Internet, Chatbots and Avoiding Bias

Everyone, left or right, feels the Internet is biased. Too much of what we see on Facebook, WhatsApp, Google search results, news sites etc is either biased or outright fake. The problem is even more dangerous with AI chatbots like ChatGPT. After all, you ask it a specific question and it generates an answer. But the answers it produces depend on the kind of input it learns from. And since the data it learns from is biased, what it learns and creates as answers is biased too.

 

What could be a solution to the bias problem? (I’ll leave fake out of this blog).

 

Google tried to address the problem in its AI chatbot named Gemini. By finetuning Gemini to give certain kinds of answers, and not give other kinds of responses. The intention was to “avoid creating or reinforcing unfair bias (e.g. sexism, racism etc)”, wrote Andrew Sullivan. But as Nate Silver pointed out:

“Unbiasedness is hard to define.”

As if that wasn’t hard enough, “cultural, social and legal norms” differ, and vary drastically across countries.

 

How then did Google’s attempt pan out? Gemini was programmed to avoid “passing judgment”. This has led to ridiculous situations e.g. ask it who was worse for society - Hitler and Elon Musk? And it refuses to answer (“It is up to each individual to decide who they believe has had a more negative impact on society.”). A blanket don’t-judge policy clearly isn’t the solution.

 

Ask Gemini for an image of a random physicist from the 17^th century and “it will give you an Indian woman, a black man, an Arab man, and a white chick with a woke dye job”! Ask for images of Singaporean women and you get images of Asian women. But ask for images of British men and it says:

“I’m still unable to generate images that specify gender and ethnicity.”

 

Many blame such outcomes on political correctness taken too far. Yes, white males unfairly and inaccurately dominate the narrative; and yes, women, blacks and Asians are marginalized. But the solution isn’t to insert those categories into everything. Gemini, for example, throws up pics of the Pope as a woman, and a request to generate a pic of a Nazi threw up a black guy in a Nazi uniform.

““When “respecting cultural norms” supersedes accuracy, there is, in fact, no guarantee of accuracy.”

 

Many now fear that if Gemini-like “don’t reinforce bias, don’t pass judgment” content becomes the norm, we’ll quickly end up in a world where many would start to believe that Popes were often women, that the 17^th century had several Asian and black physicists, that Nazis were often blacks. A 1984-like world:

“Every record has been destroyed or falsified, every book has been rewritten, every picture has been repainted, every statue and street and building has been renamed, every date has been altered. And that process is continuing day by day and minute by minute. History has stopped.”

 

Today, on the Net, you can get ideas on how to murder your spouse (and get away with it) and so on. Gemini was programmed to not answer offensive, illegal and unethical questions. Slippery slope. It has led to ridiculous situations where, as Ben Thompson wrote, “Gemini won’t help promote meat, write a brief about fossil fuels, or even help sell a goldfish”. The boundary between accurate-but-offensive, accurate-but-undesirable, seemingly-offensive-queries-can-sometimes-be-information-gathering-exercises will stop being drawn. Things will just get blocked based on the sounds-offensive/undesirable filter.

 

Are attempts to moderate AI chatbots, and the Internet in general, then impossible goals? Raghu S Jaitley wonders if the problem is fundamental to what we would like to achieve via AI, namely that AI should be: (1) useful (actionable, not vague); (2) truthful (an offensive fact shouldn’t be erased); and (3) harmless (it shouldn’t cause harm or perpetuate stereotypes). What if, wonders Jaitley, these 3 goals are mutually contradictory?

 

If that is true, any attempt to control and moderate AI chatbot could be an example of what Francis Bacon once said:

“The remedy is worse than the disease.”