Draft From an Interview Transcript With AI, Keep the Voice
To draft an article from an interview transcript with AI without losing the expert's voice, build a voice fingerprint from the transcript first, feed AI only cleared material, draft section by section against the quote map, and run a flattening check that compares every kept claim and phrase back to the transcript.
To draft an article from an interview transcript with AI without flattening the expert's voice, do the voice work before the drafting work: extract a voice fingerprint from the transcript — the expert's recurring phrases, specific numbers, named tradeoffs, and the sentences only they would say — then let AI organize and connect only material you have already marked as cleared, one section at a time, and check every AI-touched paragraph against the transcript before you keep it. Flattening happens when the model is asked to rewrite; it mostly does not happen when the model is asked to arrange.
The failure mode this protocol prevents is familiar to anyone who has pasted a transcript into a chat window and typed "turn this into a blog post." The output is fluent, organized, and interchangeable with a thousand other posts. The expert's "we burned two weeks because the staging data was three schema versions old" becomes "it's important to keep test environments up to date." That sentence is true, generic, and worthless — the interview existed to buy you the first version, and the model quietly refunded it.
Who this is for
This is for writers, editors, and content leads who already have a usable transcript — from an interview run against a run sheet that captured specifics and scoped by an interview content brief — and who want AI to shoulder the assembly work without sanding off everything the interview was for.
It assumes quote permissions and fact-checks are handled as their own step; that clearance workflow is covered in quote permissions, attribution, and fact-checking from interview to published article. Nothing below overrides it: AI drafts from cleared material only.
Why does AI flatten the expert's voice?
Because language models are trained to produce the most probable next words, an unconstrained rewrite regresses toward the average register of published business prose — the mechanism, not the vibes. Three specific things get averaged away:
- Idiolect. The expert's actual phrasing — odd metaphors, industry shorthand, sentence rhythm — is statistically unusual, so a rewrite replaces it with the median way of saying the same thing. Nielsen Norman Group's tone-of-voice work describes voice along dimensions like formality, humor, and enthusiasm; a model rewrite drags every dimension toward the middle.
- Specificity. Numbers, product names, dates, and one-off events are the least predictable tokens in a sentence. Summarization abstracts them into categories: "three schema versions old" becomes "outdated."
- Asymmetry. Real experts hold lopsided opinions — strongly for one tool, dismissive of a popular practice. Models hedge toward balance, converting a defended position into "both approaches have merits."
Once you see flattening as regression to the mean, the countermeasure is obvious: never give the model rewrite authority over the material that is valuable because it is far from the mean. Pin that material down first.
How do you keep the expert's voice in an AI-assisted draft?
Pin the voice down before the model sees anything: build a fingerprint sheet of protected material, restrict AI to assembly tasks, draft section by section, and check every AI-touched paragraph against the transcript. The artifact on this page is the voice-preservation drafting protocol: a voice fingerprint sheet, an allowed/forbidden task table for AI, a section-by-section drafting sequence, and a 10-point flattening check. Copy the three tables and run the numbered sequence for each interview-led article.
1. Build the voice fingerprint sheet from the transcript
Read the transcript once, before any AI touches it, and fill this sheet. Everything you list becomes protected material the model may move but never reword.
| Fingerprint field | What to capture | Example of what qualifies |
|---|---|---|
| Signature phrases | Expressions the expert repeated or coined | A nickname for a failure mode, a rule of thumb in their words |
| Load-bearing specifics | Numbers, dates, versions, durations, names | "Two weeks," "three schema versions," a named tool |
| Defended positions | Claims where the expert picked a side and gave a reason | "I'd never run this migration on a Friday, because…" |
| Story beats | The concrete incidents, in sequence | What broke, how they noticed, what they tried first |
| Register notes | How they talk: dry, blunt, profane, careful | Keep it; do not "professionalize" it beyond permissions |
| Do-not-touch quotes | Cleared direct quotes, exactly as cleared | Copy verbatim from the permission log |
Ten to twenty rows is normal for a one-hour interview. If you cannot fill ten rows, the problem is upstream — the interview did not yield usable material, and no drafting technique fixes that.
2. Set the AI task boundary before the first prompt
Decide what the model is allowed to do and paste the boundary into every drafting prompt. This table is the working default:
| AI may | AI may not |
|---|---|
| Cluster transcript passages by theme | Reword anything on the fingerprint sheet |
| Propose section order against the brief's outline | Generate or extend quotes, even "in the expert's style" |
| Write connective tissue between protected passages | Add claims, examples, or numbers not in the transcript or sources |
| Tighten your own linking prose | Balance a defended position with an invented counterpoint |
| Flag claims that look source-sensitive | Decide a claim is true; verification stays human |
| Suggest headings phrased as reader tasks | Change hedged statements into confident ones, or the reverse |
The second column is not a style preference. Invented quotes and smoothed positions are accuracy failures, and the same boundary that protects voice protects you from publishing fabrication. Reuters' interview guidance treats quotations as sacrosanct for the same reason: what runs inside quote marks is the source's, not the writer's.
3. Draft section by section against the quote map
Work from the quote-to-section map in your interview brief, one section per prompt. Give the model: the section's reader job, the protected passages assigned to that section (marked clearly as verbatim), and permission to write only the connective prose around them. Assembly prompts produce drafts where the expert material sits intact, in the expert's words, in the right place.
