-from HKUST Libraries
AI tools can be powerful additions to your research toolkit. However, they should be used thoughtfully, cautiously, in accordance with academic integrity and institutional policies, in conjunction with other resources like literature databases, and with human judgment and values at the forefront. According to some of the resources below, some major considerations when using generative AI chatbots for research include that they:
- “Hallucinate” (e.g. create plausible sounding but not necessarily correct material)-they are designed to give you the statistically most likely next words or output - prioritizing fluency over accuracy (the latest and paid models are better but not perfect). FACT CHECK!
- Consider how much information there is available in the training data on your topic. The more niche or unrepresented the information is in the data, the more likely hallucinations are to occur
- Tools like Perplexity, more complex GenAI "reasoning" models, and research tools based on the research literature (like Scite) often use retrieval augmented generation, in which the chatbot does a search and then bases its response on the results (a defined set of papers or links with actually existing references); this can reduce but doesn't eliminate hallucinations.
- Don't "know" about the real world and can lack non-digitized, subscription paywalled, and current information (especially if not connected to web search and/or literature databases)
- Are often a black box (even developers are sometimes unsure why they do what they do) that lacks transparency or reproducibility (important in scholarly work) and contain the biases of their human creators
- Change fast and can vary greatly in results depending on which tool you're using, whether you're using the free or paid versions, what model you're using, what data the tools are based on, how the tool operates, the frequency of updates or availabiity of web search, what you're asking it to do, how you talk to or prompt them, and more.
