This week, the Federal Bureau of Investigation allegedly used artificial intelligence (AI) to get instructions to create a bomb, two men suspected of bombing a California fertility clinic last month. The FBI did not disclose the name of the AI program in question.
This sharply focuses on the urgent need to make AI safer. We currently live in the “Wild West” era of AI. There, companies are fiercely competing to develop the fastest and most interesting AI systems. Each company wants to outperform its competitors and charge a top spot. This fierce competition often leads to intentional or unintended shortcuts, especially when it comes to safety.
Coincidentally, around the same time as the FBI revelation, one of the modern AI godfersers, Joshua Bengaio, a Canadian computer science professor, launched a new nonprofit organization dedicated to developing new AI models designed to be safer than other AI models, targeting those that cause social harm.
So, what is Bengio's new AI model? And will it actually protect the world from the harm that is enveloped in AI?
“Honest” ai
In 2018, Benguio, along with her colleagues Jan Lekun and Jeffrey Hinton, received the Turing Award for his groundbreaking research he presented in deep learning three years ago. Machine learning branches, deep learning, attempt to mimic the processes of the human brain by learning from computational data using artificial neural networks and making predictions.
Lawzero, Benguio's new nonprofit organization, is developing “Scientist AI.” Benguio states that the model is “honest and not deceptive” and incorporates design-by-design principles.
According to a preprint paper released online earlier this year, scientists AI differ from current AI systems in two important ways.
First, you can assess and communicate your level of trust for that answer, which helps reduce the problem of AI giving you an overly confident and false response.
Second, you can explain that reasoning to humans, and you can evaluate and test that conclusion.
Interestingly, older AI systems had this feature. However, in the rush for speed and new approaches, many modern AI models cannot explain their decisions. Their developers are sacrificing explanability for speed.
Bengio also intends that “scientist AI” act as a guardrail against dangerous AI. It allows you to monitor other, unreliable and harmful AI systems. This can basically fight fire with fire.
This may be the only viable solution to improve AI safety. Humans cannot properly monitor systems such as ChatGpt, which handles 1 billion queries every day. Only another AI can manage this scale.
Using AI systems over other AI systems is more than just a science fiction concept. This is a common practice in research to compare and test different levels of intelligence in AI systems.
Adding a “World Model”
Large-scale language models and machine learning are just a small part of today's AI landscape.
Another important addition that Bengio's team adds to scientist AI is the “world model” that brings certainty and explainability. Just as humans make decisions based on their understanding of the world, AI needs similar models to function effectively.
It is clear that there are no world models in current AI models.
One well-known example is “hand problems.” Most AI models today can mimic the appearance of the hand, but they cannot replicate the natural movement of the hand.
Another example is how models like ChatGpt are fighting chess, unable to win and making illegal moves.
This will even hit the best human players, despite a simpler AI system that includes models for the chess “world”;
These problems arise from the lack of a fundamental world model for these systems. They are not essentially designed to model real-world dynamics.
Alex Wong/Getty Images
Get on the right track – but it's bumpy
Bengio is on the right track and aims to build safer and more reliable AI by combining large-scale language models with other AI technologies.
But his journey is not easy. Lawzero's US$30 million funding is less than efforts such as the US$500 billion project announced earlier this year by US President Donald Trump to accelerate AI development.
What makes Lawzero's tasks difficult is the fact that scientists, like other AI projects, require a powerful, vast amount of data, and most of them are managed by major high-tech companies.
There are also some notable questions. Even if we can build an AI system that can do everything Bengio can do, how can we control other systems that could be causing harm?
Still, the project, along with the talented researchers behind it, could trigger a move towards a future where AI truly helps humans thrive. If successful, it could set new expectations for safe AI and motivate researchers, developers and policymakers to prioritize safety.
Perhaps if social media had acted similarly when it first appeared, there would be a safer online environment for the mental health of young people. And if scientists had already been introduced, it could have prevented people with harmful intent from accessing dangerous information with the help of AI systems.
