Those that have tried ChatGPT will know that it sometimes returns false information – it simply makes it up. So what is the solution, and do we need to open-source the truth?
When I asked ChatGPT who I was, it credited me with writing books I hadn’t and winning awards that didn’t exist.
So what is the solution? As more and more people (over 100 million at last count) start to use Generative AI systems, what can we do about misinformation present in these tools?
A few weeks ago, I published a podcast I recorded with Stephanie Antonian, and one of the suggestions that she had was that we open-source the truth.
Her idea is simple.
There are two “lists” – one with lists of scientific truths (eg the Earth is part of the solar system) and one that’s Law and Social truth (eg Murder is illegal).
Two separate entities maintain the lists, and each query from a generative AI system is fact-checked against these lists, and they are also used to train these systems.
Quoting Stephanie from the podcast:
“So we can fix generative AI, and we can fix the platforms if we create a backbone to the systems, where we say, this is what we know, this is what we don’t yet know, and this is what we can’t know. That’s where I think we’ve got to collaborate a little bit more because it shouldn’t be the tech companies, they don’t have to do everything.”
Would this idea work?
At last week’s Cognito event where I spoke about the future of communications in AI, I was asked about fact-checking and correcting mistakes generated by AI tools by an audience member.
I tested Stephanie’s idea out on the audience, and it resonated.
It has been said that Generative AI systems can confidently assert falsehoods, and as humans, we need to have this called out so we don’t unintentionally spread misinformation.
This is where regulation comes in, but running these “lists” could be problematic across multiple languages and jurisdictions.
There is a precedence, though. When you type in “linkedin.com” to your browser, your computer interrogates a Domain Name Server or DNS to translate the web address to an IP address.
This has been happening for years, and has been dubbed the “phonebook of the Internet.”
Early DNS were run and managed by individuals. Now The Internet Corporation for Assigned Names and Numbers (ICANN) operates servers for one of the 13 IP addresses in the root zone and has delegated operation of the other 12 IP addresses to various organisations, including NASA, the University of Maryland, and Verisign.
So one of the key components of how the internet functions on a daily basis has, in effect, been put in the hands of multiple entities.
What do you think – could this work?
Listen to the original podcast here.