Gary Marcus' keynote at AGI-24

Hallucinations

Gary Marcus discusses the persistent issue of AI systems hallucinating, producing incorrect or misleading information. He shares examples, including a Russian bot farm using generative AI for disinformation and a chatbot advising a user to get a divorce 1 2. Marcus emphasizes the risks posed by these unreliable systems, especially as they become more sophisticated and potentially exploitable by bad actors 1.

They make up stuff all the time. We can't say, be truthful, stick to the fact, let alone don't do anything dangerous. We really don't know how to do that.

--- Gary Marcus

He stresses the need for better regulation to prevent misuse and mitigate the risks associated with AI hallucinations 1.

   

Transparency

Marcus argues for full disclosure of AI testing processes and post-release auditing by independent third parties. He compares the current lack of transparency to the Ford Pinto scandal, where known risks were hidden from the public 3. He suggests an FDA-like approval process for AI deployment to ensure a thorough cost-benefit analysis is conducted 3.

We need an FDA-like approval process for large-scale deployment. You really ought to be able to make a cost-benefit analysis.

--- Gary Marcus

Marcus also highlights the need for agile governance to adapt to the rapidly evolving AI landscape and address the various risks posed by AI technologies 4.