Startup Probably raises $9 million to make AI as accurate as regular code

American startup Probably has closed a seed round of $9 million. The funds will go toward developing AI systems that prevent hallucinations and factual errors before the answer reaches the user. The main goal is to achieve accuracy comparable to deterministic systems, that is, to ordinary software code.

Why hallucinations are not just an inconvenience

Large language models regularly "hallucinate": confidently produce false facts, cite non-existent research, invent quotations. According to various independent assessments, even the most advanced commercial models make errors in 10–20% of specific factual statements. At the same time, it is extremely difficult for users to determine where the model is right and where it is making things up — it sounds equally confident in both cases.

This creates a systemic risk that limits the applicability of AI in real business scenarios. Companies are forced to choose between two bad options: hire people to verify each AI response or limit the use of models to tasks where errors are not critical. For medicine, law, financial compliance, and the public sector, this effectively means that autonomous AI systems are not applicable — the cost of error is too high.

New benchmark: deterministic accuracy

Probably is setting a non-trivial benchmark — to achieve the accuracy of deterministic systems. This is fundamentally different from how today's LLMs work. A deterministic program always produces the same result with the same input data: a calculator will never "decide" to make up an answer to 2 + 2. A language model works differently: it generates probabilistic text that may sound convincing but does not have to be true. The Probably team is building an architecture where AI does not just generate an answer but verifies it before sending it. Key principles:

• Clear separation between reliable knowledge and the model's assumptions
• Verification of each factual statement before it reaches the user
• Explicit "I don't know" instead of invented facts
• Transparency: the system explains the source of each statement
• Zero tolerance for factual errors in the final answer

This approach is opposite to the approach used in most LLM developments in recent years. Models were fine-tuned with an emphasis on "helpfulness" and "persuasiveness" — which often conflicted with accuracy.

Market and competition

The task of reliable AI is now actively attracting venture capital. In this niche, companies like Vectara, Cohere with a focus on enterprise RAG architectures, a number of stealth startups, and academic projects at major universities operate. Major players — OpenAI, Anthropic, and Google — are also investing in reducing hallucinations, but for them accuracy is only one of many product characteristics, not the main differentiator.

Probably builds everything around one idea. The $9 million seed round gives the team one and a half to two years to prove the viability of the architecture and attract the next funding. The company name is a deliberate irony about the probabilistic nature of language models and at the same time a declaration of intent to overcome it.

What this means

If Probably manages to convincingly solve the hallucination problem, it will open the way for AI into regulated industries — healthcare, legal services, financial compliance — where today autonomous systems are unacceptable due to inaccuracy. For business, this means the ability to eliminate the expensive layer of human verification and truly delegate critical tasks to machines — not in words, but with measurable accuracy guarantees.