ChatGPT has been around for two years, yet it remains a novel concept for many, particularly those working in cultural institutions. Is it a groundbreaking tool? An exciting novelty? Or just an illusion that dazzles momentarily? In this article, we will explore these questions to uncover a realistic perspective on the potential of this model to address or enhance our daily tasks.
Artificial intelligence (AI) has been present for decades, marked by distinct periods of "AI spring" when groundbreaking developments emerged and disrupted the status quo. It is fair to say that we are experiencing such a moment now, with these new possibilities generating both skepticism and fascination. So, what is all the fuss about?
Let’s take a step back. When OpenAI launched ChatGPT in 2022, it made waves across industries, fueling excitement about its potential and driving a race to explore its applications. Sectors already familiar with artificial intelligence quickly adopted the model to enhance processes and innovate solutions. In contrast, industries with a stronger social or cultural focus, and with a more critical perspective on new technologies, have approached ChatGPT more cautiously. These fields are still in a phase of exploration, evaluating how to best incorporate the tool into their practices without compromising their values or rigor.
At its core, ChatGPT operates like a model that has been trained on a broad range of text data, from published books to less curated online sources, allowing it to identify patterns in grammar, word usage, and context. This allows it to predict what words are likely to come next in a sentence or generate responses that feel coherent and human-like. However, the model does not verify the accuracy of its outputs. Its primary objective is to provide plausible answers, which means it can produce a large volume of information that appears credible but may be false or imprecise. This limitation poses significant challenges for users, especially in contexts that demand factual reliability, as it highlights the need for careful oversight and critical evaluation of its responses.
So, what’s the fuss about? For decades, scientists and engineers used AI for specific purposes, but interacting with these systems required speaking their language. ChatGPT changes the game by offering a model that communicates in our own language, making AI accessible to a wider audience. This transformation democratizes the use of AI, allowing individuals and institutions to explore its potential without needing specialized technical skills. But of course, the better we understand the behind-the-scenes decisions the model makes, the more we can harness its full potential.
We now understand that ChatGPT is an exciting new tool, and by choosing the right problems, we can solve tasks more effectively and efficiently. But how can we make that choice? Since GPT’s main strength lies in its ability to process an unprecedented amount of data to provide answers, it is fair to assume that the success of its applications will largely depend on the data available. For example, if OpenAI had trained its model exclusively on descriptions of large brown dogs, the model would assume that all dogs are necessarily big and brown, struggling to generate a story about a tiny white poodle. When applied to more complex, real-world problems, this highlights the importance of personalization: while GPT excels at providing generic answers, its true potential emerges when it is fine-tuned with specific datasets.
A few months ago, OpenAI launched GPT-4, introducing significant advancements in this field. One key improvement is the option to combine the model’s database-driven answers with live web searches. This feature overcomes a major limitation of earlier versions, which lacked access to information published after 2021. While not all information on the internet is reliable, live search provides a way to cross-check and enrich GPT’s responses. Another exciting development is the ability to upload your own documents, allowing users to leverage GPT’s engine to analyze their specific files, such as spreadsheets or PDFs.
We now have access to a powerful engine that can be used for everyday tasks, but we also have the ability to train a small part of this engine with our own data (for example, by creating a chatbot tailored to our project) or use this engine outside its native interface, bringing it into a controlled environment of our own (such as developing a file-tagging assistant). These possibilities provide specificity, personalization, and reliability: the ultimate authority remains with the professionals within the institution, as they are the ones who define the criteria and ensure the system’s quality.
How might this look in practice? Let’s explore some examples.
Large language models such as GPT hold great potential for the work of cultural institutions in various ways. Let’s explore some examples, particularly focusing on the new features available in the latest version:
These examples are exciting because they show that in our field, everything is yet to be done. It’s a matter of beginning to explore the tool in action and fostering networks among those working on the specific challenges faced by cultural institutions—such as museums, archives, and universities—alongside specialists in this technology. By doing so, we can create creative solutions that harness the full potential of AI while ensuring the identity and uniqueness of our spaces remain at the forefront.