There is a behavior so fundamental to human intelligence that we barely notice it. We do it dozens of times a day, in conversations large and small, without conscious thought. It is so natural that we consider its absence a character flaw. We even have a name for people who can’t do it: we call them overconfident. We call them bluffers. Sometimes we call them dangerous.
The behavior is simply this: knowing when you don’t know something. And saying so.
Artificial intelligence, for all its breathtaking capability, has a deep and largely unresolved problem with those three words. I don’t know. Getting an AI to say them — genuinely, reliably, and in the right context — is one of the hardest challenges in the field. It is also one of the most important. And it is something we solved with MERCED™.
Here is why it matters. And here is what it took.
Why AI is built to bluff
This is not a design flaw in the conventional sense. It is a consequence of how AI models are trained.
Most AI systems are evaluated on whether they get answers right. If they answer confidently and correctly, they score well. If they answer confidently and incorrectly, they score poorly. And if they say “I don’t know” — they score exactly the same as being wrong. Zero points either way.
The math is simple: AI learns to guess. Always. Confidently. Why? Because guessing offers a chance at a point. Honesty offers none.
MIT research published in 2025 found that AI models are 34% more likely to use confident language — words like “definitely,” “certainly,” and “without doubt” — when generating incorrect information than when generating correct information. The AI is most confident precisely when it is most wrong.
The awkward commercial reality is that if a major AI platform admitted “I don’t know” too often, users would simply go elsewhere. So the incentive to fix the problem is weaker than it should be. AI companies are caught between building systems that are honest and building systems that feel impressive. And for most consumer applications, impressive wins.
According to a 2025 global impact study, the financial cost of this problem reached $67.4 billion globally in 2024 — the price the world is paying for AI that will not admit what it does not know.
For businesses using general-purpose AI in professional settings — answering questions about shipments, compliance obligations, pricing, liability — this is not an abstract problem. Your AI is responding to your customers with serene confidence about things it may have simply invented. That is not a tool. That is a liability.
What it actually takes to build an AI that says “I don’t know”
General models learn patterns; they do not learn boundaries. To teach an AI where its knowledge ends, you must first define where the truth begins. That definition requires two decades of lived operational reality, not just a dataset.
When we set out to build MERCED™, we knew early on that this would be one of the defining challenges. Not the flashiest feature. Not the one that demos well in a conference room. But the one that determines whether an AI can actually be trusted in production — with real businesses, real stakes, and real customers.
The easy path was to ship an AI that felt confident and comprehensive and hope the gaps wouldn’t show. Our R&D team chose the harder path. I am extraordinarily proud of them for it.
Teaching an AI to say “I don’t know” with genuine reliability requires more than an instruction or a setting. It requires building a deep understanding of the boundaries of what the AI actually knows — and calibrating those boundaries against real, industry-specific knowledge rather than generic training data. In the moving and logistics world, that calibration doesn’t come from a textbook. It comes from over 20 years of operational experience with how this industry works, how it breaks, and what the right answers actually look like.
You cannot build a trustworthy uncertainty signal without a trustworthy ground truth to anchor it. That ground truth — for the specific questions our customers ask every day — exists nowhere in a general AI training corpus. We had to build it. And the only reason we could build it is because of the two decades of domain knowledge our team brought to the work.
MERCED™ can debate. And it knows when to step back.
One of the capabilities we are most proud of is MERCED™’s ability to engage in genuine debate with other AI systems. Not to simply agree with them. Not to blindly contradict them. But to reason, hold its ground where it has knowledge, push back where something is wrong — and, critically, say “I don’t know” when a question sits outside the boundaries of what it genuinely knows.
This matters more than it might sound. As AI agents become more common in business environments, they will increasingly interact with each other — comparing outputs, cross-checking recommendations, challenging each other’s conclusions. An AI that folds under pressure or doubles down on error because confidence is its only metric is a weak link. In a chain of automated decisions, that link breaks the whole.
MERCED™ was built to do neither. It will defend a position when the underlying knowledge supports it. It will update when presented with better information. And it will stand down when the honest answer is that it doesn’t know — even when another AI, or a user, is pressing it for certainty it cannot provide.
In an industry where a wrong answer can mean a financial dispute, a compliance violation, a failed military shipment, or a family’s belongings handled incorrectly — trustworthiness is not a nice-to-have. It is the whole point.
The standard we are holding ourselves to
The AI industry is beginning to recognize what we built MERCED™ around from the start. A new industry benchmark released in late 2025 takes a fundamentally different approach to evaluating AI — penalizing incorrect answers more harshly than admissions of uncertainty, and rewarding models that know when to say “I don’t know” over models that guess wrong with confidence.
The field is slowly catching up to what good AI actually looks like. We didn’t wait for the field.
Honest uncertainty is not a failure mode. It is a feature. And the businesses that rely on MERCED™ are not looking for an AI that sounds good in a demo. They are looking for an AI they can believe when it matters.
That is the standard we built to. And we are not done raising it.
