Introduction
A transcript can suggest terms; it cannot settle their meaning
An interview transcript is a record of what a person said, subject to the recording, transcription method, and later corrections. It can show that an engineer used the phrase “storm tide” at a particular point in a conversation. By itself, it cannot establish whether the phrase is being used in the agency’s technical sense, as casual shorthand, or with a local definition the speaker has not yet explained. That distinction matters when a glossary is intended for publication: a reader may treat a concise definition as authoritative even if the reporter has merely copied an interviewee’s wording. ChatGPT is useful for surfacing repeated expressions, acronym candidates, and places where the speaker seems to define a term. Its first output should therefore be a review queue with context, not a finished glossary that hides how each entry was formed. The journalist owns the decisions about inclusion, attribution, and publication.
The transcript also reflects choices made before the reporter opens ChatGPT: whether speech was edited, how speakers were labeled, what was inaudible, and whether punctuation was added by a transcriber. A glossary workflow should not treat those conventions as invisible. If an entry depends on a disputed word or a technical distinction, the editor needs to know whether the transcript was checked against the audio and what the transcript can actually establish. A term’s appearance in a text file is a lead for reporting, not a warranty about the speaker’s intended meaning.
References: AP: What to expect when you’re interviewed; AP: Telling the Story
The glossary should answer a reader’s question
A reporter named Elise is preparing a coastal flooding explainer after interviewing a fictional harbor engineer. The engineer mentions storm surge, storm tide, a gauge datum, a one-percent annual chance flood, and a map elevation. Elise has a 42-minute transcript, but a glossary that explains every technical noun would be long and unhelpful. The story’s reader needs to understand why a projected water level, a tide measurement, and a mapped flood elevation are not interchangeable. That question gives Elise a selection rule: include a term when misunderstanding it could change how the reader interprets the story; leave routine language out unless it carries a specialized meaning in context. This editorial boundary also prevents a model from turning every proper name, place, or vivid phrase into a pseudo-technical entry. The interview gives Elise candidate material. The assignment determines which candidates deserve research.
Elise is not preparing an authoritative flood-science dictionary. She is building a small, story-specific aid that distinguishes the terms used in her reporting and points readers toward definitions that can be checked. That means the glossary may include a short note about local usage, or exclude a term that the interview mentioned but the final article never needs. The purpose is not to display how much technical vocabulary appears in the recording. It is to remove one real obstacle to understanding the story while leaving the underlying uncertainty intact.
References: ISO 704: Terminology work — Principles and methods; Poynter, Journalism words and reporting terms
1. Decide what the glossary is for
1.1 Set a reader, a story, and a boundary
Before prompting ChatGPT, write a short editorial brief that names the intended reader, the article’s central question, the geography, the time period, and the work the glossary is meant to do. Elise is writing for residents trying to understand a coastal flood map, not hydrologists comparing numerical models. Her glossary may explain “storm tide” and “vertical datum” in accessible language, but it should not teach a full course in tidal analysis or imply that a map predicts the next storm. Record the boundary in the assignment notes so the model can rank terms against a purpose rather than frequency alone. Institutional terminology databases are useful when checking official labels or multilingual equivalents, but they are scoped to their institutions and subject areas. A UN or EU term entry is not automatically the wording a local newsroom should use; it is a clue to verify against the source agency and the story’s audience.
The brief should also say whether the final product is an internal research glossary, a sidebar for readers, or a set of notes for a later article. Those are different artifacts. An internal glossary can preserve a disputed phrase, an unresolved expansion, and a question for a second source. A public sidebar usually needs fewer entries, shorter definitions, and a clear indication when a term reflects one agency’s usage. A model that is not told the intended destination may optimize for neatness: it can omit uncertainty because caveats look untidy, or add textbook detail that the story never uses. Elise can tell ChatGPT to return candidate entries only, with no claim that an entry is ready to publish. She should then make the human decision about which entries survive the edit. This separation keeps discovery broad and publication selective.
