The following op-ed was written by Dr. Jeff Schwartzentruber, an industry fellow with the Catalyst Fellowship Program, and was first published in the National Post on July 7, 2023.
We’re spanning the gap from weak to strong AI — and we’re outpacing ourselves.
The Collision tech conference that took place in Toronto last week was dominated by the topic of artificial intelligence and its potential to transform (or even replace) entire industries. A month before, top scientists from around the globe published a statement with the Center for AI Safety calling for urgent mitigation of the technology on account of the very real threat of extinction from AI. They maintain that safeguarding AI technology should be a global priority alongside other societal-scale risks such as pandemics and nuclear war.
There’s no doubt AI has everyone’s attention, whether good or bad.
For those involved in AI research, none of this recent worry is surprising. With the continuation of Moore’s Law (which roughly states that computing performance doubles every two years) and the inevitable increase in data generation, researchers have predicted that the computational capability will exceed the entire human race by 2040. Are we really outpacing ourselves?
So what makes these recent developments so different? We are now spanning the gap from weak to strong AI — also known as Artificial General Intelligence, or AGI. It is quite an amazing time: if previous AI research was the incubation period, we are witnessing the birth of AGI.
What is worrisome is that the speed and accessibility at which we are transiting from weak to strong AI leaves us unprepared for the societal implications and puts AI experts on edge about the future.
Since the debut of ChatGPT only a mere eight months ago, we have seen the enablement of some truly amazing technologies including AI-generated music (and its associated infringement lawsuits), AI generative videos, full-length podcasts between AI-generated radio personalities (Joe Rogan and Sam Altman), and other large language models (LLM) such as Google’s Bard or Meta’s LLaMA.
Nothing seems to be off-limits — even extremely unethical military applications (feel free to join the petition). ChatGPT has unlocked something in the tech sphere and we are blazing ahead with almost careless disregard.
So what is wrong with our current direction, considering the amazing strides we are making in AI-generated content? Up until this point in time, AI technologies were purposefully built for a specific task, like identifying objects in a picture, or predicting an individual’s credit risk. Due to the limited scope of the problems that the technology was solving, the intricacies and limitations of the model could be well understood, documented and controlled. For these weaker models, AI researchers like myself have a plethora of tools from which we can evaluate and understand the implications.
But now, models have expanded beyond human comprehension. When talking about the size of machine learning models, we typically talk about the various “settings” that can be adjusted during training. In a more formal sense, these settings are called parameters. ChatGPT-4 has more parameters than 100 times the number of stars in our galaxy (100 trillion) making it nearly impossible to understand. Even prior to these most recent advancements in AI, there were several open and important mechanisms of the modern AI system that were not well understood mathematically.
For example, ChatGPT has been reported to have hallucinations and confabulations (terms that have been unrightfully co-opted from psychology) where the model generates false information — in other words, unknowingly lies. These new terms are in their infancy and have not yet undergone the rigorous scientific analysis expected from the AI research community.
The sheer size of these models makes them robust to a variety of applications, such that they can reason and complete tasks they were not designed for. This type of unexpected and emergent behaviour also makes AI researchers anxious, as their inventions can perhaps be used for unintended and malicious purposes. (Nuclear fusion, anyone?)
It is these extensions of new technologies (and their associated threats) that make even the most notable of AI researchers fearful of its current direction.
One recent and poignant example of this is Snapchat’s release of its AI chatbot to all of its customers, including children. This raised ethical concerns regarding AI chatbots’ impact on the developing minds of children and teens with no real understanding of the potential consequences or implications.
Similar to its technological predecessors, the use of these next gen AI technologies easily extend themselves into the realm of cybersecurity. ChatGPT and other LLM models have already been used to create malware, are capable of crafting high-fidelity phishing emails and have significant potential to damage our democracy through voter manipulation via disinformation campaigns, akin to widespread social engineering.
So where do we go from here? Sadly, there is no easy answer. These next gen AI systems have an immense capacity for doing good and will undoubtedly change our societies for the better, but their inherent dangers must be acknowledged.
My current work as a research fellow with Rogers Cybersecure Catalyst at Toronto Metropolitan University leverages and enhances machine learning technologies to automatically detect malicious cyber activity and to protect citizens and organizations from cyberattacks. As a researcher and a developer, how do I walk the line of pushing the boundaries of the field while safeguarding my contributions from unwanted scenarios?
Typically, scientists would rely on government agencies to issue some form of regulatory compliance, but this outcome looks bleak considering the speed at which the government can mobilize such activities versus the speed at which AI technology is advancing.
Due to the general lack of any real legal frameworks, policies or standards in AI research, one such solution is incorporating AI research more closely within the academic research ethic boards (REBs). REBs are commonplace within the academic community and are used to provide oversight on research activities to ensure they are enacted in a responsible, transparent and ethical manner. However, these boards are typically reserved for studies that involve humans and animals. Enacting such an approach within industry could perhaps reduce the risk of malicious use of AI and continue to promote the benefits and possibilities of such revolutionary technologies.
It’s indeed a daunting time for humanity. We need the scientific community and governments around the world to work together quickly to regulate these technologies before it’s too late.
Dr. Jeff Schwartzentruber is Senior Machine Learning Scientist at eSentire and a Catalyst Industry Fellow at Rogers Cybersecure Catalyst, Toronto Metropolitan University.