ChatGPT and policies during elections

[Biplav Srivastava]

This is courtesy Prof. Anita Nikolich.

https://openai.com/blog/how-openai-is-approaching-2024-worldwide-elections

"ChatGPT will direct users to CanIVote.org, the authoritative website on US voting information, when asked certain procedural election related questions—for example, where to vote."

Essentially, the smartest chatbot vendor in the world will punt questions to a manually created FAQ.

Regards

[Swapneel Mehta]

You raise a fair point but but when the cost of false positives is infinitely high (cue political pundits and US congress calling Altman up for testimony based on viral tweets about "chatGPT lies and says dead people can vote"), isn't it also the most responsible way to do so in the short term, until you're certain hallucinations can be prevented? 

Not bearing the burden of truth is one of the reasons Meta still employs an external network of fact-checkers ("third party fact checking") as opposed to being arbiters themselves.

Best,
Swapneel

[Biplav Srivastava]

Hi Swapneel,

No one is objecting to their approach. But which domain is 'relaxed' that the cost of false positives is low - health, finance, energy, ... ? The common approach is to thrust chatbots down the users throat (like customer support) and hope that they take off! That is the other extreme of 'responsible approach'.

More generally, in every problem one seeks to solve, one needs to know the ideal solution even if they have not met figured out the methods. That is the only way one can create a research vision, the test cases, and hope to ever solve a problem for good.

In practical sense too, the real worth of a technology is when it can solve a problem that matters, satisfactorily. Chatbots missed their opportunity to shine during COVID19 [1].

I believe elections offers another application area for AI (and other technologies) where we can frame the ideal decision-support solution so that we can help the society where needed. The current baseline is manually created FAQs.

 [1] https://www.cell.com/patterns/fulltext/S2666-3899(21)00151-3

[Swapneel Mehta]

Thanks Prof. Srivastava, 

It's a good question: I would be inclined to state it's not by domain but by application that one would define the cost of false positives. Directing a customer call to a wrong helpline number for a financial agency is objectively a lower cost than a hospital calling and telling you incorrect test results. I do buy your premise anyway, though, since I agree that most companies seldom account for the responsible deployment of tools especially in the recent market for AI products. 

The article is nicely drafted and makes a good case for chatbots to have done more than they were able to (cursory reading, didn't dig too deep). In the case of elections then do you mean 'ideal solution' in the sense there is a finite set of facts related to the event that could be offered to the end users directly instead of channeling them through a different endpoint? What would you define as the ideal structure? maybe I'm misreading your intended message, apologies if so.

Best,
Swapneel