Most of us have probably witnessed the flurry of news coverage surrounding Google’s Gemini and its blunders in generative imaging – leading incumbent CEO Sundar Pitchai to admit that, ‘I know that some of its responses have offended our users and shown bias – to be clear, that’s completely unacceptable and we got it wrong’. Similarly, Google’s co-founder Sergey Brin acknowledged that ’we messed up’, blaming much of the issue on a lack of thorough pre-release testing. As important as issues of bias and fairness are, one cannot help but ask the question: are missing something when it comes to the evolution of large language models (LLMs)?
A more technical and less anthropological point worth making from the onset is that AI gaffes are part of the development process. Some might think that we are in an AI ‘arms race’ and that developers are rushing too quickly to release products that, by normal standards, are still very much in their Beta phase. The AI feedback loop between the software companies and us as consumers is likely to be with us for some time to come.
It is also important to note that AI systems such as Gemini (currently) lack any degree of aptitude when it comes to intuition – a characteristically distinct human feature. Unless the programmers explicitly tell the LLM that, ‘there has not yet been a black pope in the history of the Catholic church’, or that, ‘there is little evidence to suggest that 10th century Vikings were anything but ethnically white’, it will continue to resort and apply the same default internal prompt set up by the programmers themselves. This is the consequence of a broad-brush, off-the-shelf racial or ethnic parity prescription. The trouble is that within certain contexts, like history, it fails miserably: the Vikings were white, the Zulu were black, the Soviets weren’t Asian, there hasn’t been a black pope, and so on.
The reality is that all LLMs, including Google’s Gemini or Microsoft’s ChatGPT will carry some degree of bias and the key lies in understanding how each is fine-tuned. We are likely to soon be faced with a situation where multiple LLMs display different characteristics: one might be appear more ‘woke’, another ‘anti-woke’ (such as Elon Musk’s self-confessed ‘Grok’ AI). Some may be better suited for artistic purposes while others might prioritise fact-checking and the natural sciences – there is a plethora of combinations and possibilities.
An interesting example of this is Character.ai – an online platform that has recently skyrocketed in popularity where people can create AI chatbots based on real or fictional characters. Examples range from Harry Potter to Vladimir Putin, but what is interesting is that despite the multitude of ‘more exciting’ possibilities, ‘Therapist’ and ‘Psychologist’ regularly feature among the most in-demand. Over 12 million messages were sent to an ‘AI therapist’ on the platform and according to a BBC article, many of its users report a very positive experience.
This, of course, points to the gravity and scale that mental health issues represent amongst young people, but it also points to a willingness to engage with an AI chatbot that is upfront and clear about its characteristics. In other words, people are less likely to be offended and more willing to engage if they are aware of the specific ‘character traits’ that make an AI behave a certain way. Openness in this regard is one of the key moral conundrums that our so-called ‘tech overlords’ will have to overcome if they wish to gather a wider degree of public trust. Fairness in presenting data and transparency in how the system is fine-tuned will represent key areas of ethical concern.
A wider problem is the reality that any interpretation of large amounts of data ultimately carries some degree of bias. It may not appear as glaringly obvious as Gemini’s image generation, but even an infinitesimally small amount of bias represents bias nonetheless. It is also not just limited to historical matters: from healthcare ethics to organisational management and business strategy, data is skewed toward the adopted position of the individual or institution that presents it. The question worth asking, then, is since we, as humans, often cannot agree on what exactly constitutes an ‘unbiased view’ (if beyond objective truths such a feat is even possible), how can we expect an AI chatbot to do so? The answer is that we shouldn’t. Instead, we should concentrate our efforts and strive for greater openness, transparency and a grounding in moral values at the foundational level. These are all but nascent steps in promoting a human-AI symbiotic future that considers both the responsible development of AI as well as human flourishing.
Andrei E. Rogobete is Associate Director at the Centre for Enterprise, Markets & Ethics. For more information about Andrei please click here.