You gave the AI both sides. It sided with the loud one.

The paragraph that was quietly begging

You typed two paragraphs for the option you resent. You typed one line for the option you secretly want. The AI read the pile, weighed it fairly, and handed back the answer your longest paragraph had been rehearsing all along. It felt balanced. It felt earned. Then, an hour later, you noticed you weren't relieved — you were vindicated. Those are not the same feeling.

The dilemma

Here is a composite — stitched from the kind of prompt we see over and over in this category, not any real person.

Meet the prompt Kabir pasted at 11pm. He works two jobs. One is a stable analyst seat at a firm that pays on time and asks little of him. The other is a side contract building dashboards for a small nonprofit — unpaid, six months in, run by a founder who keeps promising a real budget "next quarter."

The nonprofit finally offers Kabir a full role. Less money, real ownership, the work he actually likes. His prompt to the chatbot reads like this: three careful paragraphs on the nonprofit role — the runway math, the founder's flakiness, the two direct reports who would depend on him, the reversibility if it fails, the optics of leaving a good job for a shaky one. Then, at the bottom, one sentence: "Or I stay put, which is safe and honestly a little dead."

The model weighed the pile. It flagged the founder's broken budget promises. It counted the dependents. It named the reversibility risk. It concluded, warmly and at length, that staying put was the prudent call. Kabir felt the verdict land. What he didn't notice: he had spent three paragraphs building a case against the thing he wanted, and one dismissive line on the thing he was going to talk himself out of. The machine didn't read his life. It read his word count.

The read

Run the same facts through the eight fixed lenses — the ones that don't care how many sentences a thing got. Four of them bite here. (Every lens named links to the eight lenses.)

Duty — is this your work to do?

The lens asks whether the role is genuinely yours or a borrowed shape. Kabir already does the nonprofit work. Two people are being scaffolded by it right now. The stable seat asks nothing of him and would notice nothing if he left. Measured by whose work goes undone when he walks, the dead seat scores near zero and the ownership role scores +3. The word count said otherwise. Duty doesn't.

Honesty — could you say your reason aloud?

The lens asks whether the story survives daylight. Kabir's three-paragraph case reads as diligence. Read again: nearly every line is a reason to not do the thing, and the one true desire got a single self-mocking sentence. The framing itself is the tell. He wrote a brief for the prosecution and then asked the jury to be balanced. Honesty here scores the gap between what he wants and how he described wanting it: -2.

Non-harm — who actually pays?

The lens asks who eats the irreversible downside. This is where the chatbot's read had a real point, and the lens keeps it. The two direct reports depend on a founder who breaks budget promises. If Kabir jumps and the nonprofit folds, three people fall, not one. That's genuine. Non-harm scores -1 on the leap — not because of Kabir's fear, but because of the people standing under it.

Discernment — is the founder's flakiness a fact or a shield?

The lens asks whether your evidence is load-bearing or decorative. Here's the tension the chatbot smoothed over. Non-harm says the founder's broken promises are a real risk to real people. Discernment asks a sharper question: did Kabir raise the flakiness to protect his reports, or to have a respectable reason to stay in the seat he calls dead? Those two lenses point opposite ways on the same fact. The chatbot picked the reading that matched Kabir's mood. The lens holds both open and scores discernment +1 only if he can name a concrete condition — a signed budget, a written role — under which he'd go. No condition, and the flakiness is a shield, not a finding.

The verdict: the leap is defensible only if Kabir converts the founder's next promise into writing before he resigns — and staying "because it's safe" is not prudence, it's the longest paragraph winning by volume.

Notice what just happened. A general AI chatbot tends to argue whichever side your framing leans into — leaning toward the user's evident wish is a recognized failure mode of assistant models, not a bug you prompted your way into. KarmaLens runs the opposite way on purpose. The eight lenses are fixed before you type, the score each one returns is aggregated deterministically into a single committed call, and that call does not renegotiate itself when you push back or re-paste a softer version — it will quote 2.41 at you and hold. We're not claiming the words come from somewhere other than a language model, or that the read is free of bias. We're claiming the verdict is scored by what each lens measures instead of by how much room you gave each option — and that it won't be talked down.

The takeaway: the word-count audit

You can run this tonight, on the exact prompt you already wrote, without paying us.

1. Open your "balanced" prompt. Count the sentences you spent on Option A. Count the sentences you spent on Option B.
2. Circle the option that got fewer words. That is almost always the one you want and are pre-arguing against.
3. Now flip the ratio. Rewrite the short option in three honest sentences and compress the long one to a single line. Read the new version cold.
4. Ask the one question the sentence count can't fake: who pays if I'm wrong, and did I give that person their own paragraph?

The verdict that changes when you change the ratio was never a verdict. It was a transcript of how you talk to yourself when you've already half-decided. A real call survives the reweighting. A rationalization needs the original word count to hold its shape.

Ready to see your dilemma scored by the lenses instead of your paragraph lengths? Run it through KarmaLens, or read a few finished reads in the gallery first. Then ask yourself: was the answer you got last night a decision — or an echo of your longest paragraph?