Compare two competing AI writing tools on their handling of user control.

Penlight and Draftwing both polish translated documents with an AI rewrite pass. Baptiste Ferron owns a small translation agency handling business and legal documents, and has run both tools across the same kind of client work.

How to answer this, stage by stage

Six moves. Committing to a real position, with a real cost behind it, matters more here than listing every feature each tool has.

Let's learn

Say we build two AI writing tools, both of which take a translated paragraph and offer a polished rewrite. One of them lets you take back a single sentence. The other one doesn't.

Penlight and Draftwing both do this for Baptiste's agency: take a translated business or legal document and offer AI-polished phrasing across the page.

With Penlight, every AI rewrite arrives as one paragraph-sized change. Translators either accept the whole thing or reject the whole thing and redo it manually.

At its worst: a legal clause about payment terms, reworded by the AI in a way that subtly shifted its meaning, gets bundled inside an otherwise fine paragraph. A busy translator, trusting the good 90 percent of the rewrite, accepts the whole thing.

What I would leave alone: a trimmed filler word or a slightly smoothed transition sentence doesn't need the same scrutiny as a legal term. Reviewing every single line with equal intensity would slow translators down for a risk that mostly isn't there.

The lesson: user control in an AI writing tool isn't about how good the suggestions are. It's about how small a mistake you're allowed to fix without undoing everything around it.

Now here is the same thing as a story

The short version above is what you'd say defending this comparison live. Read this one for how the gap actually showed up on a real client file.

Baptiste's agency translates contracts and business correspondence for clients who care, understandably, about exact wording. His translators are good at catching subtle shifts in meaning, the kind of thing a quick skim misses.

Both tools rewrite the same paragraph. Only one of them lets you keep 90 percent of it and fix the rest.

On a rushed Friday, one translator ran a payment-terms clause through Penlight. The AI rewrite improved the sentence's flow but subtly softened a "shall" into a "may," changing an obligation into an option.

Four steps, and the third one is where the whole gap lives: there's no smaller choice available than all or nothing.

The rest of the paragraph read better than the original. Rejecting the whole thing meant losing genuinely good edits. Accepting it meant trusting one changed word buried inside four good sentences. Under deadline pressure, the translator accepted it.

Legal term rewording sits in the dangerous corner: it happens often enough to matter, and getting it wrong is genuinely expensive.

The client caught it two weeks later, during their own legal review, and it cost the agency an uncomfortable call and a rushed correction. Nobody at Penlight had done anything wrong technically. The rewrite itself was fluent and mostly accurate.

None of these three exist in Penlight's design. All three are the entire reason Draftwing's slip-through rate is so much lower.

Baptiste switched a pilot group of translators to Draftwing the next month. On an equivalent clause, misworded the same way, the tool flagged just that one line for review. The translator rejected only the risky sentence, kept the rest of the paragraph's improvements, and the whole review took under six minutes.

Someone at Penlight decided, early on, that a paragraph was the natural unit for an AI rewrite, since sentences read better in context and a whole-paragraph rewrite produced smoother prose than patching line by line. That was a reasonable design call for casual writing, where a slightly off word rarely matters.

I would take that decision back the moment the product started handling legal and business documents, where one changed word genuinely changes what a sentence promises. The unit of an edit needs to match the unit of actual risk, not just the unit that reads most smoothly.

I built the paragraph-sized rewrite because it produced the most fluent-sounding prose in every early demo, and none of those demos involved a contract. It took watching a real client catch a real changed obligation, two weeks after the fact, to see that fluent and safe aren't the same test.

PICK, in one screenFour letters, and the third one, cost asymmetry, is what actually separates these two tools.

The recap, one line per letter: position is line-level control beating paragraph-level control, impact is ten seconds versus a client relationship, cost asymmetry is optimizing against the hidden risky edit, and kill criteria is watching for review fatigue at scale.

And if you want to be sure it really works, try it somewhere elseSame four letters, a factory quality report instead of a legal contract. This time the risky word is a tolerance spec, not a legal term.

A manufacturing plant uses an AI tool to draft quality-control reports from inspection notes, comparing two competing tools much like Penlight and Draftwing. One rewrites the whole report section at once; the other lets an inspector approve one spec line at a time.

Mapped onto PICK: position is the same, line-level control over section-level control, this time applied to a measurement tolerance instead of a legal clause. Impact: rejecting a good phrasing edit costs an inspector a few seconds to redo; accepting a silently altered tolerance number, "within 0.5mm" quietly smoothed to "within 0.6mm," can mean a defective part passes inspection. Cost asymmetry: the hidden and expensive error here isn't a client dispute, it's a part that fails in the field, so the line-level tool is worth its slightly slower review. Kill criteria: if inspectors started rubber-stamping every flagged line without truly reading it, the granular control would stop providing real safety and the design would need a different fix, like requiring a physical remeasurement above a certain size discrepancy.

A factory floor instead of a law office, and the exact same underlying design call: let someone fix the one risky line without redoing everything around it.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "line-level control beats paragraph-level control, because the real risk is a bad edit hiding inside good ones" and stop.
Cost: no engineering time to rebuild Penlight's editor this quarter. Say so, and start with the cheapest version: highlighting which specific words changed within an accepted paragraph, so at least the risk is visible even without granular accept and reject.
The model gets better, for real: even if Penlight's rewrite quality improves substantially, the all-or-nothing unit remains the same risk. A better model just means the bad edit is rarer, not that it's ever fully caught before it's accepted.

Where people run it wrong.
They compare two writing tools on tone or fluency, missing that the real difference is structural: what size of edit you're allowed to control.
They assume more granular review is always slower, without checking whether it actually adds much time in practice.
They blame the translator or inspector for missing a bad edit, when the tool bundled it inside something good and gave them no smaller unit to work with.

How to use it live. When comparing two AI writing tools, ask yourself one question before anything else: if the AI gets one word wrong, what's the smallest thing I'm allowed to fix? That question reveals the real control gap faster than comparing either tool's suggestions.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0