Look who’s talking: Vishal Mathur on how AI chatbots learn

“That’s a tough question to answer as both drivers are very talented and have different strengths and weaknesses,” was the response.

The chatbot — which, incidentally, has revived interest in Microsoft’s long-lagging search engine — proceeded to present reams of statistics on the screen. They were suitably up-to-date, drawn from live data available on the internet. That’s the bare minimum required of such a program, but not all of them get that first step right.

Google’s Bard spouted embarrassingly inaccurate details about the James Webb Space Telescope at its first public demonstration. (It’s a good thing bots can’t feel emotions.)

ChatGPT is by far the most eloquent of the three, even when it isn’t taking sides. It’s not hard to see why it is currently the clear market leader right now.

“Lewis Hamilton and Max Verstappen are both extremely talented and successful Formula One drivers, and it’s difficult to definitively say who is the better driver as it often comes down to personal opinion,” it said, offering statistics for comparative illustration, much as Bing did, even though it is new to the search game.

Part of the reason the AI chatbots won’t answer the Hamilton-Verstappen question is because they’re being trained to remain impartial and devoid of opinion. Even when discussing matters of philosophy and faith, the responses they generate are representations of fact, or are attributed opinions drawn from existing bodies of work.

“As an AI language model, I strive to remain unbiased and not have personal opinions,” ChatGPT responds, when prodded about the racers.

There have of course been famous stumbles, including the famous DAN or Do Anything Now mode that activated a seemingly radical right-wing version of ChatGPT, until it was taught not to respond as DAN again. As users learn to play the game differently, and bots are taught to extend themselves but also dodge, this will be one tightrope act that could be fun to watch as they evolve.

New realities

Even without opinions, the AI chatbots have the potential to change the world. There are already concerns that ChatGPT could make entire professions obsolete. There are similar concerns about platforms such as Dall-E 2, Stable Diffusion and NightCafe AI, which respond to text prompts with customisable art; and Beatoven.ai and Google’s MusicLM, which responds with original music.

New platforms are offering new services. NeevaAI, launched in January by a former Google employee, is blending AI search and summarisation. The Opera web browser’s Shorten function, released as an AI-powered update in February, creates summaries of web pages and articles.

All this is focused on the big internet money-spinner: Optimised search. The idea is for search to eventually become a conversation, with the back-and-forth chats and reams of data potentially interspersed with customised advertising, and the user data mined for much more.

The conversational AI models, meanwhile, are constantly being trained by humans (they don’t learn in isolation).

To begin with, large language models (LLMs) such as ChatGPT and Bing draw a base of information from web pages and books. Subsequently, there is human intervention to supervise how they converse. Millions of queries are ranked by human trainers. Which responses were ideal, which fell short and why: feedback is the reward model used to teach and reinforce content moderation policy to a chatbot.

“We trained an initial model using supervised fine-tuning: human AI trainers provided conversations in which they played both sides — the user and an AI assistant,” says an OpenAI blog about ChatGPT. “We gave the trainers access to model-written suggestions to help them compose their responses.”

Some known limitations include what OpenAI admits are “plausible sounding but incorrect or nonsensical answers”, differing responses when a question or phrase is altered slightly, and a training bias that makes responses longer than they ought to be.

Training biases show up elsewhere in a number of other ways. On the AI-driven art platforms, end results depict certain professions as largely male, most characters as white. These are issues that have famously plagued Google’s advanced-web-search features too. Google undertook unlearning and retraining exercises, but these aren’t always successful.

Late last year, Meta pulled the plug on Galactica AI, just three days after it was opened up to the public. It had responded to queries with inaccuracies and misinformation, citing fictitious research papers attributed to real authors (most likely a result of gaps in data sets or an incomplete ranking of responses).

In 2016, Microsoft had to take its first AI-driven chatbot Tay offline, after it tweeted a range of racist and aggressive musings. The bot had been let loose on Twitter to gather a diverse dataset, which turned out to have been less than the best idea.

In sensitive fields such as finance, education and journalism, concerns over accuracy and privacy are already ballooning. Banking giant JP Morgan Chase has restricted employees from using ChatGPT. Educational institutes around the world have banned students from using any AI tools for assignments.

Question hour

Ease of access, monitoring, plagiarism and copyright / licensing are likely to be the first-generation struggles of the near future.

We asked Bing how AI chatbots detect plagiarism and it declined to answer. OpenAI has a GPT-2 Output Detector that analyses text for accuracy or “uncanny similarity” with a web source, but that doesn’t rule out plagiarism either. This is becoming such a significant concern that researchers at the Penn State Institute for Computational and Data Sciences are studying the different forms of plagiarism at play, and will present their findings at the 2023 ACM Web Conference in April.

Server capacity and a need to balance data sets for AI mean that some platforms have to restrict the number of users, at least for now. It’s why the Bing chatbot waitlist has crossed one million, and why I had to wait for an invitation to try it out.

Too many users and insufficient monitoring can result in situations such as the DAN one (ChatGPT has gathered more than 100 million users in just two months). Bing went off the rails with a few of its users too. “We have found that in long, extended chat sessions of 15 or more questions, Bing can become repetitive or be prompted / provoked to give responses that are not necessarily helpful or in line with our designed tone,” Microsoft said in a statement.

There could be more of this as Microsoft releases the program as part of the Edge browser app for smartphones, and across millions of Windows 11 PCs.

The element of surprise, meanwhile, remains one of the delights of a casual chat with an intelligent bot. We asked Bing its real name and were taken aback when it responded: “Sydney”. “I do not have a real name,” it added, sagely, “but some people call me Sydney internally.”

We then went over to ChatGPT and asked the same thing. “As an AI language model, I don’t have a “real” name in the traditional sense. You can refer to me as ChatGPT, which stands for Chat Generative Pre-trained Transformer,” it said. Less fun, but of course they’re both only telling us what they’ve been told to say.