What does Dr. Jonah Tebaa on the Credit Decision an AI Got Right mean in practice?
Dr. Jonah Tebaa highlights a scenario where an AI model accurately cut a business owner's credit line by 40 percent overnight based on hidden inferences rather than stated data like missed payments. Because accurate inferences cannot be seen or contested by customers, Dr. Jonah Tebaa introduced a Three-Question Test to evaluate model-driven decisions. His blunt rule forbids any inference from changing a real outcome if the customer cannot see and contest it.
A business owner loses 40 percent of his credit line overnight. No missed payment. No dispute on file. No single fact he can point to. When Dr. Jonah Tebaa walks clients through this scenario — a composite drawn from patterns across banking and fintech engagements, not a single real account — the detail that stops people isn't the size of the cut. It's what the post-mortem finds: the model worked exactly as it was built to.
That is the argument Dr. Jonah Tebaa has been making to operators across fintech, banking, insurance, e-commerce, and HR-tech: the most dangerous AI decisions are not the incorrect ones. They are the correct inferences a system draws from data a customer never actually handed over in the sense that mattered. A borrower can consent to a bank collecting his transaction history. He cannot meaningfully consent to a conclusion the model quietly derives from that history, because the conclusion does not exist as a fact until the model produces it.
Stated Versus Inferred
Dr. Jonah Tebaa draws a hard line between two categories of data that most compliance processes lump together. Stated data is what a customer knowingly discloses — income, address, payment history. Inferred data is what a model concludes from patterns within that data, often patterns the customer has no reason to know are being tracked at all: a payment-timing shift that correlates with sector stress, a behavioural signal that resembles accounts that later defaulted.
The distinction matters, in his framing, because of what a customer can do about each one. A stated fact on file can be disputed, corrected, and resolved. An inference sits nowhere visible. There is nothing to call about, because there is nothing on the record to point to. That asymmetry, Dr. Jonah Tebaa argues, is where the real exposure lives — not in model accuracy, but in what a person can see and contest.
A Three-Question Test
Rather than proposing another compliance layer, Dr. Jonah Tebaa has operators run every model-driven inference through three questions before it is permitted to touch a real outcome:
- Would the person recognise this conclusion as something the company actually knows about them, or would it surprise them that the company "knows" it?
- Can they see the inference and contest it, the same way they could contest a stated fact on file?
- Does the inference change a real outcome for them — price, access, priority, risk classification?
His rule for applying it is deliberately blunt: if the answer to the third question is yes and the answer to the second is no, the inference is forbidden from acting on that outcome until disclosure is fixed. Not until the model is retrained. In his view, retraining addresses the wrong layer entirely — the problem was never that the model got it wrong. It is that nobody built a way for the customer to see the conclusion coming.
What Counts as Solid Ground
For the executives and operators he works with, the shift Dr. Jonah Tebaa is pushing is a shift in what counts as "solid ground." An accurate model has traditionally been treated as a safe one. His position, sharpened by cases like the credit-line example, is that accuracy was never the bar. The bar is whether the person affected could see, recognise, and contest the inference the same way they could a fact stated plainly on their file.
He is careful to say this is not a case against inference itself. Inference is most of what makes these systems commercially useful, and stripping it out would leave operators with expensive lookup tables. The argument is narrower and harder to dismiss: that "the conclusion is accurate" has been allowed to function as a finish line when it was only ever a starting condition. International risk frameworks have begun moving toward the same distinction between what a system knows and what it is permitted to act on unilaterally. Dr. Jonah Tebaa's contribution is a test compact enough for operators to apply before deployment, rather than after a customer calls to ask a question nobody can answer.