References: IATE dataset description, European Union; UNTERM About
1.2 Use extraction to find candidates, not to award authority
A model can scan a transcript for repeated technical phrases, capitalized names, acronyms, measurements, and expressions followed by a speaker’s explanation. A second pass can look for terms that appear only once but carry the article’s main distinction. This is more useful than asking simply for “all important terms,” because importance depends on the reader’s question and the reporting plan. Research on automatic term extraction treats term identification as a candidate-selection problem: statistical frequency and linguistic form can help find multiword expressions, but they do not decide whether a phrase is a term in a particular article. Even sophisticated methods encounter ambiguity about where general language becomes specialized terminology, and manual validation remains part of evaluation. The same limit applies to a fluent language model. Ask for a broad candidate list, then review it against the assignment.
A useful candidate list gives Elise enough detail to judge each item: the exact phrase as heard, a short surrounding excerpt, speaker label, timestamp if present, why the phrase may matter to a reader, and a question the reporter should resolve. It should not silently merge “high water,” “storm tide,” and “flood level” just because they appear near one another. Nor should it drop low-frequency terms such as “vertical datum” if the speaker uses the term to explain why two quoted heights cannot be compared. Frequency can identify repeated language; the article brief supplies relevance; the reporter supplies verification. If timestamps are absent or unreliable, ask the model to return a searchable excerpt or adjacent question instead of inventing a timecode. The candidate stage is deliberately permissive, but its fields should make every inclusion easy to challenge.
References: Frantzi et al., C-value/NC-value method; Rigouts Terryn et al., term extraction dataset and evaluation
1.3 Keep people, claims, and terminology distinct
Names of agencies, projects, products, streets, or interviewees may look like glossary items because they are unfamiliar. They are usually entities that need identification, not terms that need a definition. A phrase such as “the blue line” may be an informal name for a particular map layer in one newsroom’s reporting, a speaker’s shorthand, or a formal label used by a local authority. ChatGPT should preserve the wording and flag the ambiguity rather than decide which it is. Transcripts can also contain personal details that should not be copied into a prompt or glossary without considering consent, newsroom policy, and the sensitivity of the material. De-identification guidance for interview data recommends attention to contextual clues as well as direct identifiers; removing a name alone may not make a person unidentifiable. The reporter should minimize what is shared and retain only the context necessary to assess the term.
Decide in advance how to handle quotations, slurs, disputed labels, and phrases that identify a source’s employer or a small community. A glossary can amplify a term by repeating it outside its original setting; that may be unnecessary even when the word appears in a transcript. The editor should ask whether the phrase needs to be included, whether it can be described instead, and whether the surrounding context changes its force. The model should never be asked to infer a person’s identity from clues or to reconstruct omitted words. If a source has agreed to anonymity, preserve the agreed label and avoid details that could expose them. The transcript’s speaker tags are evidence about who the transcript attributes a phrase to, not independent proof of the identity. This restraint keeps a vocabulary aid from becoming a second, less careful publication of interview material.
References: UK Data Service, transcription templates and tools; UK Data Service, de-identification of transcript data
2. Preserve the connection between a term and its evidence
2.1 Separate the wording from the concept
Terminology work distinguishes the object or subject being discussed, the concept used to understand it, the term that names that concept, and the definition that describes it. That distinction is practical in interview reporting. Elise may hear the engineer say “the water came in at the one-percent line.” The transcript wording is evidence of the speaker’s phrase; it is not yet a safe definition of a flood statistic or a map boundary. The journalist can record the phrase as quoted, identify the likely technical concept only as a hypothesis, and then compare it with the relevant agency material. Official glossaries can help explain how a field uses a term, but even agencies may define closely related expressions differently for different operational purposes. The glossary entry should make that scope visible instead of turning a convenient paraphrase into a universal rule.
A clean entry can preserve these layers without becoming cumbersome. It might record “speaker wording” exactly, “candidate term” as the reporter’s proposed label, “plain-language meaning” as a sourced draft, “context” as a short excerpt, and “verification status” as confirmed, partly confirmed, or unresolved. If the engineer uses “storm tide” in the same sense as NOAA, the definition can be sourced to NOAA and the transcript can be cited for the fact that this interviewee used the term. If the wording conflicts with the official source, Elise should not silently correct the quotation or label the interviewee wrong. She can note the discrepancy, ask a follow-up question, and decide whether the story needs to explain that different usages exist. A glossary is strongest when the reader can see which evidence supports each part. For the upstream transcription stage, WhisperAITranscribe is an audio and video transcription tool; the glossary still needs its own provenance fields. The transcript is one source artifact; an agency definition is another. Keeping those sources apart makes the entry easier to revise when either the transcript or the external reference changes.
