Campbell Brown has spent her career chasing accurate information, first as a renowned TV journalist, then as Facebook’s first, and only, dedicated news chief. Now, watching AI reshape how people consume information, she sees history threatening to repeat itself. This time, she’s not waiting for someone else to fix it.

Her company, Forum AI — which she discussed recently with TechCrunch’s Tim Fernholz at a StrictlyVC evening in San Francisco — evaluates how foundation models perform on what she calls “high-stakes topics” — geopolitics, mental health, finance, hiring — subjects where “there are no clear yes-or-no answers, where it’s murky and nuanced and complex.”

The idea is to find the world’s foremost experts, have them architect benchmarks, then train AI judges to evaluate models at scale. For Forum AI’s geopolitics work, Brown has recruited Niall Ferguson, Fareed Zakaria, former Secretary of State Tony Blinken, former House Speaker Kevin McCarthy, and Anne Neuberger, who led cybersecurity in the Obama administration. The goal is to get AI judges to roughly 90% consensus with those human experts, a threshold she says Forum AI has been able to reach.

Brown traces the origin of Forum AI, founded 17 months ago in New York, to specific moment. “I was at Meta when ChatGPT was first released publicly,” she recalled, “and I remember really shortly after realizing this is going to be the funnel through which all information flows. And it’s not very good.” The implications for her own children made the moment feel almost existential. “My kids are going to be really dumb if we don’t figure out how to fix this,” she recalled thinking.

What frustrated her most was that accuracy didn’t seem to be anyone’s priority. Foundation model companies, she said, are “extremely focused on coding and math,” whereas news and information are harder. But harder, she argued, doesn’t mean optional.

Indeed, when Forum AI began evaluating the leading models, the findings weren’t exactly encouraging. She cited Gemini pulling from Chinese Communist Party websites “for stories that have nothing to do with China,” and noted a left-leaning political bias across nearly all models. Subtler failures abound too, she said, including missing context, missing perspectives, straw-manning arguments without acknowledgment. “There’s a long way to go,” she said. “But I also think that there are some very easy fixes that would vastly improve the outcomes.”

Brown spent years at Facebook watching what happens when a platform optimizes for the wrong thing. “We failed at a lot of the things we tried,” she told Fernholz. The fact-checking program she built no longer exists. The lesson, even if social media has turned a blind eye to it, is that optimizing for engagement has been lousy for society and left many less informed.

Her hope is that AI can break that cycle. “Right now it could go either way,” she said; companies could give users what they want, or they could “give people what’s real and what’s honest and what’s truthful.” She acknowledged the idealistic version of that — AI optimizing for truth — might sound naive. But she thinks enterprise may be the unlikely ally here. Businesses using AI for credit decisions, lending, insurance, and hiring care about liability, and “they’re going to want you to optimize for getting it right.”

That enterprise demand is also what Forum AI is betting its business on, though turning compliance interest into consistent revenue remains a challenge, particularly given that much of the current market is still satisfied with checkbox audits and standardized benchmarks that Brown considers inadequate.

The compliance landscape, she said, is “a joke.” When New York City passed the first hiring bias law requiring AI audits, the state comptroller found more than half had violations that went undetected. Real evaluation, she said, requires domain expertise to work through not just known scenarios but edge cases that “can get you into trouble that people don’t think about.” And that work takes time. “Smart generalists aren’t going to cut it.”

Brown — whose company last fall raised $3 million led by Lerer Hippeau — is uniquely positioned to describe the disconnect between the AI industry’s self-image and the reality for most users. “You hear from the leaders of the big tech companies, ‘This technology is going to change the world,’ ‘it’s going to put you out of work,’ ‘it’s going to cure cancer,'” she said. “But then to a normal person who’s just using a chatbot to ask basic questions, they’re still getting a lot of slop and wrong answers.”

Trust in AI sits at extraordinarily low levels, and she thinks that skepticism is, in many cases, justified. “The conversation is sort of happening in Silicon Valley around one thing, and a totally different conversation is happening among consumers.”

Malware research group vx-underground, which says it has the largest collection of malware source code, said in a post on X that its archive of data amounts to about 30 terabytes.

