Scaled Cognition raises $100M to fix AI hallucinations

Key Takeaways

• Scaled Cognition raised $100 million at a $750 million valuation to tackle AI hallucinations
• The company's APT model offers alternative architecture designed for provably reliable outputs
• Healthcare, legal, and financial sectors face growing liability from AI-generated errors

Scaled Cognition, an AI lab focused on eliminating hallucinations from large language models, has closed a $100 million Series A round that values the company at $750 million. The funding will accelerate hiring for its research team and speed up product development, according to a Wall Street Journal report published Thursday.

The round reflects growing investor conviction that whoever solves AI reliability first will capture a massive slice of enterprise spending. Companies want to deploy generative AI, but the unpredictability of current models keeps legal, compliance, and risk teams awake at night.

Why frontier models keep making things up

Dan Roth, Scaled Cognition's co-founder and CEO, described frontier AI models to the WSJ as "amazing, but also sort of like schizophrenic geniuses." The problem, he explained, is consistency and verifiability.

“They can create incredible answers, and then you can ask them the same question a second time and get a completely different answer that might not even be correct. We really believe that for these systems to really be useful, you have to be able to trust them. And in order for you to trust them, they have to be provably reliable.”

— Dan Roth, Co-Founder & CEO, Scaled Cognition

Roth gave a stark example: a healthcare AI that hallucinates a single digit in a prescription could give a patient the wrong medication. One error, catastrophic consequences. This is not a hypothetical. The industry has already seen lawyers sanctioned for citing AI-hallucinated case law, and major firms admitting their attorneys relied on fabricated citations.

What Scaled Cognition is building

Roth and co-founder Dan Klein, the company's CTO, developed an alternative architecture called APT, or Agentic Pretrained Transformer. The goal is reliably accurate outputs, not just statistically probable ones.

Beyond the model itself, Scaled Cognition has built an enterprise platform for AI deployment. It includes agentic tooling, live agent monitoring, and simulation and evaluation frameworks. The pitch to enterprise buyers: you can actually audit what the AI is doing, not just hope it gets things right.

The hallucination problem has become a liability problem

A year ago, hallucinations were dismissed as teething errors. That framing has shifted. A federal judge in Wyoming threatened to sanction attorneys who submitted AI-generated briefs filled with fabricated cases. Butler Snow, a major law firm, publicly acknowledged that its lawyers had relied on hallucinated citations. What looked like quirky tech failures in consumer chat apps became reputational and regulatory landmines when applied to banking, payments, or compliance.

Leading AI developers have responded by training models to say "I don't know" rather than improvise. But the fundamental issue remains: probabilistic models will never be completely error-free. The question is whether you can reduce error rates to acceptable thresholds and detect failures before they cause harm.

Lloyds Bank and Coinbase have both increased confidence in their hallucination protections after deploying what they describe as safer generative AI systems. Insurance companies have started offering policies that specifically cover AI-related errors, including hallucinated outputs. When insurers start pricing a risk, you know the industry considers it real.

A $750 million bet on trust

The valuation is aggressive for a Series A. It suggests investors believe the market for provably reliable AI is large enough to justify the premium. Enterprise AI spending is projected to exceed $1 trillion by 2030, but much of that spending is currently blocked by trust deficits. If Scaled Cognition can credibly claim to solve, or at least significantly mitigate, hallucination risk, they have a clear path to capturing substantial enterprise budgets.

The company faces competition from the labs that created the hallucination problem in the first place. OpenAI, Anthropic, and Google DeepMind have all named hallucination reduction as a top research priority. But Scaled Cognition's argument is architectural: bolting reliability onto existing probabilistic models may never work as well as building for reliability from the ground up.

Whether APT delivers on that promise remains to be seen. The $100 million buys time to prove the thesis.

Frequently Asked Questions

What is AI hallucination?

AI hallucination occurs when a language model generates false, fabricated, or inconsistent information while presenting it confidently as fact. This can include invented statistics, fake citations, or contradictory answers to the same question.

What is Scaled Cognition's APT model?

APT stands for Agentic Pretrained Transformer. It is Scaled Cognition's flagship AI model, designed with an alternative architecture that aims to produce provably reliable outputs rather than probabilistically likely ones.

Why are enterprises concerned about AI hallucinations?

Hallucinations create legal, regulatory, and reputational risks. In sectors like healthcare, finance, and law, a single AI error can result in patient harm, compliance violations, or sanctions from courts and regulators.

How much did Scaled Cognition raise?

Scaled Cognition raised $100 million in a Series A round, valuing the company at $750 million.

Source: PYMNTS | / PYMNTS