We Detect When AI Hallucinates Your Words
You dictate: “the auth endpoint needs rate limiting set to 100 requests per minute per org.”
You expect the AI to clean up filler words and fix grammar. Instead, it rewrites the whole sentence, changes “100 requests” to “appropriate limits,” and adds a paragraph about best practices you never said.
This is hallucination in text cleanup, and it's more common than you'd think. The LLM treats your dictation as a prompt to riff on, not text to preserve. Resonant detects this and blocks it.
The problem
Cloud text cleanup routes your raw transcript through an LLM to fix speech recognition errors, remove filler words, and improve formatting. This works well 95% of the time. The other 5%, the LLM does one of these:
- •Rewrites your meaning. Changes specific numbers, names, or technical terms to more “natural” alternatives.
- •Adds content. Inserts sentences, suggestions, or explanations you never said.
- •Drops content. Removes phrases it considers redundant but that carried important nuance.
- •Prefixes with meta-commentary. Starts with “Here's a cleaned up version:” or similar framing instead of just returning the text.
Any of these silently changes what you said. The text looks polished, but it's no longer yours.
Our guardrails
Resonant's cloud processing pipeline includes three automated checks that run on every cleanup response before the text reaches your clipboard:
Bad prefix detection
Checks whether the response starts with LLM meta-commentary — “Here's,” “Sure,” “I've cleaned,” “The corrected version,” etc. If detected, the cleanup is rejected and the raw transcript is used.
Coverage check
Measures what percentage of your original words appear in the cleaned output. If coverage drops below a threshold, the LLM removed too much of what you said. The cleanup is rejected.
Content-novelty ratio
Measures how much new content the LLM added that wasn't in the original transcript. Formatting changes and minor word additions are fine. Entire new sentences are not. Above a threshold, the cleanup is rejected.
When any guardrail triggers, Resonant falls back to the raw transcript with basic formatting applied. You get slightly rougher text, but it's faithfully what you said.
Context-specific prompts
The cleanup prompts themselves are tailored to prevent hallucination. Each mode — email, message, and general text — uses a prompt with hard constraints:
- •Do not add information that wasn't in the original transcript.
- •Preserve all specific numbers, names, and technical terms exactly as spoken.
- •Return only the cleaned text. No commentary, no framing, no suggestions.
- •If the input doesn't contain meaningful content, return it unchanged rather than inventing a response.
Prompts plus guardrails. Belt and suspenders. The prompts tell the LLM what to do. The guardrails catch it when it doesn't listen.
Why this matters
Trust is the foundational requirement for dictation. If you can't trust that the text in your clipboard is what you said, you have to re-read and verify every dictation. That verification step destroys the speed advantage of voice.
Our goal is that you dictate, glance, and move on. The hallucination guardrails exist so that glance is enough.