Tidied up,
not churned out
Every AI dictation tool puts a language model between your mouth and the cursor — the same kind of model that fills the internet with "AI slop". Blabb's cleanup is tuned to take things away, never to add polish. This page documents the rules we wrote to keep chatbot flavour out of your rewrite, how each one was tested before it shipped, what that testing cost in accuracy (measurably: nothing), and where the honest limits are. As with everything cleanup-related: out of the box Blabb is set to Keep original, so no rewriting happens until you choose a style.
Your words pass through an AI. Slop is what leaks in.
"Slop" is the term that settled for AI-generated content that reads like it came off a conveyor belt: upgraded vocabulary, tidy summaries nobody asked for, invented detail, relentless evenness. Simon Willison brought the word into the mainstream in May 2024, defining it as content that is "unrequested and unreviewed". By 2025 it was Merriam-Webster's Word of the Year — the American Dialect Society picked it too. This is no longer a niche complaint; it has a dictionary entry.
Dictation is unusually exposed
Most slop is written by people prompting a chatbot. Dictation is different: you speak, and a model rewrites your sentence before it reaches the cursor. That rewrite step is exactly the machinery that turns "we use the script every morning" into "we utilise the script each morning". You said use. It wrote utilise. Nobody asked.
Recipients can tell
A Harvard Business Review study on workplace "workslop" found 40% of employees had received AI-generated content from colleagues, and each incident took about two hours to deal with. One researcher's phrase for the flavour of it: "collapsing to the mode… lukewarm emptiness." If your email, notes, or chart entries read generated, that is the mode they collapse to.
Stakes are higher in the professions
A clinical note or a matter summary that reads like chatbot output erodes trust in the record itself — and the rewrite is where precision dies quietly: a hedged finding made confident, "twice daily" turned into "BID", an invented date added to a deadline. These are not hypothetical. They are failure modes we reproduced in our own testing, which is why the rules below exist.
Your speech is already human. So the cleanup subtracts.
Spoken language arrives with everything slop lacks: specific names, real numbers, hedges, stumbles, opinions, texture. A rewrite that "improves" it is doing damage. So the cleanup model's job is defined as removal and repair only — take out fillers and false starts, fix grammar and punctuation, format the numbers you spoke — and add nothing: no vocabulary, no summary, no balance, no positivity, no sign-off.
We don't ban "AI words" — that breaks dictation
The obvious anti-slop move is a blacklist of chatbot vocabulary. We tested that idea and rejected it, because you are allowed to say leverage, crucial, robust, or strategic — people say them constantly. A model told those words are forbidden will rewrite vocabulary you actually used, which is a worse harm than slop: it changes what you said.
The rule is "never introduce, never substitute"
Every guard is phrased in one direction: the model may not add the flavour. If you said leverage, it stays leverage — same word, same formality, same bluntness. If you didn't say it, it doesn't appear. Your vocabulary is yours; the model's vocabulary doesn't get a seat at the table.
And it's enforced where enforcement works
Some habits are too deep for instructions — we proved that the hard way in testing. Those move out of the prompt and into deterministic code: a pass that runs before anything is typed maps non-keyboard typography (em dashes, smart quotes, odd Unicode spacing) to plain keyboard equivalents, and machine-shaped tokens like phone numbers and file hashes are set aside and restored character-for-character. More on that in the honest-edges section.
Five guards shipped. Each one earned its place.
These are the rules as they run today, each with the failure it exists to stop. They were candidates once — eight were tested, five survived testing, three were cut. What follows is the surviving five.
1. No invented detail
The rewrite may not add a date, time, amount, name, or citation you didn't say — and may not swap your wording for professional shorthand. "Expired in twenty twenty four" becomes 2024, never April 30, 2024. "Twice daily" stays twice daily, never BID. And the output ends where you stopped: no wrap-up sentence, no "in summary", nothing forward-looking you didn't dictate.
2. No vocabulary swaps — either direction
Plain words stay plain: use not utilise, is not serves as, start not commence. Formal words stay formal: if you said leverage or crucial, they survive untouched. The only permitted rephrasing is what your chosen style genuinely needs for concision or register — never a swap made purely to change vocabulary. Converting your spoken numbers to digits is a required correction, not a vocabulary swap.
3. Your stance survives
The rewrite may not soften a complaint, criticism, or blunt statement into diplomatic neutral-speak. If you said the delivery was a disaster, it stays a disaster. Grammar gets fixed; your position does not. A negative statement stays negative — the same never-flip rule that protects a "no" from becoming a "yes".
