Dr Steven Watson, Glenlead Head of Education Research and Associate Professor at the Faculty of Education, University of Cambridge

Few AI tools have caught the world’s attention like the release of ChatGPT by OpenAI in November 2022. In this blog, Glenlead Head of Education Research, Dr Steven Watson, gives a brief overview of what this revolutionary technology can and cannot do and calls for greater public education on its potential for both harm and good.

8 August 2023

How often is ChatGPT berated for giving incorrect answers, for ‘hallucinating’? How often has it been suggested that new and improved models will become more accurate? These questions really miss the point and betray a deep misunderstanding of what the technology is.

Since the launch of ChatGPT in November 2022, there has been significant public interest in the capabilities and implications of Large Language Models (LLMs) for society at large. Initial assessments suggest that ChatGPT and similar LLMs could be beneficial across various sectors, though comprehensive understanding of their societal impact is still in its infancy. These models build upon existing deep learning technologies but distinguish themselves with user-friendly interfaces that employ Natural Language Processing, making them accessible to non-programmers.

Media portrayals often anthropomorphize ChatGPT, attributing to it characteristics of an intelligent ‘being,’ which muddies public comprehension of its actual functions, limitations, and potential risks. As recent as July, David Brooks, quoted cognitive scientist, Douglas Hofstader in the New York Times, “We’re approaching the stage when we’re going to have a hard time saying that this machine is totally unconscious. We’re going to have to grant it some degree of consciousness, some degree of aliveness.” This demonstrates a breathtaking and almost irresponsible misunderstanding of LLMs.

On a more practical level, misconceptions abound that LLMs like ChatGPT serve as new kinds of search engines capable of generating original insights. In reality, these models lack real-time internet or database access and operate on a predetermined dataset up to a particular cutoff date. Therefore, the content generated should be understood as a recombination of existing information rather than novel research or academically validated knowledge. Distinguishing these attributes is crucial for assessing the value and limitations of LLMs in various societal applications.

When evaluating ChatGPT and similar LLMs, instances where the model generates inaccuracies or “hallucinations” are frequently cited. While such errors may raise concerns when viewed through the prism of informational retrieval, they should not obscure the primary functionality of these models as advanced language tools. In fact, their central capability lies in acting as universal translators.

The function of LLMs extends beyond mere language translation to tackle the complexities of communication in a hyper-individualized society. While the Tower of Babel sought to unify through a common language, LLMs facilitate interactions among individuals who not only speak different languages but also possess unique semantic frameworks. In a contemporary setting where individual experiences and cultural backgrounds create a tapestry of diverse meaning systems, even people using the same language can find themselves at semantic odds.

LLMs offer a form of “semantic translation,” adapting not just across formal languages but also across these individualized meaning systems. This capability is particularly relevant in modern society, where social media, subcultures, and specialized professions often create their own jargons and idiomatic expressions. LLMs have the potential to navigate these nuanced communicative landscapes, thereby serving as a tool for bridging semantic gaps and fostering greater mutual understanding.

In the process of training, LLMs like ChatGPT employ intricate recursive algorithms to detect patterns within the data they are fed. These patterns can span multiple layers of complexity, from individual symbols and sentences to broader genres, classes, and even the context in which text is used. Essentially, these models capture the semantic structures that have evolved over time through human interaction and experience. These structures guide both action and communication, serving as a form of societal ‘memory’ encoded in language. By recognizing and replicating these patterns.

In the computational process, LLMs like ChatGPT use sophisticated algorithms to identify and match patterns present in the input data. These patterns range in complexity, from single words and phrases to larger constructs like genres and thematic contexts. By doing so, the model essentially learns the semantic rules and structures that have been created over time through human communication and action. In other words, LLMs effectively “read” the existing societal ‘memory’ encoded within the data.

To illustrate, consider a user who inputs a text prompt into ChatGPT. The model analyses the semantic structure of the prompt and generates a response that aligns with it, effectively transforming the initial input into a different but semantically congruent form. This capability is not limited to textual data; it also applies to other types of content, such as images, and will likely extend to multimedia formats in the future.

For example, a user may ask the model to summarize a complex scientific article. The LLM identifies the structural and thematic elements of the article, mapping these to its training data to generate a summary that retains the essence of the original text. Similarly, a user could input a jargon-filled text from a specific subculture or profession. The LLM could then translate that jargon into more commonly understood language, effectively bridging semantic gaps between specialized communities and the general population.

Therefore, the capability of LLMs extends beyond mere language translation to include the more nuanced task of “semantic translation.” This is a crucial function, particularly in a society where individual subcultures, professions, and even social media platforms frequently develop their own sets of idioms, slang, and technical terms. In such an environment, the ability of LLMs to navigate these individualized semantic systems is of significant utility, facilitating more effective and nuanced communication across diverse societal segments.

In addition to their role in translating among individuals, genres, and specialisms, LLMs have also shown capability in bridging the gap between human natural language and computational language. In other words, they can function as intermediaries between individuals and other forms of technology, such as software or hardware systems. For example, a user could input a prompt that requests the conversion of a natural language description into a segment of code. The LLM interprets the semantic structure of the natural language prompt, and, guided by its training data, generates a piece of code that embodies the intent of the original prompt.

This function of translating natural language to code adds another layer of complexity to the utility of LLMs, making them even more versatile. This specific feature is not only important for facilitating communication between human users and technological systems, but it also holds potential for democratizing access to programming and computational processes. Individuals who may lack the specialized training in coding or software development can still perform basic computational tasks by relying on the model’s ability to translate intent into executable code.

Moreover, this computational-linguistic translation extends the influence of LLMs into areas like automation, data analysis, and even machine-human interface design. It opens the possibility for greater integration of LLMs into various technological infrastructures, thereby enabling more seamless interaction between humans and machines. Consequently, in an era characterized by rapid technological advancement, the role of LLMs as mediators among diverse forms of language and communication—be it human to human, human to machine, or machine to machine—is increasingly invaluable.

Final words

The utility of Large Language Models in enhancing various forms of communication is evident, spanning human-to-human, human-to-machine, and even machine-to-machine interactions. While the technology offers considerable advantages, it also harbours the potential for misuse, such as manipulation or exploitation. Therefore, immediate action is required to improve public comprehension of the technology’s functionalities and limitations.

The growing influence of Large Language Models in various sectors is both promising and cautionary. These tools have the capacity to revolutionize how we communicate and interact with technology. Nonetheless, the dual-use nature of such models, capable of both benefit and harm, underscores the need for public education on their capabilities and ethical considerations. Immediate measures to enhance understanding can guide more responsible usage and governance of this emergent technology.

Dr Steven Watson (Cantab) is an Associate Professor in the Faculty of Education at the University of Cambridge. The focus of his research is concerned with the impact of Large Language Model Deep Learning AI on education.

Steven is co-Chair of the Knowledge Power Politics research cluster in the Education Faculty, and the Editor-in-Chief of the Cambridge Journal of Education. He holds drees in engineering and educational studies form the University of Cambridge, the Open University, and the University of Nottingham. He holds qualified teacher status with a Postgraduate Certificate of Education form the University of Sheffield.