Although AI is new, the question of reliability is not new, and so, in many ways, the reliability or otherwise of AI outputs follows historic principles as to the reliability of information generally. So, when carrying out research in general, and more specifically, when using AI, you should ask yourself these kinds of questions:
If you want to know on which day of the week your 50th birthday will fall, you can find that kind of information out easily online. And whatever you find is likely to be correct. Almost all mentions of that date online will have the correct day listed. This kind of "true or false" information can't reasonably be debated. As a result, any AI tool is likely also to give correct information of this kind. So, AI can often be great for giving you simple facts. However, do consider the ethical issues of using AI for something you could easily find out another way. In addition, even for this simple type of data, AI can often "hallucinate" whereby it produces inaccurate information, like false references.
Computer code
Something else that can fall into this category of "true or false" information is computer code, because it will likely either work or not work. AI can often offer corrections to your code that isn't working. But, you might need to ask it several times, perhaps rephrasing your question, before you get code that works. And ultimately, AI can only be as good at such things as the documentation it is trained on. If there is lots of code in its large language model (or equivalent) that matches what you are trying to fix, you will likely be able to get a correct output. But if the code you are trying to write is unusual or niche, you may have less success.
Legal drafting
In a similar way to the computer code example, some legal clauses are very common. If you are asking an AI tool to draft a common clause for you, it is reasonably likely to be able to do so well (although you would always need to check its accuracy yourself). AI tools within legal databases are even likelier than other AI tools to draft clauses well, because usually the documentation that they are trained on is just the high-quality legal resources from within the databases themselves.
The more advanced research you are to carry out, the greater the likelihood that no-one (or not many people) has researched this specific thing before. For example, usually when writing a PhD thesis, you would be producing some completely new research within it.
If you are asking AI a question that no-one has ever answered before, you will effectively be asking it to carry out brand new research itself. Whereas, if you ask it a question about a very popular topic, it can produce an output based on this already-existing research.
AI cannot actually "research" new topics. Instead, it effectively regurgitates information based on the documentation that it is trained on (e.g. the large language model or equivalent). As a result, if it is being asked to produce brand new research, the likelihood of it producing correct information will be relatively low. So think about how niche the topic is that you are asking AI to research.
A lot of information will fall into this category, since there are often two (or more) perspectives on any given topic. Normally, when considering biases that writers might have on a topic, it is possible to have a reasonable idea what these biases might be. For example, if thinking about UK newspapers, writers in the Daily Telegraph are likely to look at news stories from a different perspective to writers in the Guardian. So, when considering the reliability of written work, you would always have to consider who has written it. This is the case, not just with newspapers, but with academic publications too. It is not uncommon for there to be competing ways of viewing academic fields. So, it is always good to try and be aware of what these competing views might be, and to try and ascertain which perspective the writer whom you are reading might be most aligned to.
The problem with AI outputs
Very often AI outputs do not inform you what sources have been drawn upon to produce the output. In which case, you will not be able to consider which potential biases the information might include. This makes assessing the quality, and in particular any biases involved, very hard to do with AI-generated outputs.
The solution
You must not rely solely on AI when carrying out research on a topic. Instead, use it only in conjunction with more traditional research, in which you can be aware of who has produced the information, and so you can investigate any potential biases they might have. Search Everything on the Library website is often a very good starting place. And there are also many specialist databases for each School at BPP. We have video guides on using these on this page.