Answer in brief: Seventeen of 22 candidate themes appeared across both prompt variants; five were unstable and were excluded from the conclusion.
Can a startup use AI to organize interview notes without inventing consensus? A useful answer has to be narrower than a product claim. This article tests a bounded workflow, publishes the scoring surface, and keeps consequential approval with a person. It does not turn a controlled result into personalized financial advice.
What we tested or analyzed
We created 12 fictional interview transcripts, removed direct identifiers, and asked OpenAI GPT-5.6 in a Codex editorial session for themes using two prompt orders. We traced each retained theme to source excerpts.
The original asset is a synthetic interview corpus, coding table, and prompt-stability test. The complete machine-readable table is available as CSV. The evidence visual below summarizes the primary criterion; its values are also written in text and shown in the table, so the chart is not the only way to obtain the result.
The measured result
Seventeen of 22 candidate themes appeared across both prompt variants; five were unstable and were excluded from the conclusion.
The row-level outcome distribution was stable: 17, unstable: 5. Those labels are deliberately more descriptive than one blended score. A partial, review, stale, exception, or unsupported row can carry a different operational risk from a plainly wrong row, so the CSV preserves the reason beside the disposition.
| Item | Outcome | Evidence or note |
|---|---|---|
| Theme 01 | stable | synthetic interview coding |
| Theme 02 | stable | synthetic interview coding |
| Theme 03 | stable | synthetic interview coding |
| Theme 04 | stable | synthetic interview coding |
| Theme 05 | stable | synthetic interview coding |
| Theme 06 | stable | synthetic interview coding |
| Theme 07 | stable | synthetic interview coding |
| Theme 08 | stable | synthetic interview coding |
| Theme 09 | stable | synthetic interview coding |
| Theme 10 | stable | synthetic interview coding |
| Theme 11 | stable | synthetic interview coding |
| Theme 12 | stable | synthetic interview coding |
| Theme 13 | stable | synthetic interview coding |
| Theme 14 | stable | synthetic interview coding |
| Theme 15 | stable | synthetic interview coding |
| Theme 16 | stable | synthetic interview coding |
| Theme 17 | stable | synthetic interview coding |
| Theme 18 | unstable | synthetic interview coding |
| Theme 19 | unstable | synthetic interview coding |
| Theme 20 | unstable | synthetic interview coding |
| Theme 21 | unstable | synthetic interview coding |
| Theme 22 | unstable | synthetic interview coding |
Reading the evidence row by row
- Theme 01 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 02 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 03 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 04 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 05 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 06 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 07 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 08 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 09 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 10 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 11 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 12 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 13 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 14 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 15 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 16 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 17 was recorded as stable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
- Theme 18 was recorded as unstable. The evidence note is “synthetic interview coding”; the disposition remains visible so it cannot be averaged away.
The expected label or control was fixed before review. The visible note explains why the row received its disposition. The chart uses the published primary criterion, but the table is authoritative because it preserves exceptions that a single percentage would hide.
How to reproduce the check
- Download the CSV and read its labels, units, and synthetic/public-data notice before using it.
- Write the expected answers or decision rule before looking at a model response.
- Use the same bounded prompt and record the model or tool, access surface, and date.
- Preserve the raw response. Break prose into atomic claims rather than grading the tone of the whole answer.
- Recompute arithmetic with deterministic formulas and verify definitions against the linked primary sources.
- Record correct, partial, wrong, uncertain, and refused outcomes separately. Do not silently repair the model output before scoring it.
- Repeat material checks after a model, source, or workflow changes.
What the result means
The value of this result is diagnostic. It shows where a structured assistant can reduce search, formatting, or first-pass review work. It does not transfer responsibility for the underlying decision. A “pass” means the row met the published rule in this test, on this date, with these inputs.
The errors and open items matter more than a polished average. In money and business workflows, one missed assumption, stale fact, false match, or overconfident definition can dominate many correct low-risk rows. That is why the artifact keeps row-level outcomes and why a human reviews exceptions rather than receiving only a percentage.
Reproducibility also has limits. A reader can repeat the steps and inspect the same answer key, but a probabilistic model may not return identical wording. A useful rerun should therefore compare atomic claims, calculations, citations, and escalation decisions—not superficial phrasing.
Why this topic needs its own boundary
A small team can move from draft to customer-facing material quickly, which makes invented evidence particularly dangerous. Interview themes, traction, market claims, and policy exceptions need a trace back to an approved source and an accountable reviewer.
The workflow favors small, inspectable artifacts: coded excerpts, slide-level comments, explicit prohibited uses, and named owners. That structure makes uncertainty visible and prevents a polished narrative from silently becoming evidence.
A safer operating workflow
- Get appropriate participant consent.
- Remove direct identifiers.
- Preserve source excerpts.
- Require traceability for each theme.
- Do not turn frequency into demand without further evidence.
How each control changes the decision
Control 1: Get appropriate participant consent. For this test, that control answers the bounded question “Can a startup use AI to organize interview notes without inventing consensus?” without extending the result into an untested decision.
Control 2: Remove direct identifiers. For this test, that control answers the bounded question “Can a startup use AI to organize interview notes without inventing consensus?” without extending the result into an untested decision.
Control 3: Preserve source excerpts. For this test, that control answers the bounded question “Can a startup use AI to organize interview notes without inventing consensus?” without extending the result into an untested decision.
Control 4: Require traceability for each theme. For this test, that control answers the bounded question “Can a startup use AI to organize interview notes without inventing consensus?” without extending the result into an untested decision.
Control 5: Do not turn frequency into demand without further evidence. For this test, that control answers the bounded question “Can a startup use AI to organize interview notes without inventing consensus?” without extending the result into an untested decision.
Keep data collection, model preparation, deterministic validation, and approval as separate stages. Use the least sensitive input that can answer the question. If removing personal or confidential data makes the result ambiguous, route the case to an approved human process instead of restoring secrets to an unapproved tool.
Calculations need an independent formula; current facts need a current primary source; classifications need an “uncertain” route; and irreversible actions need explicit authorization outside the model. Logs should capture the version, prompt, source date, output, reviewer, correction, and final disposition without retaining unnecessary personal data.
Limitations and professional boundary
Synthetic interviews cannot validate a market. Theme frequency is not customer importance, and small samples can amplify selection bias.
This publication provides general educational information. It does not know a reader’s finances, duties, jurisdiction, contracts, tax treatment, credit position, or risk tolerance. A qualified financial, accounting, tax, legal, lending, security, or other professional should review decisions with material consequences.
Primary sources
Verified 2026-07-26. Primary-source links can change; use the publication date and linked source to check for a newer version.
Bottom line
Seventeen of 22 candidate themes appeared across both prompt variants; five were unstable and were excluded from the conclusion. The practical lesson is to make AI produce inspectable work inside a controlled process—not to make fluency the final control.