sdecoret - stock.adobe.com
Meta's release of AI chatbot BlenderBot 3 raises concerns surrounding the responsibility enterprises have when using the internet as training data for their AI systems.
On Aug. 5, the technology giant released BlenderBot 3 in a blog post, claiming the AI chatbot, which was built on Meta's Open Pretrained Transformer, or OPT-175B, language model, can search the internet to converse on any topic. Meta also said in its post that, as more people interact with the system, it uses the data to improve the chatbot.
The internet wasted no time in pushing BlenderBot 3 to its limits. Headlines claimed the chatbot was not only antisemitic, but also pro-Trump, and that it spewed out conspiracies about the 2020 election. Other headlines show the chatbot bashing Meta and its CEO.
The onslaught of negative press and headlines led Meta to update its blog post on Aug. 8, saying BlenderBot 3's flaws are part of its strategy.
"While it is painful to see some of these offensive responses, public demos like this are important for building truly robust conversational AI systems and bridging the clear gap that exists today before such systems can be productionized," wrote Joelle Pineau, managing director of fundamental AI research at Meta.
Meta did not respond to TechTarget's request for comment.
Using the internet as training data
Meta's statement on public demos is both correct and incorrect, said Will McKeon-White, analyst at Forrester.
"People are extraordinarily inventive when it comes to language," he said. "Understanding things like metaphor, simile is very hard for [bots] to understand, and this might help with some of that. You do need a lot of data to train, and that is not easy."
Will McKeon-WhiteAnalyst, Forrester
"If you know what happens, then you should have taken additional steps to avoid it," McKeon-White added. "Social media does not provide a good training set, and having it open and adaptable to people on social media and learning from them is also not a good training set."
Meta's BlenderBot 3 is reminiscent of Tay, an AI chatbot released by Microsoft in 2016. Similar to BlenderBot 3, Tay was also cited for being misogynistic, racist and antisemitic. The controversy surrounding Tay caused Microsoft to shut it down a few days after the system was released on social media.
Finding other training data
Since AI chatbots, like BlenderBot 3 and Tay, are often trained on publicly accessible information and data, it shouldn't be surprising when they spit out toxic information, said Mike Bennett, director of education curriculum and business lead for responsible AI at the Institute for Experiential AI at Northeastern University.
"I just don't know how large tech companies that are investing in these chatbots are going to, in a way that's economically rational, train these devices quickly and efficiently to do anything other than speak in the mode of the sources that trained them," Bennett said.
Smaller enterprises and businesses can find other training data, but investing in the development of a curated data set to train chatbots -- and the time -- would be expensive.
A less expensive option could be for multiple smaller organizations to pool their resources to create a data set to train chatbots. However, this would cause friction since organizations would be working with competitors, and it might take time to figure out who owns what, Bennett said.
Another option is to avoid releasing systems like these prematurely.
Brands working with natural language generation (NLG) must keep a close eye on their system, maintain it, figure out its tendencies and alter the data set as needed before releasing it to the world, McKeon-White said.
Caution surrounding NLG systems
Due to BlenderBot's performance, there will likely be caution around NLG systems, McKeon-White said.
"This will probably tamper experimentation with it for a little bit," he said. This will last until providers can provide filters or protections for systems like these.
BlenderBot 3 also raises the bar for those considering AI avatars for the metaverse, Bennett said.
"We really need to see positive developments that significantly reduce the instances of these kind of vile engagements before we get into that space," he said. "Not only will it probably be a more engaging mode of interaction with digital entities, but there's also the potential for combining the kinds of unfortunate utterances that we've gotten over the last month or so from the latest version of the chatbots."