AI and Research
Did you know that Artifical Intelligence (AI) first became a field of study in the summer of 1956 in Dartmouth College, USA? It came out of an eight-week workshop attended by mathematicians and scientists who brainstormed the idea of thinking machines. Today, most AI applications are Narrow and Generative AI, although research is currently being done to develop a higher form of AI known as General AI.
Narrow AI refers to high-functioning AI systems that can replicate or surpass human output for a very specific, dedicated purpose. A large majority of AI we encounter in our everyday lives are considered Narrow AI. Think of things like facial recognition when you cross borders, chatbots on shopping websites, or even your own personal assistant, Alexa! Narrow AI is also used in the medical field, where researchers are using machine intelligence to determine cancer treatment. Essentially, Narrow AI can only perform what it is programmed to do based on the datasets it has been given.
Generative AI refers to AI systems that can create data that replicates what it has been shown before. Popular Generative models include ChatGPT, DALL-E, or Microsoft Copilot. These models can be used to create texts, images, videos, fake data, and simulations, depending on how they are trained. If they are trained with biased data, they will produce biased results.
AI: The bright side
As a researcher at Lafan, I use a variety of research methods when evaluating projects, depending on the needs of our clients. As such, for a given project, I might deal with large volumes of data in the form of survey results, interview transcripts, focus group notes, or even social media posts (with the knowledge and consent of the client, of course!). I find AI to be very useful and helpful in organising data.
For example, Microsoft Forms has an integrated AI system that creates graphs and word clouds automatically which are very convenient when it comes to generating reports. For qualitative data, I found that ChatGPT can be a useful tool for identifying patterns and themes in interview transcripts and even extract quotes from those documents. There are GPTs that have been designed by academics to perform tasks like coding for thematic analysis, which could potentially save researchers a lot of time performing intensive and time-consuming tasks.
ChatGPT also supports multiple languages. At Lafan, we are proud to be a bilingual Welsh-English company, and as a researcher, I have produced reports in both Welsh and/or English, depending on the requirements of our clients. That said, in the research process, we often collect data in both languages, depending on the preferences of our research participants. This means that I often deal with notes in both languages, and ChatGPT is able to deal with these notes and produce results in the language I require.
Areas for improvement
While AI has been incredibly helpful in my work, I’ve noticed some limitations in its performance, particularly with understanding context. For example, when processing bilingual data in Microsoft Forms, AI might highlight the word “Roedd” as significant in a word cloud simply because it appears frequently across responses. However, “Roedd” translates as “was” in English, a common word that doesn’t carry much meaning in the context of survey analysis. This highlights a key challenge: AI often struggles to discern the relevance or context of certain terms, especially in multilingual settings.
GPTs have their limitations too. One notable drawback is their tendency to provide responses to any prompt, which can sometimes lead to inaccurate or misleading results. This is commonly referred to as "hallucinations." This happens more often when the user’s input lacks specificity. For example, ChatGPT might generate incorrect quotes from interview transcripts or present information out of context. It also struggles to interpret nuances like sarcasm or humour in interview responses, potentially misrepresenting what participants actually meant.
I mentioned earlier that GPTs are able to identify themes and patterns in a transcript. While this can be helpful, I’ve found the results to be inconsistent or even insufficient at times, largely because AI struggles to detect nuance. As a result, when conducting thematic analysis, I manually review the transcript, familiarise myself with the material, and identify the themes myself. Only after identifying these themes do I use GPTs to validate my findings. This approach significantly reduces the likelihood of AI “hallucinations.”
My advice for using AI
Based on my experience using AI in research, I would strongly advise fellow researchers to always verify the accuracy of the results. It’s easy to rely on AI for much of the legwork, but it’s crucial to remember that tools like ChatGPT have limitations and still require human oversight.
Additionally, researchers should be mindful of data privacy. Avoid entering or uploading personal details into GPTs, as your prompts may be used to further train their models. Some AI tools also include an ethics page outlining their compliance with regulations like GDPR, so it’s worth reviewing this information to ensure ethical use.
Final thoughts
I’ve found AI to be an invaluable tool in my research, particularly for handling complex datasets, organising information, and streamlining time-consuming tasks. However, it does have its limitations, such as its inability to detect nuance. These challenges highlight the importance of adopting a supervised approach when working with AI tools to ensure accuracy and reliability.