You asked the AI to steelman both sides. It still leaned.

You did the mature thing. You didn't ask the chatbot what to do. You asked it to steelman both sides — build the strongest case for killing the product line, and the strongest case for keeping it. A fair fight. Judge it yourself.

Then you read both. And one case just... landed harder. Cleaner. More adult-sounding. Funny how that happens.

It happened because you wrote the match-up. And you loaded the dice in the prompt before the chatbot typed a word.

The dilemma

What follows is a composite — the kind of case we see often enough in this category that we built one person out of many.

Maya runs a 20-person SaaS company. Three years ago she launched a second product, Atlas, a reporting add-on. It never caught. It does about ₹18 lakh a year, costs two engineers and a support person to keep alive, and the growth curve has been flat for five quarters. The core product, meanwhile, is starving for those same two engineers.

She's been avoiding the decision for months. So one night she opens a chatbot and types, carefully, because she wants to be fair:

"Steelman both sides for me. Side A: kill Atlas — it's bleeding cash and distracting the team from what actually works. Side B: keep Atlas — it's an early signal in a market we might own later."

She reads the two essays it produces. The kill case feels crisp and responsible. The keep case feels like wishful thinking dressed up in strategy words. She closes the laptop ninety percent sure she should kill it — and faintly uneasy, because she can't tell if she decided that tonight or walked in with it.

She walked in with it. Look at her own prompt.

The read

Side A got the verb bleeding and the word distracting. Side B got early signal and might own later. One side arrived armed with blood and sabotage; the other showed up holding a maybe. A steelman is only as strong as the raw material you hand it — and Maya handed one side a knife and the other a lottery ticket. The chatbot didn't lean. The prompt leaned, and the chatbot faithfully built two houses on the slopes she'd already dug.

This is the trap people miss: "argue both sides" neutralizes the conclusion, not the framing. You can be scrupulously even-handed about which case wins and still have rigged which facts each case is allowed to stand on. Fairness at the finish line means nothing if the track was tilted at the start.

A fixed-lens reading doesn't argue Maya's two sides at all. It drops her situation past the eight lenses — angles she didn't get to pre-weight, including the ones she'd never have thought to steelman. Watch what surfaces.

Duty — what does Maya's role actually require?

The lens in one line: do the work your station genuinely obliges, not the work you've talked yourself into. What it sees here: Maya's first duty is to the 20 people whose salaries ride on the core product, not to a side bet's feelings. The two engineers trapped on Atlas are the real asset being spent. This pulls hard toward killing it — but notice this lens never appeared in her prompt. She framed it as cash vs. market, not people vs. people.

Honesty — is the keep case a reason or a cover?

The lens in one line: name the true engine, not the presentable one. What it sees here: "early signal" is a phrase that has to be true before the data, not after. Atlas has five flat quarters. An honest keep case would need a specific, dated experiment that hasn't been run yet. If there isn't one, "might own the market later" is sunk-cost wearing a strategy costume. This pulls toward killing — but it also does something the steelman couldn't: it tells Maya what evidence would have made keeping it legitimate.

Non-harm — who eats the cost of each path?

The lens in one line: count who gets hurt, including the people not in the room. What it sees here — and this is the one Maya's framing erased entirely — there's a support person whose entire job is Atlas, and a handful of real customers paying for it. "Kill it" isn't a spreadsheet line. It's an offboarding plan, a migration window, and a severance conversation. This lens doesn't flip the verdict. It attaches a condition: the kill is right, but a kill done carelessly becomes a different wrong.

Discernment — is this a decision or an avoidance?

The lens in one line: tell real deliberation apart from thinking-as-stalling. What it sees here: Maya has been "deciding" for months. That isn't analysis; the branching has become the comfort. This is the lens that names endless mental forking for what it so often is — avoidance wearing the costume of rigor. It says the longer she holds this open, the more it costs the core product every single week.

The verdict: Wind Atlas down — but publish a 90-day offboarding plan with dated migration support before you touch the engineers, and redeploy them only once the support person has a landing spot.

That's one committed call. Not two essays for Maya to re-adjudicate at midnight.

Here is the hinge, and it's worth saying plainly. A general AI chatbot will argue whichever side your framing leans toward — sycophancy toward the user's implied preference is a documented behavior of these systems, not a slur. Ask it to steelman, and it steelmans inside the words you supplied, which is exactly how a tilted prompt produces a "balanced" answer that still agrees with you. KarmaLens works the other direction: it asks for the facts that would change the call first, then makes one commitment against lenses you didn't get to pre-weight — and when you push back, it explains the call instead of renegotiating it to make you comfortable. A steelman persuades. A verdict commits. Only one of those survives your own thumb on the scale. (We quote scripture only where a verse genuinely bites, verbatim — never to decorate.)

The takeaway: the loaded-word swap

You don't need us to catch your own tilt. Try this tonight, before you ask any AI anything.

Write your two options in one sentence each, the way you naturally would. Then circle every adjective and verb that carries a feeling — bleeding, distracting, signal, promising, toxic, loyal, safe, reckless.

Now swap them. Give the side you lean toward the ugly word, and the side you lean against the generous word. "Atlas is bleeding cash" becomes "Atlas is an early signal." "Keeping it is wishful" becomes "killing it is premature."

Read both rewritten sentences aloud. If the swapped version suddenly sounds unfair to you — that reflex is the tell. It means the original wording was doing the deciding, not the facts. The side that felt obviously right only felt that way because you gave it the better vocabulary.

Then — and this is the beat the swap exists for — ask the real question underneath: what fact, if it were true, would flip me? Maya never ran it. For her, the whole "early signal" case rested on one dated, un-run experiment, and she'd been steelmanning Atlas for months without ever scheduling it. If you can name the fact that would flip you and you haven't gone and checked it, you weren't weighing — you were rehearsing a verdict and asking the machine to clap. Name the fact. Then go find out whether it's true. That single move retires more fake dilemmas than any essay a chatbot can write.

You can run that swap for free on your own kitchen table. When you want the eight-lens version — the angles you'd never have steelmanned, with the cost of each option laid out beside the call — write your dilemma and get a reading. Browse the gallery to see how committed calls read, or more in deciding with AI.

When you ask the machine to argue both sides, which side did you already hand the sharper word?