What does it mean to “benefit all humanity”?
To benefit all humanity
Probably one of the most important developments in modern times is the boom of NLP through Transformer Models. With the release of ChatGPT, multiple opportunities for implementing AI in our daily lives emerged. However, while these technologies have the potential to make a positive impact on the world, there are still big barriers to consider.
The more I talk to people outside the AI field, the more I realise how fast the field is moving and how scary, difficult, and exhausting it is to be up-to-date with the latest developments. Not only that, but many are even struggling with writing the right prompts to get optimal results. It was documented that the way prompts are written can affect the quality of the outcome after all.
The problem with this goes much further, I believe because even when AI is being used and implemented in new industries, the challenges of AI models remain. These challenges are related to ethical guidelines, sustainability, and fairness. The latter is a challenge that has been difficult to discuss previously as well. It is well documented how difficult it has been to avoid AI models to present patterns of discrimination and reinforce stereotypes, affecting people at a disadvantage. While doing my master’s thesis, I set the challenge to understand this problem better and understand how AGI would benefit all humanity. In this article, I will describe some of the main problems I discovered about AI algorithms and what opportunities we have to better support the vision of benefiting all humanity.
Fairness is relative
Fairness is relative, plain and simple. What is fair to me might not be fair to others. In AI research, multiple authors have proposed different definitions of the term. Some of them focus on getting every individual to get the same outcomes while others focus on group representation. Other metrics focus on measuring false positives and negatives. Many are very innovative and many are equally valid in human terms.
This problem hasn’t been analysed before. As many keep finding ways to dynamically adjust to every fairness metric. Finding variables that depending on the situation, can adapt to the different applications they may have. However, while this can solve some problems it doesn’t solve the fundamental problem that we can always create more and better fairness metrics that will not englobe the whole problem of adapting to everyone’s views.
Fairness is impossible
One of the biggest insights I got from the extensive literature research I made, was to know about, what is called, the impossibility of fairness. This term refers to the mathematical incompatibility between many of these fairness metrics and how it is difficult to know which one is more valid. It comes to be a decision of the implementer.
But then, if these metrics are mathematically incompatible, what can we do to make them be? As some authors suggest, create better metrics or create dynamic metrics. Of course, this would mean that one’s perspective will be more valued than others. How can we then benefit all humanity? How can we ensure that we can adapt to everyone’s perception of fairness? Is it possible to adapt to everyone’s perspective through the current approaches? This opens up multiple questions.
Where does this problem come from?
This problem is not only applicable to AI systems. Research about non-deterministic systems back in the 80s reflected this problem already. Somehow, fairness was used as a way to define whether a process was fair. Meaning that if one process can happen indefinitely amounts of times then it should be able to happen indefinitely amount of times. This definition though, presented the same problems of interpretation. What does indefinitely mean in this context?
This problem evolved later to appear in bandwidth allocation. How can we prioritise the division of bandwidth? What is the fair way to do it? Do we give more priority to streaming services? Do we give an equal amount to every user? Do we give an equal amount to every kind of service, even the ones that are not as used? Such questions developed into multiple papers discussing what is the right way to do it, setting new definitions that reflect the same problem we are facing today with AI systems: equally valid definitions that are not mathematically compatible.
How it is being done today
While these incompatible metrics are used today, they still present the same problems. Measuring has been automated and models can run multiple metrics at once to present an outcome. Many of these are related to what is called Justice as Fairness, a justice theory by John Rawls.
Justice as fairness presents two principles: 1) Fair equality of opportunity, meaning that offices and positions of power should be open to everybody, and 2) Difference principle, meaning to regulate inequalities by permitting inequalities that give better opportunities to those who are worse off.
This theory applies to AI fairness metrics as the equality of opportunity is measured in multiple ways through the existing metrics. By acknowledging who gets the worst outcomes models and data are adjusted to be more ‘fair’. While this theory presents points that are important to consider when developing fair systems they also present flaws. These flaws are pointed out by the Capabilities Approach.
A new opportunity through Capabilities
The Capabilities Approach, defined by Amartya Sen, is a theory that focuses on the capabilities that humans have rather than the freedom to do so. Such a theory has been commonly discussed as opposing and complementary to Justice as Fairness.
If we can apply the criticism of Justice as Fairness to current fairness metrics, we would realise that the problem is not about what outcomes people get from algorithmic systems but rather about what capabilities they are prevented from acting upon. If I send my CV to be analysed by an AI model developed in a geographical part of the world, maybe the algorithm would view me as less appropriate due to the difference in words. Fairness metrics may not always catch that as they may see me as an oddity or a percentage of false negatives that is needed to benefit the majority. The problem is then, is benefiting the majority benefiting all humanity? Instead, if we focus on the capabilities we can provide to humans, a different CV would not be seen as an oddity but instead analysed from a perspective of what capabilities a new opportunity can provide to me. Of course, this becomes much more complex to measure and it would require the analysis of qualitative and quantitative data. It requires the systems to go much beyond a single decision point and instead have a more holistic view that expands the scope of analysis.
A practical implementation?
While this problem is of course very difficult to operationalise. It requires us to think outside the box. Current NLP models are much more advanced and for sure help us solve some of these problems.
Currently, multiple vendors are focusing on developing big language models with millions of tokens that contain a large amount of knowledge with which they were trained. While such an approach has shown its benefits in making AI more widely used, many are still focusing on making these models bigger to make them much more ‘intelligent’ or better performative.
A new strategy to expand NLP models is to use RAG systems. Meaning that we can access a pool of knowledge that the model can interpret and use to give more relevant answers. If we make these knowledge bases applicable to every capability we want to provide to humans then we may have a different opportunity to implement fair algorithmic systems that can benefit all humanity. The challenge next would be then to have enough diversity in these knowledge bases so different cultures are represented and provided with equally valid capabilities.
As long as the leading companies do not share their approach to fairness and ethics, how can we trust that AGI will benefit all humanity?
Conclusion
Much of the text written above comes from the work I made for my master’s thesis at Aalborg University in Copenhagen. Having a practical implementation for this approach requires more time. If you are interested in discussing this topic, feel free to reach out, I am looking for collaborators who can help me write academic papers that lead this approach and have a practical implementation with ai-textbooks.com. If you want to read the master’s thesis, you can download it here.