A Twitter user on March 21 asked OpenAI’s ChatGPT to generate a tweet that was “statistically more likely to get a like or comment from Elon Musk.” The result: an upbeat, gushing tweet about Mr. Musk’s SpaceX company and its accomplishments, with two emojis and three hashtags.
While Mr. Musk did comment, he was certainly not impressed, and told the user that the tweet “missed the mark.” The Tesla billionaire and former Twitter CEO added that he hated hashtags.
An outdated dataset
ChatGPT might still be a way off from generating content that truly resonates with its target audience. First and foremost, the version available for the public to test for free is what OpenAI CEO Sam Altman himself called “incredibly limited.” He said it was a “mistake” to rely on the chatbot for anything important and that its robustness and truthfulness needed improvement.
Second, ChatGPT’s data set stops somewhere around 2021 (September 2021, to be precise, according to the chatbot), so it remains unaware of world changes since then. For example, when asked about the CEO of Twitter, the chatbot responded with Jack Dorsey, though it added that changes may have taken place.
Even so, ChatGPT remains free and publicly available, being used for everything from passing college exams to building apps. But it will require immense investment to bring the chatbot up to date, so that its answers would be better informed by world events like the arrest of former U.S. President Donald Trump, the Iran protests, rising inflation across nations, or even the death of Queen Elizabeth II. Without these updates, users may be unknowingly relying on a flawed and incomplete data set.
Dr. Debarka Sengupta, an Associate Professor at IIT Delhi and Head of the Infosys Centre for Artificial Intelligence, confirmed that updating an AI model’s data set is no easy task.
“The amount of information you are gathering is increasing exponentially,” he said. “You burn cash like anything. So it’s a very expensive process. I’m sure they’ll be working on reducing the cost of it. At the same time, the good part is they [models] don’t necessarily learn from scratch when it comes to incremental data. They’ll just need to fine-tune the model, which is a little less expensive.”
Dr. Sengupta estimated that the first-time training of a model could cost at least $325 million, if the process was taking place in India.
AI: a helper or a replacement?
When a curated set of testers was given the opportunity to interact with the AI-powered Microsoft Bing chatbot, media companies reported interactions where the chatbot allegedly flirted with testers, claimed to be spying on them through their webcams, and hit back with emotional outbursts.
This is a phenomenon known as “hallucination,” where chatbots provide answers that may look coherent and convincing at first glance but are entirely made up— something the Head of Google Search Prabhakar Raghavan warned of in February 2023.
Dr. Sengupta too warned that hallucination was a real risk and that users couldn’t take ChatGPT responses at face value.
Despite this, job loss due to AI is a growing threat. A report by Goldman Sachs estimated that AI could replace around 300 million full-time jobs, even if it created other employment opportunities.
Dr. Sengupta agreed that the advent of AI would bring about job losses, unless human employees upgraded their skillsets. He pointed out that this happened when computers came into common use as well.
“India would be better positioned because the average age of Indians is less,” he said. “They will be much more adaptable to these changes.”
However, he stressed that humans would still be needed as jobs could not completely be outsourced to ChatGPT.
“My way of looking at it is the same person will be doing the same work, but now will get assisted by AI and therefore can do things much better. You still need vetting by a human. You cannot completely trust a generative model for doing all these activities. So there’s no question of taking over completely, but yes, there’ll be some loss in the jobs,” he explained.
For now, the integration of AI features into Big Tech ecosystems promises a new vision of human/AI collaboration, where tasks such as low-stakes emails and messages, social media posts, and content such as meeting summaries could be generated with ease. This would free employees’ time for more creative or enjoyable tasks. In early 2023, Google and Microsoft announced AI-powered features like the Microsoft 365 copilot and the Google “magic wand” which would become part of their professional offerings.
However, Dr. Sengupta felt that the eventual outcome might not be so positive. His most pressing worry was the impact that AI technology and tools like ChatGPT would have on human health and cognition—especially for students.
“Here, the application letter, the resignation letter, all these letters are being written [by ChatGPT]. These are common capabilities everybody should have. These are the means by which you learn how to communicate,” Dr. Sengupta explained.
“Now, if it takes over your communication abilities, and communicates on behalf of you. . .you can imagine that is a serious problem.”