In the wake of the 2024 US presidential election, one fact has become clear. That means misinformation is spreading online at an alarming rate, shaping Americans' views not only of each candidate but also of topics as diverse as public health, climate change, and immigration. Generative AI only exacerbates the problem, with its ability to create deepfakes in seconds and its propensity to hallucinate facts.
Factiverse, a startup that participated in the TechCrunch Disrupt Battlefield 200 in October, is bracing for an onslaught. The company, which won top honors in the Security, Privacy, and Social Networking categories, has developed a B2B tool that provides live fact checking for text, video, and audio. The company's pitch is to help companies save time on investigations and reduce reputational risk and legal liability.
Norwegian startups are still in their early stages. Since its launch in 2020, Factiverse has raised approximately $1.45 million in pre-seed funding. Still, Factiverse CEO and co-founder Maria Amelie says they have already started working with both media and financial partners, including one of Norway's largest banks.
Factiverse also provided live fact-checking of the U.S. presidential debates that was used by several media partners, Ameri said.
“We're not an LLM (Large-Scale Language Model). We've built a different type of model based on information retrieval,” Ameri told TechCrunch.
A former technology journalist and published author, Amelie has first-hand experience in the fight against facts. She collaborated with Vinay Setty, co-founder and CTO of Factiverse, an associate professor of machine learning at the University of Stavanger, to launch a B2B-focused startup.
Amelie said Factiverse's models are based on high-quality, curated, and reliable data from trusted sources and fact-checkers around the world, rather than the “junk food data” on which the generative AI is trained. He is said to have been trained.
“We train our AI models to think intuitively like people with extensive information research experience,” Ameri says.
Based on machine learning and natural language processing, the model searches the web in real-time on everything from search engines like Google and Bing to AI search engines like You.com to academic papers and identifies claims. You can.
“The most interesting thing is that we don't show you the first thing that comes up in a search engine,” Ameri says. “We actually suggest which sources are the most or historically most reliable sources for your topic. We actually look at areas that correlate with your topic and if In some cases, we also look at who is being quoted in the article.
As of today, Factiverse says it outperforms GPT-4, Mistral 7-b, and GPT-3 in its ability to identify fact-check-worthy claims in 114 languages. The company's model is also better than LLM at determining the veracity of claims. Amelie said Factiverse's success rate is about 80% and the company's goal is to improve as it acquires new customers around the world.
“We have enough money to be the best, but we're here in the U.S. to be the fastest,” Ameri told TechCrunch. He also noted that the company hopes to raise a seed round in 2025. “We're looking for customers and investors who want to invest in trust and credibility,” she said.