So what should we call AI: A-ethical, non-ethical or unethical?

Last month during my presentation at the TINTech London Market conference, one of the terms I used was the concept of unethical AI, which I briefly defined here.

In a nutshell, I was demonstrating that by calling it ‘unethical AI’, instead of ‘a-ethical’ or ‘non-ethical’ artificial intelligence, we emphasise the idea that until humans are no longer the main trainers of AI, it may well be the case that ‘bad behaviours’ are explicitly or implicitly caused by human trainers.

This is in contrast to an idea where ‘a-ethical AI’ vs unethical, suggests a distinction similar to that between amoral and immoral.

However, a recent LinkedIn post about generative AI made me want to revisit this analysis of amoral vs immoral, or a-ethical vs unethical AI.

The post details the process of using the ‘deep research’ mode on an AI powered engine, which led to initially very promising and accurate results. This was until the final step, where the AI model not only made up a numerical figure it felt “sounded legit”, but it also ‘joked’ about the fact it had made the number up.

Now, this clearly presents a very real problem that hallucinations are causing in AI models. Because the rest of the data was accessed, presented and summarised in a very effective, clear and seemingly accurate manner, these types of models will be increasingly used.

However, if there is no way of knowing which numbers or figures have been fabricated (unless each ‘source’ is manually checked and verified by the user, also since models often interpret information from sources incorrectly or out of context), this will undoubtedly prove to have problematic consequences in the future.

Furthermore, the most surreal part of the AI interaction was that the model only admitted to the fabrication once asked to talk through its thought process.

Nonetheless, the AI is clearly not intending to be malicious here. Instead, these models would prefer to ‘make up‘ an answer (which may likely be partly/very wrong), rather than leave a gap. They have not been trained that in certain situations, for example when using AI models for research, it would be more appropriate that they didn’t provide specific answers for the information they were not certain of.

An analogy of this scenario would be a home food shopping company giving you a poor substitution, rather than no product. Here, the difference is that you can easily spot the situation and assess its suitability, while a ‘fake’ or fabricated number or statement in a document may not be as obvious.

AI models are trained on a stochastic basis (a concept which I am still trying to fully wrap my head around but will briefly try explain here), meaning they will always give what they consider to be the likely answer. Basically, AI models are learning from what they have previously been trained with, and therefore what will be the most likely answer.

For example, from looking at all the words in the English language, if you have a word with the letter Q it is almost guaranteed that it will be followed with the letter U (if the word is of english-origin and not a loanword).

Therefore, you can start to build a model of the probability of which letters come next based on the ones proceeding.

This process of learning can also be demonstrated through AI generated imagery. AI models learn though training images that cats have two eyes the majority of the time, and therefore most of the time if asked to make an image of a cat show one with two eyes.

This seems logical. However, while we have concluded that through hallucinations the AI is not acting maliciously, there is another ethical problem that arises through this - and this one may through our concept of ‘a-ethical AI’ into turmoil.

AI models are being trained on training images which are part of an internet ecosystem which is plagued with gender-based bias, sexual and racial bias, political bias, and many more forms of prejudice-based influences. Use of these forms of information will lead to AI models that have inherent biases.

Thus, if systems are being trained on these types of images and sources, it provides significant basis for arguments that technology is in fact not neutral.

This leads us to a very daunting but fundamental and essential question: how do we train AI in a way that avoid these prejudicial influences and create models that are truly neutral and a-ethical?

Tackling and confronting these questions and issues is at the heart of what these articles and this publication aims to answer.

Read more articles from ‘A Philosophy student’s take on Ethics of AI’ by subscribing to the Substack here!