Why rewriting your words is the wrong default

You can tell when a colleague's message went through an AI on the way out. The hedge is gone. The one sentence that sounded like them is gone. What arrives is clean, confident, and slightly nobody's.

For a while that was a ChatGPT problem. Now it is a dictation problem, because the category decided rewriting is the feature. Wispr Flow's homepage says it "edits as you speak" and "understands what you mean, not just what you say." Willow's says you can "describe what you want to say in rough form, and Willow turns it into polished writing." Both are quoted from their sites today, and both are honest descriptions of what they build. The output reads well. That is the point.

This post is about why that should be the option you switch on, never the default you have to switch off.

It puts words in your mouth

A rewriting tool chooses the final wording. Most of the time it chooses something close to what you said. Some of the time it chooses something you would never have sent, and there is no way to know which time is which without reading every word.

At two messages a week, that is fine. The people this is built for send eighty. Project managers, account managers, customer success leads, recruiters: people whose whole job is written communication across Gmail and Slack and Jira and Notion, all day, under their own name. At that volume a small per-message chance of the tool saying something you did not say becomes a certainty. The question is only which message, and whether it was the one to the client.

It flattens how sure you were

Hedges carry information. "I think the forecast might be off" and "the forecast is off" are different claims with different consequences for the person who sent them. Rewriting tools treat hedges as noise. They are trained to produce confident, professional text, and confidence is what they add.

The person on the other end reads the confidence as yours. When the forecast turns out to be fine, "I think it might be" was a reasonable flag. "It is" was a wrong call, and it is on your record, in your words, except they were not your words.

It taxes the speed it sold you

Dictation is faster than typing because you talk faster than you type. That gain only survives if you can send without a full read-through. A formatting tool earns that trust because its changes are mechanical: filler gone, punctuation in, layout matched to the app. You can scan and send.

A rewriting tool cannot earn it, because its changes are editorial. It may have changed the meaning, so you have to read it all, every time. The proofreading you skipped by not typing comes back as proofreading you cannot skip. This is the wall we think most people hit when AI dictation stops feeling fast: the tool was fast, and they were not.

What formatting means, and where it stops

Slack, filler gone, nothing rewritten:

Said:

“

um before I log off, I wanted to let you know that the vendor portal password has been rotated so grab the new one from the vault ... ummm basically the ticket queue is clean except two escalations that need a response by tomorrow morning ummm I flagged both in the channel

Sent:

“

Before I log off, I wanted to let you know that the vendor portal password has been rotated, so grab the new one from the vault. Basically, the ticket queue is clean except for two escalations that need a response by tomorrow morning. I flagged both in the channel.

Gmail, greeting and sign-off added, nothing else:

Said:

“

hi Brendan hope you're doing well I wanted to check if the contract came back from legal yet also do we have a date for the kickoff thanks Sarah

Sent:

“

Hi Brendan,

Hope you're doing well. I wanted to check if the contract came back from legal yet. Also, do we have a date for the kickoff?

Thanks Sarah.

What formatting does not do: it does not decide the handoff note should sound more buttoned-up, and it does not drop "basically" because a rewriter thinks it reads as hedgy. It does not add "Let me know if you have any concerns" to the Brendan email because emails usually end that way. You said none of that, so none of it gets sent.

That is the whole product. Rubil formats your words. It never rewrites or generates them. If the output is wrong, it is wrong in a way you can see in a glance, because every word in it is one you said.

When rewriting is the right default

Some people want a tool that writes. If you dictate rough notes and want a finished paragraph, if you are drafting in a second language and want the phrasing smoothed, or if you want a "make this formal" button, a rewriting tool is the correct choice and Rubil is the wrong one. Willow builds for exactly that person and says so on its homepage. Wispr's per-app tone adaptation is aimed there too.

The argument here is about the default for someone whose messages already say what they mean. That person is paying for a proofreading tax and a small, permanent risk to their own credibility, in exchange for prose that is smoother than they are. We think that is a bad trade, and we think it should be a choice you make on purpose.

Privacy

Rewriting needs context, and context has to come from somewhere. Rubil detects which app you are in without reading what is on your screen. On Chrome it reads the tab's hostname and a few attributes of the text box you clicked. On Mac it reads the identifier of the app in focus, plus the active tab's address if that app is a browser. On Windows it reads the program name and window title. None of that leaves your device. It never reads page content and never takes screenshots.

Your audio is processed for transcription and discarded. Your transcript is sent for formatting and is not stored on Rubil's servers. Your Glossary is stored, encrypted in transit and at rest, so it can follow you between devices. Every processor is named on the privacy page. Rubil does not offer fully offline processing. If that is a requirement, other tools do it and we would rather you know now.

The measure that matters

Accuracy tells you how many words a tool heard right. It cannot tell you whether the output was fit to send, which is the only thing you care about at message eighty. We publish the number that does: Post-Edit Rate, the share of dictations that needed at least one edit, live at rubil.io/quality. The definition is in What is Post-Edit Rate?.

If you want to see the difference before installing anything, try it in your browser. Say something the way you would say it to a colleague, with the ums in, and watch what comes out. Then check whether every word was yours. For a side-by-side with the tools quoted above, see Rubil vs Wispr Flow, and for what formatting looks like in one specific app, voice dictation for Slack.