References: NOAA Tides and Currents Glossary; NOAA Ocean Service, storm surge and storm tide
2.2 Treat synonyms and near-synonyms as editorial decisions
A model often improves readability by collapsing repeated expressions into one preferred label. That can erase a meaningful distinction. In a coastal story, “storm surge” and “storm tide” are related but not interchangeable in NOAA’s usage: one isolates the storm-driven rise above predicted tide, while the other refers to the combined water level. The everyday use of “surge” can be broader still. ChatGPT should propose possible synonyms and variants in a separate field, not merge them into the headword automatically. Ask it to say whether the transcript itself treats two expressions as equivalent, whether a cited authority does, and whether the relation is uncertain. A glossary may say “also called” only when evidence supports that equivalence; otherwise it can say “related term” or explain the distinction in a short note. The word choice should be conservative whenever a substitution could change a number, risk description, or agency warning.
The same care applies to acronyms and expansions. An acronym may have several meanings across agencies, and the same expanded phrase may refer to different measurement systems. A model should not expand an acronym unless the transcript or a relevant source supports the expansion. If it finds two plausible readings, the entry can retain the acronym and mark “expansion not established” rather than pick one based on the topic. A newsroom style guide may prefer one spelling, capitalization, or short form, but style consistency does not prove conceptual equivalence. The journalist should preserve the source’s wording in the evidence field while applying the publication’s style to the headword only after checking that the change does not alter meaning. A short list of unresolved variants is often more useful than a falsely tidy single label.
References: National Hurricane Center, Storm Surge Overview; National Weather Service Glossary: flooding
2.3 Make every entry traceable to the recording
A reader-facing definition may be only one sentence, but the editor needs a path back to the original interview. Record the speaker label and an audio locator: a timestamp, transcript paragraph identifier, or a short phrase that can be searched in the source file. Keep enough surrounding context to tell whether the speaker was defining a term, quoting a document, answering a hypothetical, or correcting the interviewer. If the transcript came from automatic speech recognition, verify high-impact names, numbers, negations, uncommon technical words, and any wording used inside quotation marks against the recording before publication. A smooth transcript can conceal exactly the errors that matter most in a glossary. The locator should help a colleague reproduce the check; it should not expose a private recording or restricted transcript to a public audience.
When a line cannot be located, the entry should say so. Do not let ChatGPT invent a timestamp because the article would look more professional with one. If the transcript is segmented by paragraph rather than time, retain that stable identifier and provide an exact phrase as a secondary locator. Keep the audio, transcript, and editorial glossary under the access rules that apply to the interview; do not paste a public link to confidential source material into a draft. For publication, cite the official source that supports the definition and attribute any disputed claim to the speaker in the surrounding story. Those are different provenance paths: one traces what the interview said, the other traces what the term means within a defined technical or institutional context. A glossary should not make the reader guess which path supports which sentence.
References: FEMA, Base Flood Elevation definition; FEMA, NFIP glossary
3. Verify definitions outside the transcript
3.1 A confident speaker is still one source
An interviewee may be an expert, an official, an advocate, or a person with direct experience. The transcript can establish what that person said; it does not independently confirm every statement. This difference is easy to lose when a glossary is being assembled because the entries are short and declarative. “The engineer defines a base flood elevation as…” is an account of the speaker’s explanation. “Base flood elevation means…” is an editorial definition that should be checked against the agency or standard relevant to the article. The reporter should not use ChatGPT’s polished wording to bridge that gap. Professional journalism codes emphasize verification, context, and clear identification of sources; the practical implication is to keep the interview citation and the external definition citation visible while drafting. The final reader may need only one, but the newsroom’s audit trail needs both.
