Two major problems with AI-generated content are confabulation and bias.
Confabulation is the creation of false content without the intention to deceive. AI tools, especially chatbots are prone to make up information because they are trying to provide plausible responses to the prompts they receive. For example, if an AI chatbot is asked to provide a literature review for the latest research on dementia, it may confabulate by generating a response that misconstrues information or invents plausible-sounding information, sources, and citations.
Algorithmic bias can occur if there is biased information in the AI algorithms, or if the datasets the AI tool uses contain bias. Common biases that can occur in the data AI tools pull from to generate answers to prompts can include biases against race, class, gender, sexual orientation, ability, religion, or political beliefs.
Because of both confabulation and the potential for bias, it is recommended to evaluate whether the information and citations the AI tool generates are good sources of information. See how to evaluate AI content below.
Accuracy: AI-generated content is not always accurate. It may contain errors, false claims, or plausible-sounding content that is invented and false (confabulations). AI tools may be limited by the dataset available to them, which may not include the latest information. Consider that subscription-based resources, such as library databases may offer more authoritative sources.
Fact checking is very important when using AI tools like ChatGPT!
Bias: AI-generated content may contain implicit or explicit biases introduced through the algorithmic design, data collection, data labeling, or model training processes. It’s important to evaluate AI-generated content for common types of bias, such as racial bias, gender bias, class bias, sexual orientation bias, disability bias, religious bias, and political bias.
Currency: AI-generated content may contain information that is outdated. This may result from access to old or limited datasets. For instance, some free GPTs only have access to a snapshot of the internet from several years ago, so they are unable to generate content that draws from the latest information.
Comprehensiveness: AI content may be selective as it depends on the algorithm which it uses to create the responses, and although it accesses a huge amount of information found on the internet, it may not be able to access subscription-based information that is secured behind firewalls. Content may also lack depth, be vague rather than specific, and it may be full of clichés, repetitions, and even contradictions.
Copyright: AI-generated content may infringe on copyrighted material. This is because AI tools often draw upon datasets that include copyrighted material, so new content created from these datasets may inadvertently violate copyright provisions.
Citations: AI-generated content may often lack citations to back up the information provided. If citations are included, they may be invented or misconstrued. Always double-check that citations are real and that the information provided is accurate.
Follow us on Facebook
Follow us on X (Twitter)