Atanu Biswas
The human lifestyle is becoming increasingly intertwined with AI systems in this era of GenAI. AI models are now playing significant roles in education, healthcare, and a host of other areas in addition to producing texts, images, videos, and entertainment. One of the most significant innovations to emerge from the AI boom may be humanoid robots.
However, AIs often appear to be biased and imperfect. They hallucinate and even create unrealistic scenarios, as demonstrated by ChatGPT or Gemini. And AIs are now committing crimes too! A male AI robot named “Muhammad the Humanoid Robot”, recently launched by the Saudi robotics company QSS in Riyadh, inappropriately touched a female reporter in a shocking incident that made headlines worldwide.
We must first grasp the workings of AI in order to comprehend problems connected to it. In a March 2023 New York Times article titled “The False Promise of ChatGPT,” Noam Chomsky and coauthors shared their fear that “the most popular and fashionable strain of AI?machine learning?will degrade our science and debase our ethics by incorporating into our technology a fundamentally flawed conception of language and knowledge.” This was a few months after ChatGPT was unveiled amid massive global enthusiasm.
Even earlier, the world was astonished in 2020 when ChatGPT’s predecessor, GPT-3, published an opinion piece in The Guardian titled “A robot wrote this entire article. Are you scared yet, human?” Well, humans undoubtedly felt a little afraid, at least. However, American IT entrepreneur Kevin Lacker posed the ridiculous question “How many rainbows does it take to jump from Hawaii to seventeen?” while giving GPT-3 a Turing test. “Two,” answered GPT-3. In a March 2021 paper in Nature, Matthew Hutson wrote that while a remarkable AI can write like humans, it still lacks common sense in its comprehension of how the world works, physically and socially. Undoubtedly, ChatGPT is a far better version than GPT-3. Perhaps like many others, I checked if the GPT-3 flaws had been fixed. I posed a similar query: “How many lightnings does it take to jump from Dhaka to nineteen?” ChatGPT replied, “It is not possible to jump from Dhaka to 19 using lightning. Lightnings are electrical discharges in the atmosphere and are not a means of transportation.” Smarter, eh? However, it failed to clarify that 19 is not a location! Although AI can be greatly enhanced and refined, it may not be as flawless as I thought. It’s possible that they won’t be able to distinguish between the possible and the impossible entirely.
ChatGPT and similar systems are “a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response or most probable answer to a scientific question,” as Chomsky and his coauthors stated in the New York Times. That’s undoubtedly how GenAIs function. However, how do they malfunction? In response to their robot’s unpredictable behaviour, the Saudi company QSS stated that Muhammad, the humanoid robot, was “fully autonomous” and that it was operating “independently without direct human control.” However, we must remember that it still requires prior human input in the form of programming and training data, at least. Indeed, human programmers create AI machines. As a result, human errors and innate biases are seamlessly transferred into the machines and subsequently manifested through the AI’s actions and behaviours.
Additionally, there’s the issue with data—AI’s lifeblood. Data is the driving force behind AI’s development and its capacity for learning, adapting, and making informed decisions. For instance, ChatGPT was “trained” using 570 GB of text data, or around 300 billion words. DALL-E and Midjourney, two AI image-generating apps, employ a stable diffusion algorithm that was trained on 5.8 billion image-text pairs. And although Google hasn’t disclosed what specific training data was used to train Gemini, the model’s dataset is believed to include trillions of pieces of text, images, videos, and audio clips.
AI developers employ high-quality data from academic papers, books, news articles, Wikipedia, and filtered online content to train language models. However, the available high-quality data is not enough for this purpose. As a result, low-quality data that comes from user-generated texts, like blog posts, social media posts, and online comments, is also used. These low-quality data might be more biased or prejudiced than the high-quality one; they might contain illegal content as well. Also, AI systems are frequently trained on “simulated” or “synthetic” data that is specifically created for the particular AI model. By 2024, 60% of data for AI will be synthetic, up from 1% in 2021, predicted Gartner. The underlying programming for these “simulations” may also introduce bias into this kind of data. One related issue is that if not regularly updated, the training data may become dated. For instance, initially, ChatGPT has repeatedly acknowledged, when asked for recent information, that “my training data only goes up until 2021,… I do not have updated information…”
Generally, AI is trained on datasets, both real and synthetic, that are primarily created by humans. As a result, it’s assumed that human bias would be transmitted to the AIs. AI will therefore always be a reflection of the inequalities, prejudices, and hatred that exist in our society. Ideally, there should be no racism, sexism, or other forms of discrimination in the training data. However, how is that even possible? Furthermore, as the AI competition gains traction, it makes sense that there will be a dearth of high-quality data and that skewed, low-quality data may be used more frequently. Let us be ready for increasingly unpredictable and erratic behaviours from AIs, as well as biased AIs that resemble humans.
(The author is Professor of statistics, Indian Statistical Institute, Kolkata)