AI Podcast Source Accuracy: A Fact-Checking Checklist

An AI-generated podcast can sound confident before it is correct. Conversation adds flow, examples, and transitions; those same additions create opportunities to widen a claim, merge two sources, or turn uncertainty into a neat conclusion. Voice quality makes the result easier to trust, not more trustworthy.

Source accuracy review is therefore a separate production stage. The reviewer’s job is not to decide whether the episode sounds plausible. It is to trace every material statement back to the evidence and confirm that the spoken version preserves scope, magnitude, and uncertainty.

Freeze the review package

Collect the exact inputs used for generation:

• source files or saved web pages;
• version, date, author, and URL or DOI;
• cleaned or OCR-derived text;
• generation instructions;
• script;
• final rendered audio;
• final transcript.

Do not review against a newer source without labeling the difference. If the script came from draft v3 and you check final v5, you may approve a statement that was unsupported at generation time.

Give the package an identifier. Store corrections beside it. This turns “we checked it” into a reproducible record.

Separate claim types

Mark the script using four labels:

• Q — quotation: exact words attributed to a source.
• P — paraphrase: the source’s meaning in new wording.
• I — inference: a reasonable conclusion drawn from the source.
• T — transition or teaching example: connective material created for the episode.

Each type has a different test. A quotation must match exactly. A paraphrase must retain meaning and boundaries. An inference must be identified as interpretation. A teaching example must not be presented as an event that actually happened.

This labeling is especially useful in two-host scripts. Casual phrases such as “what happened was” can transform an illustrative scenario into a factual claim.

Verify numbers with their full frame

For every important number, check:

1. value and sign;
2. unit and denominator;
3. time period;
4. population or segment;
5. comparison baseline;
6. whether it is measured, estimated, projected, or illustrative;
7. uncertainty or range.

“Conversion increased 20 percent” may mean a relative increase from 5 to 6 percent, a twenty-percentage-point increase, or a forecast. Those are different claims.

Listen to the rendered audio, not only the script. Speech systems can drop a decimal pause, pronounce a range as subtraction, or read an acronym as a unit. For high-risk numbers, have a second reviewer follow the source while listening.

Check scope words

Small words carry scientific and policy meaning:

• can, may, tends to, is;
• some, most, all;
• associated with, predicts, causes;
• in this sample, generally;
• preliminary, final;
• not, no, except.

Generated summaries often compress these distinctions. Search the transcript for absolute language such as “always,” “proves,” “everyone,” “safe,” and “guarantees.” Do not ban the words mechanically; demand source support.

Confirm that population and setting remain attached to the result. Evidence from one country, age group, dataset, or software version should not silently become universal advice.

Compare claims with surrounding context

Finding a similar sentence in the source is not enough. Read the paragraph before and after it. Check footnotes, figure captions, definitions, and the limitations section.

A report may state a headline result and immediately explain that data quality was poor. A policy may grant permission only under conditions listed in the next clause. A paper may describe a subgroup result as exploratory. Context determines whether the podcast’s wording is fair.

When the source uses a table, verify the row and column labels. When it uses a chart, verify axes and legends. Do not rely on OCR output for a critical decimal, minus sign, or superscript without viewing the page image.

Audit quotations and attribution

Search each quotation in the source. Confirm wording, speaker, date, and location. Ellipses must not remove a qualification that reverses meaning. Shorten quotations for audio, but never join distant fragments as one sentence.

For paraphrases, identify the source explicitly enough that a listener can distinguish:

• what the document says;
• what a cited third party says;
• what a host concludes.

Do not attribute generated language to an author. Do not invent a customer story, expert reaction, or personal experience to make the dialogue warmer. If an example is hypothetical, say so.

Compare multiple sources without blending them

When an episode uses several sources, create a claim ledger:

| Claim | Source | Location | Review status | | --- | --- | --- | --- | | Definition of active user | Analytics spec v2 | §3.1 | Verified | | Q2 retention value | Board report | Table 4 | Verified | | Cause of decline | None | — | Remove or label hypothesis |

Two sources may use the same term differently. Define each before comparing. Publication dates matter: a later document can supersede an earlier one, but that does not mean the earlier author knew the later result.

If sources conflict, preserve the conflict. Do not average them into a smooth answer unless a valid method justifies doing so.

Review uncertainty and limitations

Every episode does not need a gloomy disclaimer. It does need the limitations that change interpretation or action. Ask:

• What cannot be concluded from this design?
• What data are missing?
• Which assumption drives the result?
• Is the estimate precise enough for the recommendation?
• Did the source itself flag a competing explanation?

Place the relevant limitation near the claim, not only in a closing block. Listeners may not retain a generic warning delivered ten minutes later.

For research content, use the paper listening workflow to map design, measurement, and figures before scripting.

Check transformations introduced for audio

Audio production adds changes beyond summarization:

• rounding numbers;
• expanding acronyms;
• replacing tables with descriptions;
• reordering sections;
• adding analogies;
• translating or transliterating terms;
• assigning statements to different hosts.

Review each transformation. A rounded value should not cross a decision threshold. An analogy should preserve the important mechanism and boundary. A translated technical term should match domain usage. Reordering should not make a later result appear to justify an earlier decision.

The two-host script guide recommends stable roles and questions grounded in the source; those choices reduce, but do not eliminate, the need for verification.

Perform a listen-through

Text review misses audio-only errors. Listen at normal speed and mark:

• names or terms pronounced as another word;
• numbers whose pauses create ambiguity;
• speaker switches that change attribution;
• sentences clipped during stitching;
• repeated or missing lines;
• background sound that masks a qualification;
• chapter titles that promise content the section does not contain.

Follow the transcript during a second pass for important material. Update the transcript to match the final audio, not the pre-render script.

Use risk-based review depth

Personal review of a public article and a clinical protocol do not deserve the same release process.

• Low stakes: sampling plus full review of names, numbers, and quotations.
• Moderate stakes: full claim ledger and listen-through by someone familiar with the subject.
• High stakes: qualified domain review, source inspection, documented approval, and an explicit statement that audio does not replace authoritative instructions.

Some documents should not become a generated podcast at all. Exact legal, medical, safety, or operational instructions may lose too much through conversation and summarization.

Release with corrections in mind

Publish the source link, transcript, episode date, and a correction contact or process. Keep version history. If an error changes meaning, update both audio and transcript rather than silently fixing only the text.

Before release, confirm:

• all material claims have a source location;
• numbers retain unit, period, population, and baseline;
• quotations are exact and attributed;
• inference is labeled;
• uncertainty appears beside the relevant claim;
• visual and OCR-derived evidence was inspected;
• the final audio, not just the script, was reviewed;
• the transcript matches the final render;
• sensitive or restricted information is absent.

DuoCast can accelerate the conversion from source to two-host audio. It cannot turn generation into verification. The most reliable workflow keeps those jobs separate: prepare the source, generate the draft, trace the claims, listen to the render, and preserve a correction path.

Fluency is a production quality. Accuracy is an evidence relationship. Review the relationship.