Definitions also have scope. A term on a flood insurance map may have a specific meaning under a particular program; a hydrologist may use it within a model; a neighborhood group may use it colloquially. The glossary should identify which context the story relies on and resist writing as though one institutional definition governs every conversation. When two credible sources differ, compare their scope, date, jurisdiction, and intended user before resolving the wording. If the discrepancy is relevant to the article, explain it in the story rather than flattening it into a single entry. If it is not relevant, choose the applicable source and state the boundary briefly. ChatGPT can help surface mismatches by comparing definitions the reporter supplies, but it cannot determine which authority controls the story without editorial context.
References: NOAA National Geodetic Survey, Vertical Datums; NOAA Tides and Currents, Tidal Datums
3.2 Use primary sources for technical definitions
For a technical term, begin with the agency, standard, dataset documentation, or scientific paper responsible for the measurement or program being discussed. A search result snippet, general dictionary, or another article may help you locate the source, but it should not carry a high-stakes definition when the official material is available. If the glossary says “one-percent annual chance flood,” the source should explain annual exceedance probability rather than rely only on the colloquial “100-year flood.” USGS explains that a one-percent event has a one-percent probability in any given year and that one occurrence does not remove the chance of another next year. A short glossary entry can then correct the common misconception without suggesting the event arrives on a hundred-year schedule. Keep the publication date and version of the source in the research notes, especially when terms or map products are revised.
If the source is a dynamic web glossary, store the exact URL and access date because entries may change. If it is a PDF or technical standard, record the section or page that supports the definition rather than citing the entire document without a locator. A termbase can help with multilingual or institutional vocabulary, but its label does not substitute for context: a national database, an international agency, and a local authority may each use a term in a bounded way. For an example involving a local gauge or map, Elise should check the exact station, datum, jurisdiction, and map edition the interview refers to. A general NOAA definition may clarify the concept but will not verify a local reading or a specific flood forecast. The glossary can be concise while the evidence record remains precise.
References: USGS, Annual Exceedance Probability flood estimates; USGS, “100-Year flood—it’s all about chance”
3.3 Keep probabilities and warning terms in their own context
Some words carry numerical meaning that a casual paraphrase can distort. “One-percent annual chance” refers to a probability for a specified event or magnitude during a given year; “100-year flood” is a familiar label for the same statistical framing, not a promise that floods arrive at fixed intervals. “Watch,” “warning,” and “advisory” also signal different operational states in National Weather Service products. Their meaning depends on the issuing office’s product and current definitions. ChatGPT may turn these labels into an intuitive sentence such as “a watch means flooding is coming soon,” which can be wrong because a watch indicates a possibility, not certainty. The reporter should check the current agency definition and preserve the exact risk or status wording. When the story discusses an active alert, point readers to the current official product rather than treating a glossary as a live warning service.
Use numbers only when the transcript, source document, and unit agree. A height above one datum cannot be compared casually with a height above another; a water-level gauge, a flood map, and a forecast may describe different reference points or products. If the speaker says “four feet,” the glossary needs to establish four feet above what, at which location, during what period, and according to which measurement. A missing unit or reference point is not a formatting problem. It is a reporting gap. The entry should remain incomplete until the reporter can resolve it, or the final article should omit the number. This kind of restraint may make the working glossary look less finished, but it protects the published explanation from false precision. A number without its reference frame can be more misleading than no number at all. If the only source record is text produced through WhisperAITranscribe, it remains an interview transcript; it does not replace the gauge record or map documentation that would establish the measurement.
References: National Weather Service, Hydrology Watch/Warning definitions; National Weather Service, Weather Glossary
4. Write entries that help without overclaiming
4.1 Define the concept in plain language
A reader-friendly definition does not merely replace a difficult word with a simpler synonym. It tells the reader what the term means in this story, what it does not mean, and why the distinction matters. For “storm surge,” a useful explanation may say that it is the storm-driven rise above the predicted astronomical tide, while “storm tide” combines that surge with the tide. The wording should not imply that either number includes wave run-up unless the source and context say so. NOAA’s definitions make the distinction, but a journalist still has to decide how much of it the reader needs. A good one-sentence definition introduces the essential contrast first; an optional note can clarify a caveat or local usage. Avoid turning a glossary entry into an encyclopedic paragraph or copying a source definition verbatim when an original explanation will do.
