An accountable algorithm comes from knowing your topic
Each algorithm is different and linked to the objectives of its respective system. For this reason, it is difficult to formulate universal guidelines for building a safe and accountable algorithm other than “know the target phenomenon or population and its features and conditions, and take these into account as objectively as possible in the algorithm’s parameters and their weightings”.
The more complex the target, the more parameters an algorithm has. In some very narrow AIs, there can be just ten or so parameters, whereas for example the GPT3 model, which ChatGPT is based on, is known to have around 175 billion parameters. That alone indicates that the rationale behind the individual outputs of ChatGPT cannot be fully explained or made transparent by human efforts.
In fact, ChatGPT is a typical “black box”, a type of system we’ll discuss more closely on the next page.