4. Your certainty survives
"I think", "probably", "maybe", "not sure" are meaning, not mess. The rewrite may not strip them to make a tentative claim sound confident, and may not add them to make a confident claim sound tentative. One allowance, tested both ways: a redundant stack of the same hedge — "maybe perhaps I think" — may trim to one. The uncertainty must survive; the stutter may not.
5. Your voice stays yours
"I approved the budget" is never recast as "a decision was made". First-person active statements stay first-person active; passive you actually used stays passive. And the output never turns around and addresses you — you're the speaker, not the audience.
One rule at a time, or you learn nothing
Anti-slop rules are easy to write and easy to get wrong: every "don't" you add is also a new way the model can over-correct and start mangling clean input. So no rule shipped on conviction. Each went through the same loop, alone.
Write the failing test first
Before a rule exists, the slop it targets gets a detector test — a dictated scenario that can only fail if the model does the bad thing. Some of those tests failed on the first run, which is how we knew they were testing something real. The ones that passed immediately stayed anyway: free armour against future regressions.
Measure, add, measure again
Run the suite before the rule. Add exactly one rule. Run it again after. Read every changed result by hand, because a small model's output varies run to run and a percentage alone will lie to you. Keep the rule if the targeted slop died and nothing measurably broke; revert if it didn't. Three candidates died this way — including a prompt-level ban on em dashes and exotic Unicode that the model simply ignored, even with the characters named in the instruction. It got cut, and the fix moved to the deterministic typing layer where it actually works.
Then double the length
Short sentences are easy to keep honest; paragraphs are where polish creeps in. So the scenario suite grew to 440 general cases — 220 of them roughly double the original length, dense multi-sentence dictations with real names, figures, and disfluencies — plus the medical and legal domain suites. Longer text is exactly where the rules have to hold.
The safety cost nothing we can measure
This is the question that matters: guardrails that quietly make the rewrite worse are not guardrails. With all five rules in and the suite at double length, every writing style held the score band it had before the anti-slop work began — concise 96.4%, friendly 97.5%, professional 94.8% of 440 scenarios each, medical 91.1% and legal 95.7% of their domain suites. Within the run-to-run variation of a small model, the rules cost no accuracy. They only removed failures.
What the rules caught
Invented boilerplate in legal phrasing — gone. Clinical shorthand like "PO BID" substituted for your spoken words — gone. Certainty-stripping, where "I think it's probably fine" came back as a confident assertion — caught and corrected, with the hedge-stack allowance kept honest in both directions.
What doubling the length exposed
The long-form run surfaced slop that short tests can't see: labels invented out of nowhere ("Bullet points:", "Headline:"), accented spellings you never typed ("rocio" becoming "Rocío"), and vocabulary upgrades resurfacing under pressure on long inputs. Every one of those is now a pinned, failing test — which is the correct state for a known issue: visible, measured, and feeding the deterministic checks rather than lurking.
Why we'll say "held" and not "improved"
A small model's scores move a point or two between identical runs. Claiming a gain inside that noise would be marketing, not measurement. What we can stand behind is stronger: the targeted slop is gone at the case level, and the suite totals didn't move. Nothing was traded away.
What prompt rules can't do — and what we refuse to claim
Instructions shape a small model; they don't rewire it. Some habits survived being named explicitly in the prompt — odd spacing before units, a stray semicolon — and for those, Blabb doesn't rely on the prompt at all. Before anything is typed, a deterministic pass maps non-keyboard typography to keyboard equivalents, so em dashes, smart quotes, and exotic Unicode spaces can't reach your cursor. Machine-shaped tokens — phone numbers, hashes, licence keys — are set aside and restored character-for- character, with your raw words as the fallback if a token can't be proven whole. That's the pattern throughout: prompts where prompts work, code where they don't.
Two things we won't claim. First, we don't chase AI-detector scores and won't promise your text "passes as human-written" — detectors are unreliable moving targets, and gaming them isn't what this is for. The goal is narrower and checkable: the output stays your words, in your voice, with nothing added. Second, none of this is a safety net watching every rewrite: cleanup is off by default (Keep original until you choose a style), the design rules run only when a style is active, and the optional Protect meaning check that compares rewrite against original is off by default too. The broader faithfulness story — permitted edits, injection handling, the testing behind all of it — is on its own page: how we tune for faithfulness.
Your words, cleaned up. Still yours.
Two weeks free, models in the installer, dictation that runs on your PC.
FREE TIER 2,000 WORDS/DAY · 14-DAY UNLIMITED TRIAL IN-APP · CANCEL ANY TIME