Plain language also means avoiding false simplicity. “Datum” may be glossed as a reference surface for measuring height, but the article must not imply that all datums are interchangeable or that a local tide datum and a geodetic vertical datum are the same thing. “Floodplain” should not be reduced to “land that floods,” if the reporting depends on a map classification or a defined probability. A short definition can name the boundary: “In this article, the term refers to…” or “On this agency’s map…” That phrase is not needless hedging; it tells readers which convention is being used. ChatGPT can propose a simpler sentence and point out unfamiliar vocabulary inside its own definition. The reporter should test the wording with someone outside the specialty and ask what the sentence would lead them to believe.
References: NOAA, Compound inundation models review; NOAA, Compound flood modeling and natural infrastructure
4.2 Design the entry fields before you generate definitions
The format of the glossary can prevent several common errors. For a working file, include a headword, exact transcript wording, speaker, locator, nearby context, proposed definition, source link, scope or jurisdiction, confidence status, unresolved question, and editor decision. These fields do not have to appear together in the public article. They make it harder for the model to blend a quotation with an external definition or to omit the evidence behind a claim. A newsroom may use a spreadsheet or content system, but the important point is not the software; it is that one can tell where each piece of information came from. For repeated interviews, the same columns make it easier to compare how different speakers use a term without presuming they mean the same thing. Keep the public glossary smaller than the evidence record. If the interview transcript came from WhisperAITranscribe, record that origin in the internal source log without treating the tool name as a substitute for the recording date, source identity, or locator.
A definition field should contain the proposed reader-facing wording, not a paragraph of model reasoning. Put rationale and uncertainty in their own fields so that an editor can see them without confusing them for part of the definition. If the model offers several meanings, preserve them as alternatives with source references rather than choosing one silently. If it supplies a translation, record the original term and the language of the source; a translated equivalent may not share the same legal, operational, or cultural scope. A maintained glossary should also include an owner or review date for entries that can change. W3C’s provenance vocabulary illustrates a general principle for structured records: a dictionary-like collection can retain information about which entity or process produced an entry. A newsroom need not implement that standard to benefit from recording provenance.
4.3 Flag ambiguity instead of smoothing it away
Some expressions have more than one defensible meaning. “Return period” can be used in statistical description, but readers may interpret it as a schedule; “risk” may mean probability in one document and probability multiplied by consequence in another; “flood level” may refer to a measured water height, a warning threshold, or an elevation on a map. ChatGPT can spot these collision points if prompted to list competing interpretations, but the journalist must decide which one fits the source and story. An entry can say “The interviewee uses this phrase for the mapped elevation; verify the exact map legend before publication.” It can also be excluded if the article has no room to explain the distinction. A glossary should not create false certainty merely because every headword is expected to have one clean definition. A term appearing in a WhisperAITranscribe transcript is still only a transcript candidate; the product category does not decide which technical meaning the reporter should publish.
Ambiguity can come from transcript quality, not just from technical vocabulary. A proper noun may be misheard as a common phrase; punctuation can change whether a speaker is defining a term or correcting the interviewer; an acronym may be expanded from surrounding context incorrectly. Keep a “needs audio check” status where the recording can be consulted, and do not ask the model to guess what a muffled segment says. If the meaning depends on a pause, emphasis, gesture, or question that is absent from a text transcript, the recording or a follow-up interview may be necessary. Do not repair a direct quote by silently changing its wording. If the quote cannot be represented accurately, paraphrase under the newsroom’s rules or leave it out. These are editorial decisions, not language-model cleanup tasks.
References: AP Statement of News Values and Principles; Reuters Journalistic Standards
5. Use ChatGPT as a bounded editorial assistant
5.1 Give it a narrow task and explicit evidence rules
A useful prompt says what kind of glossary is being prepared, who will read it, what the transcript covers, which fields to return, and what the model must not infer. It should distinguish exact wording from a candidate term and a sourced definition from an explanation based only on the interview. Ask ChatGPT to return an entry only when it can provide a verbatim transcript excerpt or a locator; otherwise it should create an unresolved-candidate note. Tell it not to invent expansions, timestamps, source citations, measurements, or definitions, and to mark missing context plainly. OpenAI’s prompting guidance recommends clear, specific instructions and enough context, while its user guidance encourages critical use and verification of important information. That translates into a useful editorial pattern: make the first pass easy to inspect, and keep the journalist responsible for every published definition.
