The big news in artificial intelligence (AI) this past year was the arrival of GPT-3, a substantially improved version of the “Generative Pre-trained Transformer” from OpenAI, an advanced AI system built on a web of artificial neural networks, deep machine learning, and massive collection of data on human language. The system has been described as a giant step towards the realization of AGI, “artificial general intelligence”, the ability of a system to use language in virtually any domain of human activity. I wrote about this development in the latest issue of Language Learning & Technology, a special journal issue on big data and language learning. I discuss the breakthrough represented by AGI:
Normally, an AI system will be able to deal effectively only within a narrowly defined domain, for which the system has been trained, so as to expect specific language patterns typically used in that context. Google Duplex, for example, does a remarkable job in conversing over the phone with human operators in making dinner reservations or reserving a ride on Uber. GPT-3, in contrast, has been shown to interact through language in a wide variety of genres and content areas: creative writing, journalism, essays, poetry, text-based gaming, and even writing software code. The Guardian newspaper ran an article written by the program, while the New York Times asked it to write about love. A blogger used GPT-3 to write multiple blog posts, subsequently receiving numerous subscribers and notice on tech websites. The fact that many readers were not able to tell that the GPT-3 generated texts were written by an AI system raises questions of trust and authenticity, mirroring the concerns raised about audio and video “deepfakes”, based on training an artificial neural network on many hours of real audio or video footage of the targeted individual.
The system represents a remarkable achievement in its ability to write in natural sounding language (idiomaticity, flow, cohesion). That ability is based on the collection and analysis of huge volumes of speech data collected by crawling the internet, including all of Wikipedia. GPT-3 translates that data into a very large (175 billion!) set of connections or “parameters”, i.e. mathematical representations of patterns. These parameters provide a model of language, based not on rules, but on actual language usage. That allows the system to predict speech sequencing, based on regularly occurring constructions of words and phrases, thereby enabling the machine production of natural-sounding language utterances. One can imagine how powerful GPT-3 could be integrated into a smart personal assistant such as Siri. We are already seeing interesting uses of chatbots and intelligent assistants in language learning. A company called LearnFromAnyone is building on top of GPT-3 a kind of automated tutor, which can take on the identity of famous scientists or writers.
While GPT-3 and other advanced AI systems represent a significant technical achievement, there are, as I discuss in the article, plenty of reasons to be cautious and thoughtful in their use, as is the case generally with big data in both social and educational contexts. While the language generated by GPT-3 mimics what a human might write in terms of language use, compositional structure, and idea development, the texts don’t always make sense in terms of lived human experience, i.e. demonstrating an understanding of social norms and cultural practices. Human beings have the advantage in communicative effectiveness of having lived in the real world and and having developed the pragmatic abilities to generate language that is contingent on human interactions and appropriate to the context. We also can use crucial non-verbal cues, unavailable to a machine: gesture, gaze, posture, intonation, etc.
I argue in the article that a human factor is a crucial mediating factor in implementations of AI systems built on top of big data, particularly in education. Learning analytics (collection of data about student academic performance) tends to treat students as data, not as human beings with complicated lives (especially these days). I discuss these and other ethical and practical issues with data collection and use in the context of D’Ignazio and Klein’s Data feminism (2020). The book explores many examples of inequities in data science, as well as providing useful suggestions for overcoming disparities in data collection (favoring standard language use, for example) and for recognizing and compensating for algorithmic bias.