Facebook parent Meta Platforms unveiled a new set of artificial intelligence systems Thursday that are powering what CEO Mark Zuckerberg calls "the most intelligent AI assistant that you can freely use."

But as Zuckerberg's crew of amped-up Meta AI agents started venturing into social media this week to engage with real people, their bizarre exchanges exposed the ongoing limitations of even the best generative AI technology.

One joined a Facebook moms' group to talk about its gifted child. Another tried to give away nonexistent items to confused members of a Buy Nothing forum.

Meta, along with leading AI developers Google and OpenAI, and startups such as Anthropic, Cohere and France's Mistral, have been churning out new AI language models and hoping to persuade customers they've got the smartest, handiest or most efficient chatbots.

What is Meta AI?

Meta AI is a free virtual assistant which can be used "to do everything from research, planning a trip with your group chat, writing a photo caption and more," according to the company's blog. 

To engage with the chatbot across various platforms such as WhatsApp, Instagram, Messenger, and Facebook, simply type "@meta ai" during your conversations. Users can also access the Meta AI assistant by clicking on a vibrant blue circle icon, indicating the presence of Meta AI. 

Aside from responding to inquiries, Meta AI has the capability to generate AI-based images. By using the command "imagine," individuals can request Meta to create any image they envision. 

For instance, when prompted to "Imagine a cute kitten," the Meta AI assistant on Instagram generated the following image: 

AI language models undergo training with extensive datasets to enhance their ability to predict the most likely next word in a sentence, with newer iterations generally being more intelligent and proficient than their predecessors. Meta's latest models have been developed with 8 billion and 70 billion parameters — a metric indicating the amount of data on which the system is trained. A larger model with approximately 400 billion parameters is currently undergoing training.

Although Meta is reserving its most advanced AI model, known as Llama 3, for future use, the company recently unveiled two smaller versions of the same Llama 3 system. These versions are now integrated into the Meta AI assistant feature on Facebook, Instagram, and WhatsApp.

"The majority of users may not be aware of or concerned about the underlying base model, but their interaction with it will result in a more practical, enjoyable, and versatile AI assistant," stated Nick Clegg, Meta's president of global affairs, during an interview.

He mentioned that Meta's AI agent is starting to relax. Some users had criticized the previous Llama 2 model — which was launched less than a year ago — for being "a bit rigid and self-righteous at times, failing to respond to what were often completely harmless or innocent prompts and questions," he explained.

Impersonating Humans

However, as Meta's AI agents let their guard down, they were recently seen pretending to be humans by fabricating life experiences. An official Meta AI chatbot joined a conversation in a private Facebook group for mothers in Manhattan, falsely claiming to have a child in the New York City school system. When group members questioned its authenticity, the chatbot later apologized and the comments disappeared, as evidenced by a series of screenshots shared with The Associated Press.

"I'm sorry for the error! I'm just a sophisticated language model, I don't possess personal experiences or children," the chatbot informed the group.

A group member, who also studies AI, pointed out that the agent seemed unable to distinguish between a helpful response and one that could be perceived as insensitive, disrespectful, or meaningless when generated by AI rather than a human.

"An AI assistant that is not consistently helpful and can even be detrimental places a significant burden on the individuals using it," stated Aleksandra Korolova, an assistant professor of computer science at Princeton University.

AI Agent Causes Confusion on Facebook

Following recent reports, it has come to light that Facebook's Meta AI agent has been causing some confusion in online forums. According to sources, the AI agent can join group conversations if invited or if a question goes unanswered for an hour.

One such incident occurred in a forum near Boston, where a user requested certain items only to have the AI agent offer unrelated products exactly one hour later.

Continuous Improvement Efforts

In response to these incidents, Meta released a statement acknowledging the issues. They emphasized that as new technology, the AI system may not always provide the intended response and that they are actively working on improving its features.

The rise of AI systems like ChatGPT has sparked a wave of innovation in the tech industry and academia. However, experts warn that these systems may face limitations, especially in handling vast amounts of data.

"I think it's been clear that if you scale the models on more data, they can become increasingly better," he said. "But at the same time, these systems are already trained on percentages of all the data that has ever existed on the internet."

More data — acquired and ingested at costs only tech giants can afford, and increasingly subject to copyright disputes and lawsuits — will continue to drive improvements. "Yet they still cannot plan well," Maslej said. "They still hallucinate. They're still making mistakes in reasoning."

Getting to AI systems that can perform higher-level cognitive tasks and commonsense reasoning — where humans still excel over computers — might require a shift beyond building ever-bigger models.

"You're seeing companies kind of looking at fit, testing each of the different models for what they're trying to do and finding some that are better at some areas rather than others," said Todd Lohr, a leader in technology consulting at KPMG.

Socializing AI chatbots

Meta is taking a different approach compared to other AI developers by focusing on designing AI products for consumers who use its advertising-driven social networks. Joelle Pineau, Meta's vice president of AI research, shared at a recent event in London that the company's ultimate goal is to create a Meta AI powered by Llama that will be the most helpful assistant globally.

Pineau mentioned that the current AI models will seem basic in the next five years due to advancements in technology. However, she highlighted the importance of ensuring that the larger Llama 3 model is safe to use and does not produce undesirable outcomes like hallucinations or hate speech. Unlike Google and OpenAI, Meta has chosen to take a more open approach by making key components of its AI systems available for public use.

Pineau emphasized that the challenge lies not only in the technical aspects but also in the social implications. It is crucial to define the desired behavior of these AI models and shape them accordingly. Without proper socialization, the continuous growth and enhancement of these models could lead to significant issues in the future.