AI in Art Conservation

Artificial Intelligence (AI) represents a generational paradigm shift on how we work and live. How can art conservators use it? Here is how I understand it so far...

When I first heard of AI in early 2023, I was very excited by its promise to interact with an “information robot” that would learn from the user!

As I started to try and apply it to my work as a conservator, and relying on the kindness of my tech savvy brother, I soon realized it was more complicated than that. This substack will relate on my findings as I navigate using this new technology in support of art conservation work and research.

What do I mean by AI?

AI as a term has become shorthand for any tool that uses a Large Language Model (LLM) coupled with Natural Language Processing (NLP) algorithms to generate response, using a chatbot interface such as ChatGPT. These tools allow users to interact with the digital assistant to ask questions or generate content based on information the LLM has access to in its knowledge base. Most often we will encounter these as chatbots in online shopping. An LLM is a neural network with an extensive understanding of human language, which it acquires through training on a massive amount of internet text data. It's essential to note that while AI can mimic human responses, it operates on learned patterns rather than absolute truths.

This technology has great potential for academic research use as well as to access information quickly for one’s daily tasks. Using the paid version of OpenAI, for example, it is possible to create a customized Generative Pretrained Transformer (GPT) to retrieve specific information from a set of documents uploaded by the user, and give the assistant instructions as to the tone and nature of the response you are looking for. This model is able to produce text that is both logical and contextually appropriate, similar to what a human might, and can handle tasks like text completion, translation, and summarization. GPT is like a smart text writer that, based on inputted large amounts of datasets it was trained in, starts with a sentence and predicts what words come next, keeping everything in its "memory" until it reaches the end of a sentence. So, if you ask GPT a question with some context, it can give you a response that sounds good, but it might not always be entirely correct because it guesses the words based on patterns it has learned, rather than knowing the absolute truth.

GPT Architecture
GPT models learn language by breaking it down into units called tokens, which can include clusters of words, punctuation, and other text components. This process helps the AI understand and process language more effectively. GPT is a ‘next-token predictor’ meaning it works as a predictor for the next part of a sentence, making educated guesses based on probabilities, starting from a given beginning. As it generates tokens (which can be words or parts of words), it stores them in its "memory" along with the initial starting point. However, it's important to note that GPT has a limit of what can be imputed into a custom chatbot. GPT-3.5 initially had a 4,096 token limit (corresponding to about 3,500 English words), but newer models can go up to 32,768 tokens, with pricing based on token usage, meaning more tokens equal higher costs. If you add too many documents or text beyond this limit, it might start to forget or lose track of some of the earlier information as it needs to fit within the allocated memory space.
One way to reduce the amount of text uploaded is to shorten the text into "embeddings." These are special codes that are assigned to words or groups of words, akin to turning words into numbers, so the computer can process them better. These codes are created in advance, and they're like a map that helps the model navigate and understand the knowledge we want it to work with. Once we've done that, we upload the knowledge along with these code-like embeddings to the GPT model. Now, the model can use these codes to understand and generate text more efficiently. It's like giving the model a set of tools to work with the information. These codes allow the model to compare pieces of text to see if they're similar. So, when you ask a question, the model can find the parts of the knowledge that relate to your question by checking how similar they are. This process, called "similarity search," is a key part of the document helper we're developing, helping the model find and provide the most relevant information to you.

AI Applications in Art Conservation

For art conservators, AI serves as a digital ally in accessing practical data such as preparation of adhesives or insights into the use of new materials. This technology can support us in staying up to date to the latest research on a topic and make informed decisions during the conservation process.

There is also an aspect in this new tool that democratizes access to knowledge, transcending geographical barriers and fostering global collaboration within the art conservation community. Its multilingual capabilities and nuanced responses enrich conservation practices, offering sophisticated insights not readily available through conventional search engines.

Creating a collaborative Application Programming Interface (API) for art conservation would be ideal, leveraging knowledge from databases like the AIC wiki or CAMEO, providing real-time access to relevant and updated information.

For daily tasks, the basic, free model of ChatGPT can be used it to help write abstracts and letters of recommendation, translate texts, or any other writing task. I also use it to help me build my own custom GPTs with knowledge datasets that I upload. I have created a public one and several private GPTs that I use to “interact with my data”.

The more familiar I become with how to use these, the better I become as a user, and the more useful the technology becomes.

Challenges and Caveats
Although AI algorithms, such as large language models (LLMs), can provide valuable information, caution is necessary. The accuracy and reliability of the data they provide should always be confirmed, as information provided by LLMs may require revision and scrutiny. The quality of the response is often a direct result of the quality of the question. Users must possess a nuanced understanding of conservation practices that allows interpreting AI-generated output with critical thinking. Continuous updates and maintenance of AI models are imperative to uphold the field's standards.

Another obvious issue is data copyright: if I am uploading datasets that rely on other people’s copyrighted work, am I infringing those limits? I believe the answer might be yes both legally as well as ethically.

While AI opens doors to innovative approaches, it is crucial to exercise judgment, validate results, and continually refine AI models to uphold the ethical as well as practical standards of the conservation field. By embracing AI responsibly, art conservation professionals can leverage its benefits and pave the way for a more advanced and inclusive future in art preservation and restoration.

This article was written with AI assistance.