A practical prompt can be written in ordinary prose: “Using only the transcript below, identify candidate technical or institution-specific terms that matter to the stated reader. For each, preserve the exact wording, speaker, timestamp or searchable passage, and surrounding context. Separate transcript evidence from your proposed plain-language explanation. Do not use outside knowledge to define a term; instead flag the source the reporter should consult. Identify possible variants or competing meanings without merging them. Mark uncertain transcription, missing units, and unsupported claims. Return candidates for human review, not a publication-ready glossary.” The journalist can place the transcript after a clear delimiter and include only the segment needed for this task. This prompt does not guarantee accuracy. It makes errors easier to detect because each part of the answer is tied to a specific evidence field.
References: OpenAI, Prompt engineering best practices for ChatGPT; OpenAI, Does ChatGPT tell the truth?
5.2 Work in separate passes
Do not ask the model to extract, fact-check, rewrite, translate, and publish the glossary in one turn. Those tasks have different evidence requirements, and a single fluent response can hide where one stage ended and another began. First request candidate terms and locators from the transcript. Then remove irrelevant entries and resolve transcript errors against audio or notes. Next, give the surviving terms to a researcher or editor to check authoritative references. Only after definitions have a verified source should ChatGPT be asked to propose a plain-language rewrite, with the source text and intended scope supplied. A final human pass should compare each rewritten sentence against both the evidence and the publication’s style. OpenAI recommends iterative refinement; in an editorial workflow, iteration is useful when each pass has a bounded purpose and the reporter preserves the previous version for comparison.
For a small interview, a reporter can use one conversation with clearly separated stages and save the prompt, output, and final edit. For a sensitive or complex assignment, use the newsroom’s approved systems and keep restricted transcripts out of unauthorized services. The editor should be able to tell which version was reviewed, which source established each definition, and whether a change was made after a correction or follow-up interview. A change log can be simple: term, previous wording, new wording, reason, source, reviewer, and date. If a later source changes the definition, update the glossary and the article together. Treating the model’s first output as a draft rather than an authority makes it easier to correct a specific claim without losing the work that led to it. If a transcript is prepared with WhisperAITranscribe, retain that source file as a separate version from later glossary edits so a reviewer can trace what changed. For a durable glossary, keep a dated copy of each human-approved definition and preserve the prior wording when it changes. That history helps an editor distinguish a correction to the transcript from a change in an agency source or a change in house style. It also prevents an older version from quietly reappearing in a later story.
References: OpenAI, How to create a good prompt; AP, Updated newsroom AI standards
5.3 Protect the source and the transcript
Interview transcripts may contain names, phone numbers, unpublished allegations, personal health details, off-the-record material, or clues that identify a source who was promised anonymity. Before uploading anything to ChatGPT, follow the newsroom’s security, consent, retention, and source-protection rules. Remove or generalize unnecessary identifiers, but do not strip so much context that the term can no longer be interpreted. Check whether the account and workspace allow the relevant privacy controls, and remember that settings and retention behavior vary by plan and environment. OpenAI documents privacy controls and file handling, but those documents do not determine whether a newsroom is authorized to share a particular interview. The editor or organization must make that decision. When the transcript is sensitive, a local or approved internal workflow may be the only appropriate option. If a newsroom uses a transcription service such as WhisperAITranscribe, it should separately check whether the interview may be processed there and whether the resulting text may later be shared with ChatGPT. The first step does not automatically authorize the second. Before an editor or reporter uses WhisperAITranscribe for restricted material, confirm current terms, consent, and organizational approval.