A reply by Bernardo Quintero, founder of VirusTotal, an online service that scans files for malware across multiple antivirus engines at once, said his service has about 31 petabytes of malware samples that users have contributed to date. (A petabyte is ~1,000x larger than a terabyte.)

In both cases, that’s a lot of data. For context, cybersecurity companies, AI researchers, and threat intelligence firms treat repositories like these as critical for training detection models and understanding how attacks evolve. But this had us wondering: What would these enormous datasets actually look like stacked as hard drives one on top of the other and side by side? And how would they compare to, say, the Eiffel Tower?

Someone in our newsroom asked an AI chatbot this question, and it got it incredibly wrong.

Instead, we did some rough back-of-a-napkin math to figure out how tall these data banks would be. Since vx-underground and VirusTotal both have “about” that much data each, “about” is good enough for us in this case. 

Let’s say we’re using 1 terabyte capacity internal hard drives, since these are generally designed to be the same physical size to fit inside any computer. These standardized 3.5-inch internal hard drives are 1 inch in height, which for the sake of stacking one on top of the other is really what we want to know here.

We’re also assuming that the hard drives we’re using in this example are exactly 1 terabyte, because in reality the total usable file capacity of a hard drive is generally somewhat less. 

Using this online conversion tool, it looks like vx-underground’s 30 terabytes of malware data could fill 30 hard drives stacked on top of one another, reaching 30 inches, or about 2.5 feet tall.

For reference, this reporter is 6 feet tall. (See visual below, and yes, terrible opsec, I know.)

With that same logic, VirusTotal’s 31 petabytes of submitted data would fill 31,744 hard drives, which stacked on top of one another would reach about 2,645 feet.

The world’s tallest building, the Burj Khalifa in Dubai, is slightly taller at 2,722 feet.

The Eiffel Tower is 1,083 feet tall. By that logic, VirusTotal has about two and a half Eiffel Towers’ worth of data.

While AI is now being applied to everything from healthcare to customer support, no single use case has yet been nearly as popular or lucrative as code writing.

Jack Newton, co-founder and CEO of Clio, a Canadian law firm management software company, is convinced that legal tech is poised to be the next big winner of the LLMs era. That’s a self-interested claim — 18-year-old Clio is a legal tech company — but the numbers are hard to dismiss.

Clio saw its revenue growth accelerate sharply after integrating AI into its offering in 2023. The company surpassed $200 million in annual recurring revenue (ARR) in mid-2024, doubled that figure by late last year, and just announced that its ARR reached $500 million.

“LLMs are so excellent for coding because all the existing code in the world is a huge repository to train on,” Newton said. “The analogy to legal is really clear.”

Law firms hold massive corpuses of contracts and agreements, providing a rich basis of text-based data for AI models to learn from.

“Tech companies and lawyers alike are recognizing what a huge amount of upside there is for legal with LLMs,” Newton said.

Clio isn’t the only legal tech company seeing a massive revenue surge driven by AI.  

Four-year-old Harvey, which offers LLM AI for law firms, hit ARR of $190 million by the end of 2025, co-founder and CEO Winston Weinberg shared on LinkedIn. Harvey’s main rival, Legora, announced last month that it reached $100 million in ARR a mere 18 months after launching its platform.

Although the legal tech community’s definition of ARR has been under scrutiny recently, the opportunity to apply AI to law makes clear sense, given that LLMs can automate the field’s most time-consuming tasks, such as document review and drafting.

Legal tech companies aren’t the only ones recognizing how valuable AI could be for lawyers. Earlier this week, Anthropic announced a suite of new legal-specific features, expanding Claude for Legal — the law-focused plug-in whose debut earlier this year sent legal tech stocks tumbling.

Both Harvey and Legora rely on Claude as a core model among others, which makes the dynamic an uncomfortable one: a key supplier is now also a competitor.

For Newton, these are all signs of the vast potential of the legal AI market. He has reason to be optimistic. The Canadian-based Clio was valued at $5 billion when it raised a $500 million Series G last November. The company provides law firms with time-tracking, invoicing, and payment tools. It $1 billion acquisition of data intelligence platform vLex last year now allows lawyers to use Clio’s AI for research, as well.