You Ask, I Answer: Reducing Generative AI Hallucinations?

Today's episode of You Ask I Answer was recorded in front of a live studio audience at the Digital Now conference in Denver, Colorado, in November 2023. The session title was appropriately, You Ask, I answer live. Generative AI QA. Enjoy. Other questions?

Yes. It's fine. Someone else can go. Everybody's having right now I have a staff member who is taking content from content specialists and putting it into Chat GPT and saying, okay, write me 10 social media posts and advertise a copy, etc. etc.

What I would really love is that it is if we had a mechanism to implement the social posts and say, now schedule these posts, you know, two per month over a three-month period, etc. etc. Where are we at in achieving that? And does it already exist and I don't know when we see it for it? It depends on whose platform you're using.

So Agora Pulse, for example, has that utility where you can it has generative AI prompts and you can say here's what it want and do the thing. I don't know if it fits that exact use case, but it will get you awfully close. The if it can't, almost every platform has bulk import of stuff at some in some way. So what again, what you would do is say, here's this the thing. Make me social content, but you're gonna format it the output as a CSV file.

Column one is the URL, column two is a social post, column three is this, at the date and time I want it scheduled. Process this, provide me a link to download. So that's your prompt. It will spit out the the output that you want, and now you can just bulk load that into your social schedule. Yeah.

Beautiful. What about for the images that would go with those? So most social hosts have images anymore. You know, we all are using that. Um, if it's not, you know, like just the web link that generates the image.

So is there a mechanism for that as well? That with a JPEG or Panglab? It depends on your social scheduler. Um, if your social scheduler can take an image link, then you would put your visual assets on some publicly accessible server and then to provide the reference links to those things. Um that's probably the easiest way to do that.

And the AI tool you're describing could grab the leap on the website and what was the one you mentioned outside of the website. Uh there's one called Agora Pulse. Yes. They're based in France. They they I've been using them for years.

Um all the social scheduling tools right now are are struggling to figure out how to integrate AI because they all want to be able to say that they have it. Most of them are putting some implementation of open AI software in it, but they haven't really figured out yet how to make it integrated into the product. So it's it's that particular part of the industry is still a very nascent space. Yes, here and then here. So it sounds like so many people in the room are well along their AI journey, and I am not.

Um the last session I was in, they mentioned that they had started down the process using chat beats GPC4, got to the end of it, and said this is giving us untrue responses, and we can't make it work. I have to keep my data, I mean it's legal guidance, so I can't risk untrue response. Does that negate the use of AI? It would take you back to pure search, or is there a better tool? Um I would I would use Bing because you'll at least get citations for where it's getting its information, or Google Bard is the other one.

So let's go into BARD here. Um ways that derivative works retain their copyright and the conditions under which a derivative work would lose its copyright, such as a transformative work. Cite relevant cases. So one of the things that BARD in particular has, they just added it not too long ago, is a little button down here called, hey, are you lying? It's called double check my response.

Um but what it does is it then goes and crawls the Google's index catalog and it highlights in green, hey, this is where I found this information. And then this one here, this goldsmith versus Hearst says, I found content that differs. I think I lied. Right? So but this one here, in this case, the Google versus Oracle America, it found a citation that you can then go and check out to make sure it's so the the search-based language models now have some level of, hey, here's where I got this information from.

I would absolutely not use chat GPT for finding relevant data because it just hallucinates. And it's not intentional, it's not malicious. The way it works is it's pulling those word clouds and it finds associations that have the greatest strength and it assembles an answer. In very early versions, when you ask a question like who was president of the United States in 1492, it pulls 1492. What are the words associated with that?

Well, there's like this Christopher Columbus person. It pulls the United States. What are the words associated with that? And the president, well, that's an important person. So it would answer, Christopher Columbus was president of the United States in 1492, even though it's factually completely wrong.

But the statistical associations made that the logical answer. So pure language models like Claude and like ChatGPT's GPT-4 model, they have no fact-checking, right? Whereas the search engine-based ones have some citation. So I would always use that anytime you need to say, where did you get this information? If you enjoyed this video, please hit the like button.

Subscribe to my channel if you haven't already. And if you want to know when new videos are available, hit the bell button to be notified as soon as new content is live.