By Jaspreet Bindra
Chat GPT created intense excitement when it was released about 10 months back, on November 30, 2022. The first few months the press, analysts, scientists, pundits, technologists, people like you and me were overcome with shock and awe by the capabilities that Chat GPT, Bard and others exhibited. They had almost a human-like way to interact with us, have conversations with us. They could do things that we never thought an AI could do. They appeared almost intelligent, and that, in fact, gave rise to some fears about AI in general, and GenerativeAI, in particular. Many of us have grown up watching movies like Terminator or Hal where, a super intelligence which humans invent, eventually rules over humans and becomes much more superior to human beings. That fear started coming in and the excitement which we had slowly started to turn into some nervousness and some fear. While the prospects of super intelligence are probably there, they are, in my opinion, very dim. What we have to worry about far more are the clear and present, the here and now ethical issues which surround Chat GPT, GenerativeAI and AI in general. I classify these issues broadly into three different kinds.
The first, and probably the most talked about right now, which is the most out there is about how GenerativeAI models plagiarize. There are problems related to copyright and content ownership. Getty images, for example, is suing Stable Diffusion in the London High Court, accusing it of illegally using its images without permission. A bunch of authors have sued OpenAI, Google, Microsoft and others saying that their works were used to train these models and they got nothing in return. The New York Times is also contemplating doing the same – if the New York Times wins in court, theoretically, OpenAI would be liable for USD 150,000 of fine for every time that it violated copyright. If you can imagine the sheer amount of data that Chat GPT is trained on, this could translate to billions of dollars.
What happens if you, for example, have Stable Diffusion or Dall-E to comb the web and combine multiple images, say a Pablo Picasso and a Mona Lisa, who owns it? You, the AI model, Pablo Picasso or Da Vinci whose original pictures and compositions were mashed together. So, this is a very, very thorny issue. We do hear news that BigTech is working with publishers to have some kind of compensation mechanism. Even if that happens, we do not know whether it will be fair. What will happen to small creators? People like you and me who have created material that is being used, without our permission or compensation, to train these models. So, plagiarism is #1.
The second big issue is that these models are inherently biased. While OpenAI and others have done a reasonably creditable job to build ethical guardrails around Chat GPT, therefore, the model does not spout racist or sexist content. Many experts, including Gary Marcus and Timnit Gebru, argue that these guardrails are thin and the model is our moral and we are sitting on an ethical time bomb. Many times, these models have been jailbroken. Famously, when a New York Times reporter interacted with Chat GPT, and Chat GPT professed its love for the reporter and told him about how the reporter’s marriage was a sham. Models like GPT have been trained on Reddit and Wikipedia and all these massive web sources. Well, 67% of Reddit users in the US are men, less than 15% of Wikipedians are women. Obviously, the inherent biases which exist in us human beings and which are translated into the web are then reflected in these models. Dall-E2, too, for example, tilts towards creating images of white men and sometimes over sexualizes images of women. There are doctors and then there are women doctors. The second big ethical issue, therefore, is bias, and we need to take care of that.
Finally, GenerativeAI models are quite bad for the environment. Blockchain gets the bad rap, but a transformer model which builds an AI, 1000th of GPT3, for example, can spew carbon dioxide equivalent to that of 125 NewYork-Beijing flight round trips. That is a lot of carbon released. Models like ChatGPT, Bard, etc. live on the cloud like Microsoft Azure, Google Cloud or AWS. This cloud is nothing but hundreds of data centres, guzzling water and power at alarming quantities. An article in The Guardian revealed that data centres currently consume 200 Terawatt hours per year, roughly equal to what South Africa currently consumes. This is the third issue – of the environment.
As we go forward to build all these models, we need to figure out how we can keep these ethical issues in mind and develop the science and technology further. Do we make smaller models? How do we make green data centres? How do countries regulate so that companies are constrained to not be so abusive to the environment or have bias or manage the plagiarism issues that come up.
The author is managing director , founder, The Tech Whisperer
