In a recent paper researchers at Microsoft evaluated the performance and behavior of GPT-4, the latest large language model from Open AI (Sparks of Artificial General Intelligence: Early experiments with GPT-4, https://arxiv.org/abs/2303.12712).
The results are impressive. Fellow developers in the ML community rightly point out that these models are based on quite simple advancements in the underlying technology. However, the outputs of GPT-4 as witnessed in the wild and presented in the paper, such as the Socratic dialogue (including a discussion of that very dialogue compared to a less well-written one from ChatGPT), are simply stunning.
The authors claim GPT-4 shows signs of general intelligence, “demonstrated by its core mental capabilities (such as reasoning, creativity, and deduction), its range of topics on which it has gained expertise (such as literature, medicine, and coding), and the variety of tasks it is able to perform (e.g., playing games, using tools, explaining itself, …).”
The definition of intelligence used is of course debatable (as the authors acknowledge). True general artificial intelligence (AGI) is still far away. Yann LeCun (e.g. “A Path Towards Autonomous Machine Intelligence”, https://openreview.net/forum?id=BZ5a1r-kVsf) and others propose a more complex architecture including many interacting components like predictive world models and intrinsic motivation as prerequisites for AGI.
Nevertheless, the consequences of ChatGPT and other large generative models cannot be overestimated. As Bill Gates puts it: “The Age of AI has begun“.
Increasingly powerful AI systems will be developed; ‘superintelligence’ might be just a matter of time. Now is the time to think about how the responsible use and development of these systems should look so that AI benefits all people.
As the authors of the above-mentioned paper state, “It is easy to use GPT-4 to generate biased outcomes”. And bias, clearly, is just one of the many problems that we face when dealing with such powerful models.
We at Daiki aim to provide guidance on responsible AI. Our team of experts in machine learning, philosophy, law, and design helps developers build trustworthy AI systems.
We do not want to simply warn of the potential dangers of AI, certainly not block. We are not a team of doubters. AI is a fascinating technology that will transform the world we live in. Let’s use AI for good.
In this recurring series, we will present examples of how AI systems can be used to benefit people. There are many cases where it’s indeed ethical to use AI, for example, to increase the likelihood of saving lives with systems that can reliably detect diseases at their early stages, or where AI can be more objective than humans, e.g. to make decisions in sports.
AI referees in sports
Let’s start with the latter area and discuss the role of AI referees. These are certainly not life-and-death decisions (some sports fans may disagree), but they are nonetheless very important and consequential. After all, sports are about fairness, and adherence to the rules is central to trustworthy, competitive, and entertaining amateur and professional sports. Referees are supposed to guarantee fairness. In this sense, they bring applied ethics to sports.
The FIFA world cup 2022 relied on machine learning to detect offside. This AI referee automatically determines if an offside player touches the ball, this decision is then presented to the video match officials and confirmed by the human referee team.
Offside is a very important rule in (professional) football that fundamentally affects offensive play; it is not uncommon for close offside situations to lead to clear scoring opportunities. Referee calls about this rule therefore can (and do) decide matches – with direct mental and financial consequences for athletes, club managers, and fans.
According to FIFA the new AI technology “was able to support the video match officials by helping them to make more accurate and more reproducible decisions in a shorter period of time.” Decisions by the AI referee are visualized in 3D animations and presented to both the (video) referees and the spectators.
This is clearly a very good example of how AI can be embedded in a critical process to achieve better, more reliable, and more transparent results. It’s also a good example of how to use AI responsibly: the human oversight (video referees evaluating the AI referee decision) and 3D presentation elements (transparency for all) provide trust and confidence.
As a consequence in professional football leagues that use technology-assisted offside decisions, this rule – in stark contrast to other video-supported decisions like illegal handball – almost eliminated debate, and decisions are accepted by players, team officials, and fans. Of course, it helps that offside is a very clear rule where only the positions of the players, the position of the ball, and the timing of the pass matter. But that just means that for other rules and other sports, there are more technical challenges to be solved.
As long as the AI systems are robust, reliable, and used transparently they can help make (professional) sports better and more equitable.