The Voice Audit: Catching AI Patterns Before Your Client Does

Your client knows what AI writing looks like. They may not be able to articulate it precisely, but they know when they are reading it. The goal of a voice audit is to catch the patterns they would notice before they see the piece.

A voice audit is not a general quality pass. It is a systematic check for a specific list of patterns that mark text as machine-generated. It runs after the substance pass, when you already know the facts are right and the argument holds. The question at this stage is whether the piece sounds like a person wrote it.

The checklist

Hedging phrases that do not add information. "It's important to note that," "it's worth mentioning," "as with any approach, results may vary," "it depends on your specific situation." Flag every one. Most can be cut entirely. A few can be replaced with the actual claim the writer was hedging around.

Identical paragraph length. Read the piece and note whether all the paragraphs are roughly the same number of sentences. If they are, the piece was generated with a regularity bias. Some paragraphs should be longer because they carry more weight. Some should be one or two sentences because the point does not require more.

Symmetrical argument structure. If every section makes exactly three points, or every argument has exactly one counterargument, the model imposed a structure for balance rather than following the actual shape of the argument. Uneven structure is human. Perfect balance is not.

The dead connector paragraph. Any paragraph whose function is to summarize what just happened and announce what comes next. "Now that we have covered X, it is time to explore Y." These exist because the model generates transitions rather than writes them. Cut them.

The positive-hedging opening. "There is no one-size-fits-all approach." "Every writer's situation is unique." "It depends on your goals." If the piece opens or closes with any variation of these, the opening or closing was generated by a model trying not to say anything falsifiable.

Generic claims where specific ones belong. "Many content writers struggle with consistency." Struggle with what kind of consistency? For which kind of output? At what scale? If a claim applies equally to everyone in a category, it applies to no one in particular. Find the specific version and use it.

How to run the audit

Read the piece once for the list above, making notes. Do not edit during this pass. You are cataloguing problems, not fixing them. Editing during a read interrupts the ability to see patterns across the whole piece.

After the read-through, go back and fix in order of severity. Hedging phrases are usually the highest-concentration problem and fix quickly. Structural issues take longer because they require decisions about what to cut or expand.

If the piece is going to a client who has seen your work before, add one more check: read it against something you have written yourself at roughly the same length. Do the sentence lengths vary the same way? Are the transitions handled similarly? The comparison surfaces things the systematic checklist misses because it is calibrated to your specific voice, not to generic AI patterns.

The failure mode

The voice audit catches AI patterns in AI output. It does not fix a piece that had no argument to begin with. A piece that is well-edited for voice tells but still says nothing specific will sound like a polished version of nothing in particular.

Voice and substance are separate problems. The audit addresses voice. The brief addresses substance. When clients push back on AI-assisted work, they are usually catching a combination of both. Fixing only the voice leaves the substance problem intact. That is why the audit runs last, after the substance pass has confirmed the piece is worth polishing.