By Chelsi Slotten
ChatGPT took the world by storm at the end of 2022 and has grown significantly since then. Companies are integrating AI into search functions, new models are released somewhat regularly by a variety of existing companies, schools are stressing about plagiarism, startups seem to be spinning up every day using this technology, and it’s being touted as a REVOLUTION in work. For all the excitement that surrounds it there is still a lot that the general user doesn’t know about ChatGPT or other generative AI models. Before we get into the risks and benefits of ChatGPT to archaeology, let’s talk about what it is.
What is ChatGPT
ChatGPT is a chatbot that allows the users to communicate with a Generative Pre-trained Transformer(GPT) that has been trained using a type of Natural Language Processing(NLP), which is itself a type of artificial intelligence (AI). In other words ChatGPT is given a question and, using the data it was trained on, it predicts the answer word by word. Another popular type of AI is image generating AI such as DALL-E which turns language into pictures. Importantly, the predictions these models make can be wrong, which varies by what the models are trained on. ChatGPT is trained on… the internet. More specifically it is trained by scrapping publicly available information off the internet. This is a common method of training AI models.
The problem with this is the internet can be a pretty dark place full of biases, misinformation, and trolls. This has, unsurprisingly, led to a model that replicates those biases and very confidently hallucinates incorrect answers. Not everything it says is incorrect but it doesn’t differentiate between what is correct and not. While a new model GPT-4 has been launched that is apparently better at telling the difference between true and not true, I haven’t tried it out. But enough about what ChatGPT is. How can it be used in archaeology?
ChatGPT…The future of archaeology?
Luckily for archaeologists the resounding answer here is no, ChatGPT will not be replacing archaeologists any time soon. That is not to say that AI models won’t impact our work. Even ChatGPT admits “GPT has the potential to aid archaeologists in various aspects of their research, but it is important to note that it should not replace the human interpretation and analysis of archaeological data.” Startling to me, in 1 year of asking this bot about the impact of AI on education, research, coding, statistics, publishing, and writing, this is the first time I have ever seen it mention not replacing humans. It can help us though.
Some areas that AI models like ChatGPT could come in useful are in text and image-based tasks. That could be assisting with translations, text mining documents to identify patterns or specific examples we need- although it would need to learn how to cite for that to be helpful, creating potential images of what sites may have looked like based on descriptions in old texts, decoding ancient scrolls, and improving science communication. I am particularly excited about the potential to improve science communication. As long as the information being fed into it is already verified, it will help avoid the hallucination issue.
AI models like ChatGPT could also be useful in helping with the business administration of archaeology companies, museums, heritage societies, and field schools. This work is not archaeology specific and has broad ramifications for impact in fields like HR, accounting, and data analysis.
Drawbacks to using ChatGPT in Archaeology
There are three main issues with using ChatGPT, and other AI models like it: bias, hallucination, and attribution. These exist because of how the models are trained and the scale of the models.
Attribution is at the heart of a lot of the issues that exist around AI models. The question of where this information comes from isn’t well addressed. If you ask ChatGPT for an answer to a question, the answer might be predicted based on looking at academic papers on ResearchGate, Wikipedia entries, blogs, and social media prior to 2020, which is when the model was first released. It doesn’t tell us where it is getting the information from, so it is difficult to assess its validity.
AI models could also run afoul of copyright issues. ChatGPT is training itself ons ome text that is copyrighted. Does it owe a fee and attribution for using those materials? Models like DALL-E are trained on artwork that are created by people, stealing their work and style. Generating images in the style of “insert living artists name” steals their creativity and livelihood without compensation or recognition. How do we ensure artists are properly compensated when their work is used to power AI technology? These questions don’t yet have answers.
Another issue that arises from attribution problems is hallucination. Hallucination in an AI model is when the model confidently tells you an answer that is wrong. The model has predicted what the next word in the sentence should be incorrectly. If we don’t know where the information the model is trained on comes from, there is no guarantee that it is factually accurate, which leads to hallucination. If the models told the user where the information came from they could make a better judgment on the answer. Without that transparency it can be difficult for users to understand what is true and what is hallucinated by models like ChatGPT.
Lastly are the issues of bias in AI models. Again, this comes from issues around what text the model is trained on. ChatGPT is trained on the internet, the internet can be a pretty seedy place full of misinformation and prejudice. If you train an AI model on biased data you will end up with a biased model. This has happened before with Tay, a chatbot that became racist after being exposed to social media. There is also the issue of bias among the programmers who created it. Software engineering tends to be dominated by white males. The lack of diversity in those who are creating the models used by AI will show up in a lack of diversity in how the models work. This risks further disenfranchising or marginalizing underrepresented groups.
We also know that biases change. ChatGPT was trained on a specific set of data from 2020 and earlier. So it will reflect the biases that existed in the world at that time. Those might not be the biases of 50 years ago, or today although they may be similar. They also will be the biases of a select group, the people who wrote the content it is trained on. The model is essentially a snapshot, and it will create content with the biases of that snapshot. Using the model to undertake our work carries the very real possibility of reinforcing and perpetuating biases. The same biases that many archaeologists are working to break down.
The Future of ChatGPT
“Yeah, but your scientists were so preoccupied with whether or not they could, they didn’t stop to think if they should.” – Dr. Ian Malcolm, fictional character, Jurassic Park.
11 months ago several leaders in the technology space signed an open letter asking for a 6 month pause to the development in AI models. That didn’t happen. There are more models, and AI based companies than ever before. The letter does point to concern over what this technology could mean for how we live. We know that AI is having a negative impact on cyber security, something that should concern anyone who conducts their business online. We also know that it is making google worse, an annoying occurrence for those of us who use google for everything.
Which is not to be all doom and gloom. AI has the potential to help people and companies grow faster, automate the boring stuff, and reduce errors in some processes. It could help free humans up to do more of the creating and innovating. Within archaeology, ChatGPT and AI models like it could help handle the business side of things so archaeologists can spend more time in the field, analyzing, researching, and understanding the past. The key for me, is that it needs to be created and used in such a way to respect privacy, limit bias, minimize inequality (or at least not make it worse), and respect safety and security issues.
Leave a Reply