A quiet crisis has been brewing under the AI boom. For all the dazzling demos of GPT-5, Gemini 2.0, and Claude 4, enterprise adoption of large language models has hit a stubborn wall: they keep making things up.
Now a startup called Scaled Cognition is betting $100 million that it can finally fix that.
The Hallucination Problem in Numbers
The numbers are sobering. OpenAI's own system card for o3, released in April 2025, showed a hallucination rate of 33% on PersonQA — a benchmark for answering questions about people. That's one in three answers being wrong. Its predecessor o1 scored 16%. Newer isn't always better.
According to The Business Research Company, the global market for AI hallucination detection and mitigation software is on track to hit $2.47 billion in 2026. A February 2026 study of 847 production deployments found that RAG (Retrieval-Augmented Generation) pipelines reduce hallucination rates by a median of 71% compared to standalone LLMs — but that still leaves plenty of room for error in high-stakes enterprise settings.
Seekr, another player in the space, recently published research showing that in agentic workflows — where AI systems take actions rather than just answering questions — hallucination rates are actually rising, not falling. The very complexity that makes agents powerful also makes them harder to ground in truth.
Dan Roth's Thesis
Scaled Cognition was founded by Dan Roth, a veteran AI researcher and former VP at a major tech company. The company's thesis is simple: model improvement alone won't solve hallucinations. You need a dedicated reliability layer that sits between the model and the enterprise application.
Backed by Khosla Ventures with a $100 million Series A — a staggering round by any measure — Scaled Cognition is building what it calls a "factuality engine." The system doesn't try to make models smarter. Instead, it cross-references model outputs against verified sources in real-time, scoring each claim for accuracy before it reaches the user.
This approach mirrors what the market is already demanding. Crunchbase data shows that AI startups captured 81% of all global venture funding in Q1 2026, with a massive pivot toward reliability infrastructure and guardrails. Investors are no longer throwing money at the shiniest demo — they want systems that won't hallucinate a customer invoice.
Why This Matters for Enterprise AI
The implications are significant. Every hallucinated output costs the same as a correct one in tokens. But the real cost is trust. A single confidently wrong answer from an AI assistant handling customer support, legal research, or medical triage can erode months of deployment progress.
Enterprise procurement teams are starting to ask a new question: not "how smart is your model?" but "can you prove your answers are right?" This shift is creating an entire new category of AI infrastructure — one that Scaled Cognition, Seekr, and a handful of others are racing to own.
The EU AI Act's phased enforcement adds regulatory urgency. Starting in 2026, companies deploying AI in regulated sectors must provide provenance and contestability for model outputs. A reliability layer isn't optional anymore — it's becoming a compliance requirement.
The Verdict
Scaled Cognition's $100 million raise signals that the AI industry is maturing. The first wave was about raw capability — can we build a model that sounds human? The second wave is about reliability — can we build a model that is correct?
If Dan Roth and his team deliver on their factuality engine, they won't just have built a successful startup. They'll have unlocked the enterprise AI market that everyone has been waiting for.
The hallucination era may finally be ending. It's about time.
Photo by BUDDHI Kumar SHRESTHA on Unsplash
