Forget AGI if AI models stay inconsistent with jagged skills: Bengio

Yoshua Bengio says AGI may be irrelevant if AI systems continue showing jagged intelligence - excelling at complex tasks but failing at simple ones - as he warns of rising risks, unsafe capabilities

Joshua Bengio, the Canadian computer scientist often hailed as the godfather of artificial intelligence (AI), on Friday said it would be better to forget artificial general intelligence (AGI) as the next big thing if modern systems continue to have jagged intelligence – extremely good at many complex tasks but really bad in many simple ones.

 

“We can escape this vision of an AGI moment if AI continues to have these jagged capabilities. We could well be in a world where AI already has these dangerous capabilities and that is why there is a need for careful scientific evaluation to determine the risk and capabilities of these machines,” he said on the final day of the India AI Impact Summit.

 

Academicians, researchers, and policymakers have cautioned against the AI juggernaut, saying while its prospects remain exciting in many fields – especially in areas of scientific research and drug discovery or curing cancer – it is also prone to hallucinations, biases, and botched outcomes.

 

“Maybe a few years ago, it would have been exciting to think of AGI reaching human levels, but now it is meaningless as we are going to have things that will be extremely stupid and weak in some ways and dangerous in the wrong hands. Businesses also need to determine whether AI is good for what they are trying to do,” Bengio said while launching the International AI Safety Report.

 

His comments come days after Google DeepMind Chief Executive Officer (CEO) and Cofounder Demis Hassabis said the era of AGI is still about five to eight years away, and the systems need to be trained further and possess capability to learn on their own.

 

While general purpose AI capabilities have continued to improve — especially in mathematics, coding, and autonomous operation — some of the dangers of these models and systems are already out in the open. Deepfakes are on the rise along with fraud and scams. Bengio expressed particular concern over the amount of AI-generated non-consensual intimate images, which affect women and girls — something that brought Elon Musk’s Grok into public attention last month. He also warned that closed models are not immune to attacks just like open source ones, and that the guardrails are still not up to the mark.

 

The AI Safety Report added that AI systems can discover software vulnerabilities and write malicious code. “In one competition, an AI agent identified 77 per cent of the vulnerabilities present in real software. Criminal groups and state-associated attackers are actively using general-purpose AI in their operations. Whether attackers or defenders will benefit more from AI assistance remains uncertain,” it said.

 

The report also said that general-purpose AI will likely automate a wide range of cognitive tasks, especially in knowledge work, with early evidence showing no effect on overall employment, but some signs of declining demand for early-career workers. Bengio also expressed his concern that governments are not doing enough to address the issue of job losses due to AI.