- Meta's latest family of AI models, Llama 4, can wade into more contentious territory than its predecessor.
- Llama 4 is also "dramatically more balanced" in the prompts it refuses, Meta said.
- AI models have long struggled with bias, with Musk labeling ChatGPT as "woke."
Meta's latest family of AI models, Llama 4, is designed to answer more "contentious" topics like politics than its predecessor, the company said on Saturday.
AI companies typically build guardrails so chatbots, like Meta's AI or ChatGPT, don't wade into overly controversial territory. It's a tricky balance, because too much prompt-dodging can annoy users or leave out important context.
Meta said Llama 4 is less likely to dodge hot-button questions. While the previous version, Llama 3.3, refused to answer 7% of politically or socially charged prompts, Llama 4 turns them down less than 2% of the time, per Meta's tests.
The model is also "dramatically more balanced" in the prompts it refuses, Meta said.
The Llama 4 models consist of the Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. The Llama 4 Scout and Llama 4 Maverick were released on Saturday, while the Llama 4 Behemoth is still training, Meta said.
The Llama 4 Scout and Llama 4 Maverick were distilled from Llama 4 Behemoth, which Meta said is its "most powerful yet and among the world's smartest LLMs."
Meta tested Llama 4 with a set of debated topical questions — ones in which people often take opposing sides. In these tests, Meta checked whether the model would answer one side but refuse the other. This happened in just 1% of the test questions, Meta said.
The Llama 4 models — including the Llama 4 Scout and Llama 4 Maverick released on Saturday — are multimodal AI systems, Meta said. Multimodal systems are capable of processing and integrating various types of data, including text, video, images, and audio.
Meta called the Llama 4 Scout and Llama 4 Maverick its "most advanced models yet," noting that both are "open-weight" AI models.
Open-weight models sit between open-source and proprietary AI, sharing pre-trained parameters but keeping key development details under wraps. This allows developers to fine-tune and deploy the model without access to its training data or architecture.
On a "contentious" set of political or social topics, Llama 4 responds with a "strong political lean" at a rate comparable to Grok's, one of its competitors. This rate is half of what it was in Llama 3.3, Meta said.
"While we are making progress, we know we have more work to do and will continue to drive this rate further down," the company added.
Meta did not respond to Business Insider's request for comment.
"Woke" chatbots
The company said on Saturday that all major LLMs have struggled with bias and they have historically leaned left on contentious issues. "Our goal is to remove bias from our AI models and to make sure that Llama can understand and articulate both sides of a contentious issue," Meta added.
Elon Musk has criticized chatbots like OpenAI's ChatGPT for being "woke" and championed his own xAi's Grok as an alternative.
xAI's training methods for Grok appeared to heavily prioritize right-wing beliefs, some employees told Business Insider's Grace Kay in February.
Meanwhile, OpenAI updated its model in February to embrace "intellectual freedom" and respond objectively to contentious topics.
Llama, Meta's open-source large language model that competes with proprietary models from other companies, has been a key initiative for the company.
CEO Mark Zuckerberg aims to make Llama the industry standard worldwide and said Meta's AI chatbot, available across Facebook, Instagram, and WhatsApp, could reach a billion users this year. As of December, 600 million users accessed Meta AI each month.
Zuckerberg has committed as much as $65 billion to AI projects this year.
