What potential do large language models such as ChatGPT have for cultural institutions such as archives, libraries and museums? What are the false expectations surrounding them?
ChatGPT is the new cool kid on the block. The public launch of OpenAI’s new language model has triggered a wide range of reactions, from excitement to fear. For those of us who see it as a tool with enormous potential, there is still a lot of confusion and a feeling that ChatGPT can and knows everything.
It is said that AI is moving from being an external tool or prosthesis that enhances human capabilities to becoming more anthropomorphic or human-like, with chatbots and virtual assistants such as Siri or Alexa being the best examples of the latter. This motivates questions such as “will AI take our jobs?” or, on the other hand, “will AI free us from all our unwanted tasks?” and “does this mean that AI is some kind of mechanical slave?”.
Leaving aside the ethical and philosophical considerations, this post argues that it’s healthy and appropriate to ask if and how AI can help us automate some of the tasks we find boring, repetitive or don’t have the time to do at all.
We argue that people working in cultural institutions with a mission to preserve, understand and make accessible cultural heritage can benefit from new technologies without compromising the high scholarly standards of their work.
The cultural and educational sectors are often underfunded and understaffed, facing new challenges to keep up with current forms of cultural consumption and to retain the engagement of their audiences, so it’s worth considering how technology can support them.
In our work with cultural institutions, we’ve often heard about the impossibility of cataloging new collections, even if they are the most meaningful and exciting you can imagine – because of staff shortages or lack of resources. If you are reading this and work in a museum, archive, library or similar, you probably have many other tasks in mind that you never get to do, either because your to-do list is full of boring administrative stuff, or because your team is simply not big enough to take on a new and exciting project.
This post explores what AI language models like ChatGPT can and cannot do to help cultural institutions fulfill their mission and make their staff more satisfied with the process and outcome of their work.
ChatGPT is the superhero of language models – but what are its superpowers?
The first step in harnessing the potential of AI language models for our work in cultural institutions is to understand what they are and what they are trained to do.
Let’s start with ChatGPT, as it has become famous enough that most people have at least a rough idea of what it is, but most of what follows applies to many other models out there. To define it briefly, ChatGPT is a language model that has been trained to generate human-like responses to text input. It is its conversational ability that has impressed us so much, i.e. its ability to “understand” the question or request that we write on the chat interface and to respond with text that actually makes sense. Whether it is accurate beyond making sense is another question. For now, let’s focus on why it reads so human-like, and we’ll come back to accuracy later.
If ChatGPT is able to process speech so well, it is because it has been trained on an unprecedented amount of data (something like the equivalent of several million books). What does it mean that a model can “learn”? The concept of learning, one of the many human-like metaphors used in artificial intelligence, actually refers to the process of exposing the model to large amounts of text data and instructing it to recognise the patterns and structures of language. During the training process, the language model is presented with examples of text and the language patterns associated with them, such as grammar rules, word usage and contextual meaning. The model then analyzes these patterns and uses them to generate new text, make predictions about what words will come next in a sentence, or answer questions based on the information available in the training data.
We just need to ask nicely
Now that we know roughly how ChatGPT has been trained, we can deduce what it will probably do well. As you may have seen, ChatGPT is very good at tasks that require a general knowledge of how language is structured internally (grammar, vocabulary, how words combine). Examples include summarizing or rewriting a text, translating into different languages, generating a text based on a given topic and instructions on format and style, or in response to a question, etc. The model can also generate keywords that describe the subject of a larger text, or extract the words that make up the names of people or places.
All of this can come quite handy to support the work of cultural institutions, right? However, ChatGPT won’t do it just out of the box, but rather we need to know what to ask from it and how.
We need to think of language models as bricks with which we can build something new, something that we need. The chat implementation available on OpenAI’s website offers us a playground to see the potential, but to actually have the model do what cultural institutions do, we need to integrate it into a system that is designed to achieve our goals. If, for example, we want it to help us index large collections, we need to instruct the model to perform classification tasks to match textual context in the objects with terms in a vocabulary. If we want a system to help us draft content for exhibitions, we need to ask it to generate new text based on selected input material and a determined style.
What about hallucinations?
We can now also talk about what not to expect from ChatGPT and touch on the issue of accuracy.
Those who have played around with the chatbot might have seen that (as OpenAI clearly informs) the information in the responses might not be accurate and it might refer to facts or people that do not exist.
There is a clear explanation for this: the implementation available in the chatbot can only extract information from its training data, which is huge, but still limited. So if you ask about something on which there is not that much information out there, the chances are high that the answer will be wrong. Different would be if we would be interacting with an implementation of ChatGPT that has access to the internet and can search for relevant responses to a question. If we want the system to provide information only based on certain sources, we can also limit the corpus the model will take into account. For example, we could ask questions about art history and instruct the model to search for possible responses in a corpus of art history books.
How can ChatGPT be useful for cultural institutions
Large language models such as GPT have great potential for the work of cultural and heritage institutions in a number of ways:
- Because they are trained on a huge amount of data, they allow systems to be built to perform complex tasks without the need for large training datasets. For example, in order to index and catalog large collections, a large language model such as GPT can help us build a text classification system that is able to assign categories to text fragments and thus automatically generate metadata. This leaves us with the question of how to deal with automated metadata and whether we need to review, moderate or label it to make it transparent.
- Because of its ability to “understand” questions and formulate coherent answers by gathering information from a defined corpus, GPT can be used to develop new ways of searching within catalogs that are not limited to exact keyword matches.
- Thanks to GPT’s ability to write coherent texts and mimic the style of given examples, the model can be used to write summaries of long documents, enabling a new form of checking the relevance of a result in a search without having to read/listen/watch the whole document.
- GPT models are multilingual, meaning that they are trained on data in many languages and are able to translate and switch from one language to another, or receive a question or query in one language and be instructed to answer in another. For example, GPT can help us to subtitle videos to make them accessible, or to search collections in the user’s language.
Feel free to get in touch with us if you are considering building solutions for the cultural sector using language models and AI.