Probably raises $9M to tackle AI hallucinations head-on

Another day, another AI startup chasing the hallucination problem. But Probably is taking a slightly different angle than most.

The company just announced a $9 million raise to build what they call a “more reliable kind of AI.” The pitch is straightforward: stop hallucinations and factual errors from ever reaching the user, rather than just cleaning up after the fact.

What’s the actual approach?

Most AI reliability tools work like spellcheck on steroids — they flag potential errors after the model generates text, hoping a human reviewer catches the rest. Probably wants to bake correctness checks directly into the generation pipeline. Think of it as a guardrail system that runs in real-time, intercepting likely hallucinations before they appear in output.

This is harder than it sounds. Large language models are fundamentally probabilistic — they guess the next token based on patterns, not facts. Getting them to behave like deterministic systems requires either constraining the model heavily (which kills creativity) or building an external verification layer that’s fast enough to keep up with generation.

The $9M round suggests investors believe Probably has something here. But I’ve seen this movie before. Several startups have tried similar approaches — some built custom fine-tuning pipelines, others used retrieval-augmented generation (RAG) as a crutch. The results have been mixed at best.

Where Probably might have an edge is focus. They’re not trying to solve all hallucinations everywhere. They’re targeting specific enterprise use cases where accuracy is non-negotiable: legal document review, medical record summarization, financial compliance. Narrower scope means tighter verification rules, which means fewer false positives and faster execution.

Still, I’m skeptical about the “deterministic” claim. True determinism in AI is an oxymoron — the moment you introduce a neural network, you’re dealing with probabilities. What Probably is really selling is predictable behavior, not deterministic guarantees. That’s a meaningful distinction, and I hope their marketing doesn’t blur it.

If they can get accuracy to 99.9% on structured tasks, that’s genuinely useful. But the last 0.1% is where the real disasters live. A hallucination in a legal contract or a medical diagnosis doesn’t need to happen often to cause serious damage.

The $9M will likely go toward expanding their engineering team and running more rigorous benchmarks against existing solutions. I’d love to see them publish independent evaluation results rather than just internal metrics. The AI reliability space is full of companies claiming 95%+ accuracy on their own test sets, only to fall apart in real-world conditions.

Probably’s approach is worth watching, but I’m not ready to call it a breakthrough. Let’s see how they perform outside the lab first.