Show me how you would redesign a confident-sounding AI answer to be honest instead.

Solum Finance is a budgeting app. Amadou Kessler is a freelance graphic designer who has used it for eight months. Here is the affordability answer Solum gave him the day he asked about a used car, and the redesign that would have caught what it left out.

How to answer this, stage by stage

Nobody is grading whether you can spot a bad sentence. They're grading whether you can rewrite it into something specific enough to actually act on.

Let's learn

Solum is a budgeting app with a feature that answers one question: can I actually afford this purchase, right now?

Before Solum, freelancers like Amadou tracked irregular income by hand in a spreadsheet, checking their own numbers before any big purchase, a slow, careful habit that took real time but caught real problems.

The whole method, in five rows. The hard one is the third.

Now Solum answers instantly, pulling in real transaction history, no spreadsheet required.

Here's the turn: the extra convenience was never the issue. The real problem showed up once Solum's answers started sounding more certain than the underlying number actually was, and users stopped doing their own math, because a confident tone reads as a confident fact, whether or not it is one.

At its worst, someone makes a large purchase believing "Yes, this fits" is a settled fact, then finds out three weeks later, at the worst possible moment, that it never was.

The checking rate didn't drift down slowly forever. It fell off a shelf the month the copy changed.

What I would leave alone: small, low-stakes purchases don't need this treatment. A quick, confident yes on a 12 dollar subscription is fine, since being wrong about it costs almost nothing.

The lesson: a confident tone is a writing choice, not a fact about the world. The moment the writing sounds surer than the model actually is, it's lying by tone, even if every number behind it is technically accurate.

Now here is the same thing as a story

The short version above is what you'd say defending this redesign to Solum's product council. Read this one for how the trust actually built up.

Every other Sunday, Amadou Kessler used to sit down with a spreadsheet and his last two months of invoices, checking what he could actually afford before he let himself buy anything big. He's a freelance graphic designer, and for years before Solum, he could tell within about fifty dollars what a slow month would do to his plans.

Solum's early affordability answers were careful. "Likely affordable, but your last invoice hasn't cleared yet," the kind of hedge Amadou's own spreadsheet would have caught too. He kept checking his own numbers for the first few months anyway, out of habit, and Solum kept agreeing with him.

After a while he checked before every big purchase, then only the ones over a few hundred dollars, then, once Solum's answers started sounding flatly certain, a copywriting update rewrote its hedges into plain yeses, he stopped checking at all, since the tone alone read as more sure than any spreadsheet he could build.

Solum's team assumed Amadou had a dial he could turn down slowly. He only ever had a switch.

A new engineer on Solum's product team, running a routine audit of the affordability copy, asked in a meeting, "wait, why does this always say 'Yes, this fits' no matter what the underlying number actually is? Did the model get more confident, or did we just make the copy sound more confident?" Nobody on the team had a clean answer.

The audit that followed the new hire's question is what turned up Amadou's case, three months after it happened.

Three months into the confident-tone version, that audit found it: Amadou had asked Solum whether he could afford a 28,000 dollar used car. Solum said "Yes, this fits your budget! You'll have 340 dollars left over each month after this purchase," with no hedge at all. Amadou signed the loan that afternoon.

Three weeks later, an insurance renewal came in 180 dollars higher than expected, the same month one of his two biggest clients paid nine days late. His account went 95 dollars into overdraft, a gap Solum's own transaction history could have shown months in advance, if the answer had ever mentioned it.

With the redesigned answer, restored hedges and named specifics, Solum tells Amadou: "Likely affordable, with two things to watch: your income varied by more than 30 percent over the last three months, and your car insurance renews in six weeks at an estimated 180 dollar increase. In your average month you'd have about 340 dollars left over, but in your lowest month this year, you'd have been about 95 dollars short. Want to see the math, or talk to a real advisor?" Run the same afternoon forward: Amadou waits two pay cycles before signing, and never sees an overdraft at all.

The old copy asked Amadou to trust a tone. The new one shows him the actual month that would have broken his budget.

I signed off on the confident rewrite because it tested well, people liked how sure it sounded. It took a new hire's plain question, and a stranger's overdraft, to see that sure-sounding and sure were never the same thing.

The five steps, if you want to remember it

The recap, one line per letter: find the person is Amadou and his old spreadsheet habit, locate the habit is the cross-checking he quietly stopped doing, identify the flip is over-trust triggered by a more confident tone, pinpoint the old decision is an onboarding promise nobody revisited, and show the replay is the redesigned answer catching the risk three weeks before it would have mattered.

And if you want to be sure it really works, try it somewhere elseA different flip family, a home-insurance claims assistant instead of a budgeting app. Confidence gets gamed from the other direction this time.

Covent Assurance runs an AI claims assistant that gives renters a fast read on whether a claim is likely to be approved. Noor Falkenberg filed a stolen-bicycle claim through it. This story runs on a different family: the pre-editing flip, not over-trust.

Over months, claimants like Noor learned which phrasing got a fast, confident "likely approved" read from the assistant, and started pre-editing their own claim descriptions to match it, stripping out the messy, true details, whether the lock was cut or simply left unlocked, a small gap in the timeline, that made a case genuinely harder to assess, in exchange for a smoother, more confident-sounding answer. The old decision behind it: Covent's early claims copy explicitly told users that certain word choices "help us process your claim faster," a well-meant onboarding tip that quietly taught claimants to write toward the model's comfort zone instead of writing what actually happened.

The habit forms in the third box. By the fifth box, the model is confidently approving a story that isn't quite the real one.

The same tree that fixed Solum's affordability answer also tells Covent's assistant when a claim needs a real human look instead of an instant, confident yes.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "name the specific risk factors, show the worst realistic case, and never let the copy sound surer than the model's real number," and stop.
Cost: there's no time to rewrite every answer template right now. Say so honestly, and start with the highest-stakes decisions, a car, a claim, a loan, since that's where a confident-sounding mistake costs the most.
The model gets better, for real: if Solum's underlying affordability model genuinely gets more accurate, that's still not a reason to let the copy sound more certain than its actual confidence number, the redesign constrains the words to the number, whatever the number turns out to be.

Where people run it wrong.
They test confident-sounding copy for how much users like it, without ever checking whether it matches the model's real number.
They add one disclaimer at the bottom of the screen and call the honesty problem solved, when users have already learned to skip past it.
They assume a wrong answer will generate a complaint, when most people who get burned quietly blame themselves instead.

How to use it live. When someone hands you a confident-sounding AI answer and asks you to make it honest, ask yourself one question before rewriting a single word: what does the number underneath this sentence actually support? Rewrite until the words say exactly that, no more.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0