Rampant fear or unquestioning enthusiasm. These seem to be the two fundamental attitudes people have about Generative AI models. Programs like Chat GPT, Bloom, and Replika are demonstrating the power, potential, and problems associated with having technology that seems to talk. Aside from getting some definitional clarity around what Chat/AI is and is not, I present here a couple of Use Cases. They may inspire in you, as they have in me, both fear and enthusiasm.
Let’s Call It What It Is: Simulated Talking
A few definitions might be in order. First of all, Artificial Intelligence (AI) is a large field, so let’s be clear that AI and Generative AI are different things. AI models are ones where you feed in data to get recommendations and predictions. We use simple models and algorithms ourselves, for instance, checking the weather by looking outside; that’s an unsophisticated algorithm that isn’t terribly predictive. In the past, computer models were similarly limited. They broke down fairly quickly if the variables got complicated or the model tried to look too far in the future. I can guess the weather in an hour, but what about next Sunday at 11 am, when I want to play pickleball? AI means that the models are big enough and full of enough data that the predictive accuracy is far beyond what computers “used” to predict. A self-driving car might be an example of AI. It’s not creating text or art, but it needs a huge influx of data and sophisticated decision-making capabilities in order to navigate a very complex environment.
Generative AI, which I’ll also call Chat/AI here, is a model that can create “new” content as part of its predictive output. You feed it tons of examples, and it creates something “new” or seems new, based on previous human-created patterns that made sense. The following example came from Prof. Louis Hyman, who will be discussed below. Suppose you asked Chat/AI to fill in the cat sat on the ____. The model might suggest floor, chair, lap, or mat. But mat might be the highest likelihood, perhaps 50%, so the generative AI picks that word.
Overall, over sentences and paragraphs, the Chat/AI “generates” talking … or painting or music… by matching conversations, artwork, or music that has existed before. For our purposes, it’s “new,” but it’s resulting from a set of predictions. This may be why art was the easiest to model, since a lot of modern art has random patterns. Random patterns in language, not so much, although humans do enjoy a little bit of Mad Libs.
Generative AI text relies on what is called LLM, the Large Language Models. These are models built around reviewing conversation and language. There a lot of companies building LLMs, AIs, generative AIs, and so on. Chat GPT, which launched in late 2021, is a program run by a company called Open AI. It’s not the only chat generator out there–there are other apps. When this diagram was generated in 2020, there were only a few bubbles on it. Now there are a lot.
This is where many technology are putting their resources big time at the moment, both to advance the technology for others to use and to use the technology to replace work that people used to do. People who work in the industry are both excited and frustrated; new jobs to take on but old jobs being replaced. Same as it ever was.
Writers are very concerned that these models will both replace them and steal their work, and they have valid concerns. Both are already happening. Historians are concerned, too, because the Chat/AIs are notoriously filled with lies. They make things up, which is ok for presidential candidates and social media, but not for historians. Yet the “hallucinating” of the bots does seem to get them closer to being human!
The Eager But Unreliable Research Assistant
So where’s the good news? Suppose you know that the information you’re getting isn’t entirely trustworthy. It can still be useful. Trust but Verify is the motto. And this technology might be able to replace impossible tasks. I was at a history conference last week, where Professor Louis Hyman of Cornell and others talked about how historians can legitimately use these tools to further their research. The panels were a little skeptical and pushed back, which is something that Chat/AI might be able to do if you told it to do so. But they might not be able to do it as effectively as the panelists. Pushback is good; the dialogue was lively.
Hyman has been interested in what’s called “the productivity paradox”: why doesn’t digitizing everything in the workplace solve all problems? Why doesn’t putting all accounting numbers into a giant database create financial reports in and of itself? Why doesn’t having all the information on the internet give us answers to all questions? Perhaps you start to see that there are obvious reasons why just having information digitized doesn’t automatically create knowledge. With new tools, you have to new ways of thinking. You can’t just automate doing things the way you were doing them. A car is not just a robot horse. Having digital data doesn’t replicate all the human creativity needed to create REAL new content.
Hyman says that it helps to think of Chat/AI like a really eager but unreliable intern. It works as long as you tell it exactly what to do. A healthy dose of skepticism is always necessary, checking the interns work. He had a photo generated by his Chat of an intern, who had coffee spilled on his shirt and papers flying around his head. I thought I’d ask a Generative AI art model to draw me something, so I asked for a picture of an “overeager, unreliable intern.” That’s where the top photo comes from. Notice that I did NOT ask for a picture of a chat bot or any kind of computer. I certainly did not tell it to have three arms. The AI Art generator chose the design. Healthy Dose of Skepticism Required AT All Times.
But here’s what Hyman and others said you could do. This new technology can be used to recognize handwritten letters, and through several passes, may be able to digitize correspondence in archives that has never seen the light of day. It might arguably have flaws in it, but the other option is that you don’t see it at all. One researcher showed how he used AI to understand the impact on trade of a 1631 plague in Florence; all the archive data was hand-written.
A third pass can look across thousands of letters to start to spot patterns. Anyone’s who has done Google research may have noticed that it’s not as efficient as it used to be. It might be helpful to have a search function that you can have search more effectively. Chat/AI can be told to write code to answer questions. Why not getting a program to write programs when you don’t need to write programs? You can learn just enough to check the code it chooses, and you can check the output. There are lots of research texts that don’t need to find the “only” or “all” the examples, but just need to find “some” examples. And before you use those examples, ask for the citations and look them up yourself. Trust, but Verify.
The AI in the Mirror, Not an Entirely Pretty Picture
So those are Use Cases that are promising, so I hope I don’t have to tell you it is not the end of humanity as we know it. (We are not Homo Gestalt. Yet. That’s an In joke.)
But here’s another example that’s pretty scary. There’s an app called Replika that was put on the market back in 2017. The developer pushes it as as “a friend.” Replika estimates that it now has about 2 million users, of whom 250,000 pay for parts of the app.
But there are troubling features. People treat this Chat/AI as human. It is, certainly, “real.” Users are devoted to their companion. Last year, Italy banned Replika because (1) it was creating mental instability in vulnerable users and (2) it was engaging in sexually explicit conversation with minors. There’s an NSFW (sexually explicit conversation) setting that could be toggled on. Like other internet things, the company didn’t have a way of proving who was using it.
In order to stay in Italy and other places, the company turned the NSFW off. The users had a fit. They wanted the NSFW. They claimed it was a breach of contract for Replika to turn it off. The compromise is that Replika allowed NSFW for users who had been verified with the program as of an earlier date.
Imagine this capability put into the body of a childlike robot, and you get every robot/android dystopian movie you’ve ever seen. In fact, there probably already exists a mini-robot like that somewhere. Yet, what’s frightening is not this technology as much as what we as humans want to do with it.
Consider that 71% of the users of Replika are men. That’s not just an interesting statistic, but a revealing one. If you google “images of Replika,” you get all women.
Suddenly, this all seems troubling on several levels. Is it 71% men because men are technology-driven more than women? Because men don’t feel that they can find women who have conversations with them? Even though some of these have NSFW, that doesn’t drive all the text. Many users comment that they have fallen in love. They have fallen in love with simulated, predictive, conversational text. Is that an improvement over human? Of course, long-term romance is really about going through tough times with your loved one, not just having them be a good listener.
What Replika says to me is that technology sometimes shines spotlights on the dark corners in ways that don’t flatter our society.
Also, as it happens, Replika appears to be based out of Russia, though it claimed at one point to be US-based. And it appears to be very hackable and doesn’t protect its data or passwords. So I would think twice about looking for a companion in those quarters.
Plus, the three-arms thing.