”
The language being talked about here is Bengali, which, besides being spoken in Bangladesh, is predominantly spoken in the state of West Bengal and to some extent in other states such as Tripura and Assam in India.
The internet fact-checks
In the video, James Manyika, SVP, Google, goes on to say that having fed the system with very few prompts in Bengali, it had learned to now translate all of Bengali.
However, former Google researcher Margaret Mitchell took to Twitter to point out, with evidence, that the statement is incorrect. She pointed out that Google’s PaLM, the AI model which is the forerunner to Bard, had been trained to understand Bengali. A quick look at PaLM’s datasheet does reveal that Bengali is one of the languages it is trained on.
(Source: PaLM research paper)
Google had announced PaLM at their Google I/O event in 2022. During the event, Google had revealed that they trained the model to communicate in Bengali, according to Mitchell.
In the long Twitter thread, Mitchell puts forward the question, “So how could it be that Google execs are making it seem like their system magically learned Bengali, when it most likely was trained on Bengali?” This eventually leads us to ask who among Google and CBS got it wrong?
Interestingly, Mitchell was fired by Google in 2021 when she used an automated script to comb through her emails in search of evidence to corroborate the assertions made by her fellow AI ethics researcher, Timnit Gebru.
It is worth mentioning that Gebru was also dismissed by Google during the same period because she sent a critical email to colleagues about the company’s approach to diversity and AI ethics.
Adding to AI hype
In the video, Pichai goes on to say that society is not prepared for the rapid advancement of AI. While statements like this could appear fascinating for some, many, on the other hand, believe it just adds to the AI hype.
Emily M. Bender, professor of linguistics at the University of Washington, referred to Pichai’s CBS segment saying that we should not accept AI hype as news.
Since the launch of ChatGPT, AI hype has gone into overdrive and time and time again, experts and researchers have pointed out how this could turn out to be detrimental for the development of the technology. For example, it has led to unrealistic expectations around what AI is capable of achieving in the short-term.
However, private corporations stand to benefit from AI hype. It invariably leads to more investments or creates a demand for their AI-powered products or services.
According to some, claiming that an AI model learned Bengali autonomously further contributes to the hype surrounding AI by spreading misinformation. It’s no doubt that Pichai was hoping to garner attention for Google’s products and services that incorporate AI with his claims.