Drafting the whole article in one prompt reliably fails because the model optimizes for global coherence — and global coherence is exactly the pressure that rewrites protected material to match the surrounding register. Small prompts keep the model on connective tissue.
4. Run the flattening check
Before edit review, compare the draft against the transcript with this 10-point check. Any "yes" sends the section back.
- Did any number, date, version, or duration become a vaguer category word?
- Did any signature phrase get replaced by a synonym?
- Does any quote differ by even one word from the cleared version?
- Did a defended position acquire a "however, others argue" it never had?
- Did a story lose its sequence — what broke, how they noticed, what they tried?
- Did hedged statements become confident, or confident ones become hedged?
- Did the expert's register shift toward generic professional prose?
- Does any paragraph read as advice that could exist without this interview?
- Did AI-typical connective phrases creep in where the expert's transitions existed?
- Could the expert read this draft aloud and recognize themselves?
Point 8 is the roll-up test. A paragraph that survives points 1–7 but fails point 8 usually means the section was built from the model's general knowledge instead of the transcript — cut it or rebuild it from fingerprint rows.
5. Hand off to clearance and editing as usual
The draft now goes through your normal pipeline: quote clearance against the permission log, claim checks against the source ledger, and the house style pass. The editorial style guide governs everything the fingerprint sheet does not; where they conflict, the fingerprint wins inside quoted and attributed material, and the style guide wins in your connective prose.
Worked example: one passage, three treatments
Assumption: the passage below is our own sample material, not a real quote. Suppose the transcript says: "We burned two weeks because staging was three schema versions behind prod. Nobody looked. The dashboard was green the whole time."
- Flattened (what a summarize prompt returns): "Keeping staging environments synchronized with production is critical, as outdated test data can cause significant delays despite passing status checks."
- Overcorrected (fake voice): "We torched half a month because staging was hopelessly ancient — classic, right?" The model invented intensity and a verbal tic the expert never had; this fails the clearance step as a misquote.
- Preserved (assembly with connective tissue): The dashboard said everything was fine. It wasn't. "We burned two weeks because staging was three schema versions behind prod," followed by your prose explaining what readers should check so their own green dashboard is telling the truth.
The preserved version is the only one that keeps the incident's evidentiary value — and the only one Google's people-first guidance would call first-hand experience rather than restated consensus. Google's AI-era guidance points the same direction: unique, useful content is the durable input, and a transcript full of protected specifics is about as unique as inputs get.
Mistakes that flatten voice even with a good protocol
- Feeding the raw transcript instead of the fingerprint sheet plus assigned passages. The model treats everything unmarked as rewritable.
- Letting AI "clean up" quotes for grammar. Cleanup is a human clearance decision made with the permission log open, not a drafting default.
- Accepting a beautiful paragraph that cites nothing from the interview. Fluency is not evidence; delete it or source it.
- Running one flattening check at the end instead of per section. Flattening compounds — later prompts imitate earlier flattened sections.
- Professionalizing register without asking. If the expert is blunt, the article should be blunt within the bounds they cleared; ask them, not the model.
- Using AI to fill gaps the interview left. A gap means a follow-up question or a narrower article, never synthesized expertise — the same rule the interview brief enforces before drafting.
FAQ
Can AI draft an article from an interview transcript at all?
Yes, and it is good at the assembly parts: clustering passages, proposing structure, and writing connective prose around material you have protected. It fails when given rewrite authority over the transcript itself, because rewriting regresses unusual phrasing and specifics toward generic prose.
What does "flattening the expert's voice" actually mean?
It is the loss of the three things interviews exist to capture: the expert's own phrasing, their load-bearing specifics, and their asymmetric positions. A flattened draft is fluent and true but could have been written without the interview.
Should I paste the whole transcript into the prompt?
Only for the clustering step, where the model groups passages by theme. For drafting, feed one section's cleared passages at a time with the fingerprint sheet, so the model never has license to rework protected material for global smoothness.
How do I keep AI from inventing or extending quotes?
State it as a hard boundary in every prompt, mark protected passages as verbatim, and verify at the flattening check: every quoted string in the draft must match the cleared version character for character. Anything else is treated as a misquote, not a style issue.
Does using AI for assembly need to be disclosed to the expert?
Tell the expert how their material will be handled if they ask, and honor whatever handling conditions were set during permissions. The accountable facts — quotes accurate, claims sourced, attribution as agreed — are the same regardless of what tooling produced the connective prose.
Where this fits in the interview pipeline
This is the drafting stage — the last step before clearance and publication. Upstream: the interview content brief decides what the article must prove, and the interview run sheet gets the specifics onto the transcript in the first place. If one conversation should become several articles, split it with the interview-to-article splitting map before you draft, so each draft has its own fingerprint sheet. Downstream, every draft clears through quote permissions, attribution, and fact-checking. The broader standards — what experience means when you are not a famous expert and how to source claims without making them boring — apply to the finished page like any other.
Sources and last-reviewed notes
Last reviewed: 2026-09-14. Sources checked: Google guidance on helpful, people-first content and its AI-search guidance on unique, useful content; Nielsen Norman Group's tone-of-voice dimensions for the voice-register framing; and Reuters' interview guidance on the handling of quotations. The fingerprint sheet, task table, flattening check, and every transcript passage in the worked example are our own illustrations, not real interview material.