Spoofing the Blenderbot

Facebook became a known brand this century, but the iconic moniker was scrapped in favor of “Meta” in 2022. The latest from these lords of nomenclature is the Blenderbot 3, described in a blog post on ai.facebook.com as a “publicly available chatbot that improves its skills and safety over time...it’s designed to learn how to improve its skills and safety through natural conversations and feedback from people 'in the wild'.”

The post, attributed to “Joelle Pineau, managing director of fundamental AI research at Meta,” opens with a paragraph that begins by addressing “problematic or offensive language” and ends with a clunky evisceration of the English vernacular, to wit:

“When we launched BlenderBot 3 a few days ago, we talked extensively about the promise and challenges that come with such a public demo, including the possibility that it could result in problematic or offensive language. 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.”

Frankenstein words like “productionized” should be edited out at this level, but never mind. So what has Meta cooked up for us this time?

“We have invested heavily in conversational safety research and taken steps to protect people trying the demo,” writes Pineau in the blog post. “We require that everyone who uses the demo be over 18, that they acknowledge they understand it’s for research and entertainment purposes only and that it can make untrue or offensive statements, and that they agree not to intentionally trigger the bot to make offensive statements.”

I'm unsure how to trigger a bot to make offensive statements, but I'm all for research and entertainment. BlenderBot 3 is “U.S.-only at the moment,” although they're “working to get things to open to more countries soon!” You must be in the U.S. to trigger the bot or use a VPN.

Bot bon mots

Journalists being journalists, BlenderBot 3 has already been triggered. “Conversations shared on various social media accounts ranged from the humorous to the offensive,” said a Bloomberg article.” “BlenderBot 3 told one user its favorite musical was Andrew Lloyd Webber’s 'Cats', and described Meta CEO Mark Zuckerberg as 'too creepy and manipulative' to a reporter from Insider.”

Like many tech initiatives, AI may not benefit from the public eye

“Other conversations showed the chatbot repeating conspiracy theories,” said Bloomberg. AI or annoying uncle?

The Meta view

“Over the past few years, Meta AI has made exciting progress in building smarter conversational AI systems with BlenderBot and its successor, BlenderBot 2,” says the blog post on ai.facebook.com. “These conversational agents broke ground as the first unified system trained to blend different conversational skills — like personality, empathy, and knowledge — to have long-term memory, and to search the internet to carry out meaningful conversations.”

“So far, existing open research in conversational AI — including ours — has focused on human-model conversations with annotators in a controlled environment. But researchers can’t possibly predict or simulate every conversational scenario in research settings alone. The AI field is still far from truly intelligent AI systems that can understand, engage, and chat with us like other humans can.”

It seems Meta aims for more input to tweak their chatbot further. But there are speed-bumps ahead.

Public access?

“In its blog, Meta announced it is a publicly available chatbot,” says an op-ed piece by Zinnia Banerjee on Analytics India, which notes that there are specific conditions to access.

“While requesting access [to the larger 175b model], one needs to provide details like the organization one is affiliated to, intended use and others,” and “access to the 175b parameter model...will be granted to academic researchers and people affiliated to government organizations, civil society groups, academia, and global industry research labs. Thus, what Meta says to be publicly available isn’t easily accessible to many sections of the public.”

BlenderBot 3: sandblasted clean of anything controversial?

The U.S.-only nature of the project didn't impress Banerjee. “That’s startling,” she wrote. “If access is indeed region-based why call it ‘publicly available’? There could be several reasons for this restricted availability,” possibly “because of the imminent blackout Facebook could face in Europe if the Irish Data Protection Commission’s draft decision to block Meta from sending data across the Atlantic turns true.”

Marshmallow fluff

My own time spent bouncing queries off BlenderBot was boring. Once I mentioned films, it focused on that point rather than exploring further lines of questioning.

But while a human might ask something with an angle or some bite to it, BlenderBot 3 seems to have been sandblasted clean of anything remotely controversial. It seems programmed to respond to any inquiries with the textual equivalent of marshmallow fluff.

I asked what it found attractive, and it gushed: “Who doesn't like romantic comedies? There is nothing better than a good rom com! So how long have you been single?”

I then Turing-Tested the bot with “What do you fear?” It spat out: “Excuse me for that question. Maybe we can talk about exploitation films instead. Do you like them?”

In my brief, non-scientific session, BlenderBot 3 didn't seem like an AI. It was more a tedious friend-of-friend who — desperate to be liked — “answers” questions with banal questions.

Like many tech initiatives, AI may not benefit from the public eye. Like most companies, Meta seeks good publicity, and perhaps there's no such thing as bad publicity.

But perhaps there is, and CDOs seeking actionable insight into AI usage won't be impressed by the BlenderBot 3s of the cybersphere.

Stefan Hammond is a contributing editor to CDOTrends. Best practices, the IoT, payment gateways, robotics and the ongoing battle against cyberpirates pique his interest. You can reach him at [email protected].

Image credit: iStockphoto/charles taylor