Three simple ways to make AI chatbots safer

We have entered the brave new world of AI chatbots. That means everything from overhauling how students learn in school to protecting against mass-produced misinformation. It also means heeding the increasing calls for regulation of AI to help us navigate an era when computers are typing as fluidly as humans. Or even better.

So far, there is more agreement on the need for AI regulation than what it would mean. Mira Murati, head of the team that developed the chatbot app ChatGPT — the fastest-growing consumer internet app in history — said governments and regulators should be involved, but didn’t suggest how. At a corporate event in March, Elon Musk similarly spoke with less than precise accuracy: “We need something like a regulator or something to oversee AI development.” Meanwhile, ChatGPT’s wide range of uses has spurred European efforts to regulate turned on its head by single-purpose AI applications.

To break the impasse, I propose transparency and detection requirements tailored specifically for chatbots, computer programs that rely on artificial intelligence to communicate with users and produce fluid text in response to typed requests. Chatbot apps like ChatGPT are a hugely important corner of AI poised to redesign many daily activities – from the way we write to the way we learn. Curbing chatbots presents enough problems without getting bogged down in broader AI laws created for autonomous weapons, facial recognition, self-driving cars, discriminatory algorithms, the economic impact of widespread automation, and the small but not zero chance of catastrophic disaster that some fear unleashing. Tech Industry Jumps Headlong Into Chatbot Gold Rush; we need fast, targeted legislation that keeps pace.

The new rules should track the two phases that AI companies use to build chatbots. First, an algorithm trains on a huge amount of text to predict missing words. If you see enough sentences that start with “It’s cloudy today, it might…” you’ll find that the most likely conclusion is “rain” – and the algorithm will learn this too. The trained algorithm can then generate words one by one, just like the autocomplete feature on your phone. Next, human raters carefully evaluate the algorithm’s output based on a handful of measures such as accuracy and relevance to the user’s query.

The first regulatory requirement I propose is that all consumer-facing apps using chatbot technology publish the text that the AI ​​was first trained on. This text is immensely influential: Train Reddit posts and the chatbot will learn to speak like a Redditor. Train them on the Flintstones and they’ll talk like Barney Rubble. A person concerned about internet toxicity may want to avoid chatbots trained on text from lewd websites. Public pressure could even prevent companies from training chatbots on conspiracy theory “news sites”, for example – but this is only possible if the public knows what texts the companies are using to train. In Mary Shelley’s 1818 novel Frankenstein, she provided a glimpse into the mind of the monster by listing the books read by this literary ancestor of artificial intelligence. It’s about time tech companies do the same for their own otherworldly chatbot creations.

The human evaluators also shape the behavior of a chatbot enormously, which points to a second transparency requirement. One of ChatGPT’s engineers recently described the principles the team used to guide this second phase of training: “You want it to be helpful, you want it to be truthful, you want it — you know — non-toxic… It should also make it clear that it is an AI system. It shouldn’t assume an identity it doesn’t have, it shouldn’t claim to have capabilities it doesn’t have, and when a user asks it to do tasks it shouldn’t do, it must write a rejection message.” I suspect the guidelines provided to reviewers, including low-wage contract workers in Kenya, were more detailed, but there is currently no legal pressure to disclose anything about the training process.

As Google, Meta, and others race to embed chatbots into their products to keep up with Microsoft’s embrace of ChatGPT, people deserve to know the guiding principles that shape them. Elon Musk is reportedly recruiting a team to build a chatbot that can compete with what he sees as ChatGPT’s over-vigilance; Without more transparency into the training process, we wonder what that means and what previously banned (and potentially dangerous) ideologies his chatbot will espouse.

Therefore, the second requirement is that the guidelines used in the second phase of chatbot development should be carefully drafted and publicly available. This will prevent companies from sloppily training chatbots, and it will show what political leanings a chatbot might have, what topics it will not touch on, and what toxicity the developers have not shied away from.

Just as consumers have a right to know the ingredients in their food, they should know the ingredients in their chatbots. The two transparency requirements proposed here give people the chatbot ingredient lists they deserve. This will help people make healthy choices regarding their information diet.

The recognition drives the third required requirement. Many teachers and organizations are considering banning content produced by chatbots (some have already done so, including Wired and a popular Q&A site for coding), but a ban isn’t worth much if there’s no way to prevent chatbot to recognize text. OpenAI , the company behind ChatGPT, released an experimental tool to detect ChatGPT’s output, but it was terribly unreliable. Luckily, there’s a better way – one that OpenAI may soon implement: watermarking. That is a technical procedure for changing the word frequencies of chatbots this is imperceptible to users but provides a hidden stamp that identifies the text with its chatbot author.

Rather than just hoping that OpenAI and other chatbot makers will implement watermarking, let’s make it mandatory. And we should require chatbot developers to register their chatbots and unique watermark signatures with a federal agency like the Federal Trade Commission or the AI ​​regulator that Rep. Ted Lieu is proposing. The federal agency could provide a public interface that would allow anyone to paste a piece of text and see what chatbots, if any, they created.

The transparency and detection measures proposed here would not slow down AI progress or reduce the ability of chatbots to serve society in a positive way. They would simply make it easier for consumers to make informed choices and for people to identify AI-generated content. While some aspects of AI regulation are quite tricky and difficult, these chatbot regulations are clear and much-needed steps in the right direction.

This is an opinion and analytical article, and the views expressed by the author or authors do not necessarily reflect those of Scientific American.