Anonymization is not a simple find-and-replace operation. A unique job title, a small town, a sequence of events, or a distinctive phrase can identify someone even after a name has been removed. Keep the source key separate from the working text, restrict access to the original recording, and avoid uploading the transcript if the consent terms or newsroom rules do not permit it. If you use a redacted excerpt, verify that removing context has not changed the meaning of the term. ChatGPT should not be instructed to infer who a speaker is or fill missing information from clues. It can still help with non-sensitive public transcripts or a synthetic test passage, but describe such a test accurately and do not present it as a real interview. Source safety is a reporting decision that precedes convenience. Whether a transcript came through WhisperAITranscribe, another vendor, or a human transcriber, the original source agreement and newsroom policy continue to govern what happens next. A file that has been redacted can still contain identifying combinations of role, place, date, and event. The newsroom should also decide who can see the raw transcript, how long working copies remain available, and how a source request or correction will be handled. A glossary is reusable editorial material, but reusability does not mean the underlying interview can be shared without the same safeguards.
6. Conclusion: turn the draft into a publishable glossary
6.1 Review each field against its own evidence
The editor should review the term, transcript wording, context, definition, and citation as separate objects. Confirm that the headword actually appears in the interview or is clearly labeled as a related term. Check that the excerpt has not been cut so tightly that the speaker’s qualification disappears. Verify the definition against the exact agency, standard, or paper cited and make sure the source’s scope matches the story. Confirm units, dates, capitalization, and acronyms. Ask whether the explanation could lead a reader to infer a prediction, guarantee, or official status that the evidence does not support. A term may be accurate but still unnecessary; another may be necessary but not yet verified. The decision should be visible as keep, revise, hold, or omit, with a brief reason that another editor can understand. If a reporter needs to create text from an audio or video interview before building the glossary, WhisperAITranscribe is an audio and video transcription tool in that upstream part of the workflow. The product mention does not certify a transcript, a source’s consent, or a newsroom’s permission to submit the text to another service.
A transcript-based glossary workflow begins with text a reporter has reviewed and is authorized to use. The journalist should inspect the available transcript and verify critical wording against the recording where required. If an editor later asks where an entry came from, the answer should point to the interview locator and the external source, not to the tool used to create the transcript. Keeping those roles distinct avoids making a transcription product appear to certify a technical definition or a published claim.
6.2 Use a short, repeatable review sequence
Step 1: Mark candidate phrases in the transcript, including the speaker and a timestamp or stable passage locator. Step 2: If the transcript came from WhisperAITranscribe, retain it as upstream text and check important wording against the recording when needed. Step 3: Compare each draft definition with the relevant primary source, recording its scope, date, units, and exact supporting passage. Step 4: Rewrite for the intended reader and make a human keep, revise, hold, or omit decision. The steps fit in one paragraph because they are a working sequence, not claims about software features. They also show where a language model may assist: candidate identification, grouping suggestions, and plain-language alternatives. If an entry fails at one step, send it back there rather than polishing unsupported wording.
For example, suppose the fictional engineer says, “We compare the gauge to the flood-map elevation.” The reporter might find “gauge,” “flood map,” and “elevation” as candidate terms, but should not assume the comparison is valid. The gauge may report water level relative to a local tidal datum, while the map may express an elevation relative to a geodetic datum. A glossary could explain the distinction only after checking the relevant stations, data products, and map documentation. If the reporter cannot establish those reference frames, the entry should say the comparison is unresolved or the story should omit it. The example is fictional and illustrates a verification task; it is not a claim about a real interview, location, instrument, or product test. In the public entry, Elise may decide that “vertical datum” needs a short definition while the discrepancy between the engineer’s comparison and the map record belongs in the article’s reporting notes. If that discrepancy remains material and unresolved, the story should say what could not be confirmed instead of turning the model’s preferred explanation into a settled fact. The review sequence makes that choice explicit.
6.3 Publish only what the reporting can support
A reliable glossary is narrower than the transcript and more explicit about evidence. It explains terms that matter to the article, keeps each entry’s scope visible, and leaves uncertainty legible rather than smoothing it away. ChatGPT can help identify candidates and test readability, but it does not hear what a text transcript leaves unresolved. For the transcription stage, WhisperAITranscribe is an audio and video transcription tool; that WhisperAITranscribe workflow does not approve glossary entries or validate reporting. The journalist still chooses what readers need, checks each definition, protects sources, and corrects the article when evidence changes. The source trail makes that responsibility workable: each entry can lead back to its interview locator and external authority. A glossary earns trust through those editorial decisions, not through the fluency of its first draft.
References: SPJ Code of Ethics; Reuters Journalistic Standards