For people who work with English-language materials, research international topics, and produce knowledge-based content. Availability, free quotas, and features vary by account, region, and product strategy — always confirm against the actual product pages and terms of service.
Introduction: What Global Tools Really Teach You Is Not Just "Writing Better"
Many people try global AI tools to compare which model is smarter. For content creators, the more valuable comparison is not feeding every model the same riddle, but observing how each participates in a real workflow: which one is better at brainstorming angles, which one is better at reading materials, which one preserves sources more reliably, and which one can explain complex material to a general audience.
The free route has limited resources, which actually pushes us toward a more robust method: collect materials first, then put AI to work; demand sources before demanding polish; produce one small, credible piece before chasing more elaborate visuals.
This article walks through a complete case study: turning three public English-language sources into a 90-second explainer video for your own audience, titled "Why You Should Check the Sources Behind AI Assistant Answers." The final deliverables include a source list, an evidence table, the localized script, a bilingual glossary, a shot table, and a pre-publish checklist.
graph LR
A[Collect public sources] --> B[Import into Notebook]
B --> C[Extract evidence and disagreements]
C --> D[Cross-checking]
D --> E[Localization]
E --> F[Visual explanation]
F --> G[Fact review]
G --> H[Publish]1. The Minimal Free Tool Stack
| Tool | Role in this article | Key usage points |
|---|---|---|
| ChatGPT Free | Brainstorm angles, explain concepts, rewrite voiceover copy | Be explicit about requirements, constraints, and output format |
| Gemini free plan | Assist with search, compare information, generate explanation frameworks | Open and inspect the cited sources; never rely on summaries alone |
| NotebookLM / Gemini Notebook | Ask questions grounded in uploaded or added sources; generate organized output with citations | Build the habit of "answers from within the materials" |
| Built-in image features available on your account | Produce concept diagrams, cover drafts, and visual metaphors | Features and quotas depend on your actual account |
| Local editing tools | Subtitles, assembly, voiceover, and export | Using global AI does not mean you must use international editing software |
OpenAI currently offers a free ChatGPT plan; Google's official Gemini page lists a free plan with writing, planning, and image capabilities; NotebookLM is positioned as a tool for researching and organizing the sources you provide, with answers that trace back to those sources. Free plans typically come with usage caps, differences in model access, or feature restrictions, so never build your entire production plan on a fixed quota.
The core of this stack is not "multiple models voting against each other," but assigning different jobs to different tools: the Notebook owns the evidence scope, Gemini handles discovery and comparison, and ChatGPT rewrites confirmed content into language that fits your audience.
2. First, Define "What Is Allowed into the Script"
When researching international topics, the most common problem is not mistranslation but mixing information of different tiers. Sort your materials into four categories:
- Primary sources: product announcements, official documentation, research papers, regulatory filings, raw data.
- Reliable interpretations: readings of primary sources by professional institutions, universities, and mainstream media.
- Opinion material: commentary, interviews, personal experience, and industry judgments.
- Lead material: social posts, reposted summaries, and claims without a complete provenance.
Primary sources go into the fact table first; reliable interpretations aid understanding; opinion material can only be quoted as opinion; lead material is used to trace primary sources and must never become a key conclusion on its own.
Build a source table for this case study:
| ID | Source type | Purpose | Usable as factual evidence? |
|---|---|---|---|
| S1 | Official product description | Understand the AI assistant's capabilities and limits | Yes, but only for that product itself |
| S2 | Official research-tool description | Understand how source-grounded answering works | Yes |
| S3 | University or research-institution guide | Learn how to verify AI output | Yes, as methodological advice |
| S4 | Personal commentary | Understand user questions | Cannot support facts on its own |
Do not let the model "find three links that look like sources" and call it done. You need to open each page and confirm that the title, publisher, date, body, and conclusions are consistent.
3. Use a Notebook to Build a Controlled Material Space
The value of NotebookLM is not browsing the entire internet for you, but putting a chosen set of materials into one space so that answers stay anchored to those materials as much as possible. Some official pages currently also use the name Gemini Notebook; the interface and naming may change, but the method of "asking questions around your sources" stays the same.
After adding the three sources to the same Notebook, do not ask it to write the script yet. First ask about the coverage of the materials:
Answer only from the sources in this Notebook.
1. What problem does each source address?
2. Which conclusions are supported by at least two sources?
3. Which claims appear in only one source?
4. Which important questions are not answered by these sources?
5. Produce a three-column table: "confirmed / needs verification / opinion only".Then ask it to generate an evidence table:
Around the topic "Why you should check the sources behind AI assistant answers," extract at most 8 pieces of evidence usable in a 90-second video.
For each item include:
- evidence_id
- A concise conclusion
- The source_id that supports it
- The location or quotation hint within the source
- The conditions that must be preserved when using this evidence
- Whether the sources disagree on it
Do not pad the list to 8 items by adding information from outside the materials.The evidence table exists to establish a "fact budget" before the script. A 90-second video does not need twenty conclusions; three to five are usually enough.
4. Cross-Checking: A Citation from the Model Does Not Mean the Citation Supports the Conclusion
Source-driven tools reduce the risk of fabrication, but they do not eliminate manual verification. Check at least the following:
- Does the cited passage actually support the sentence it follows?
- Where the original says "may," has the script turned it into "will"?
- Where the original covers a specific scenario, has the script generalized it to all scenarios?
- Which year is the data from, and does it still apply to the current discussion?
- Is the source introducing its own product, or making an independent comparison?
- When several sources rely on the same primary material, have they been miscounted as multiple independent pieces of evidence?
You can ask Gemini or ChatGPT to play the opposing side, but you must provide the evidence table along with it:
Below are my evidence table and draft conclusions. Review them as an adversarial editor. Do not add new facts.
Tasks:
1. Find places where a conclusion exceeds the scope of the evidence
2. Find places where conditions have been omitted
3. Find possible translation ambiguities
4. Find claims that appear multi-sourced but actually trace back to a single primary source
5. Rewrite over-confident statements so they match the strength of the evidenceIf the two tools disagree, do not settle it by "majority vote." Go back to the original text and determine whether the dispute is about facts, terminology, or angle of interpretation.
5. From Foreign-Language Material to Content for Your Audience Is Not Line-by-Line Translation
Localizing international material involves at least four layers of work: terminology translation, background supplementation, example replacement, and tone adjustment.
For example, when adapting English materials into Chinese, the phrase grounded in sources could be rendered literally as "基于来源的接地回答" ("grounded answering based on sources"), which sounds unnatural in Chinese. In context, a better expression is "answers built around the provided materials, with a path back to the source for verification."
Build the glossary first. Here is the English-to-Chinese glossary from this case study, kept as an illustrative example:
| English term | Usage in this article | Not recommended | Notes |
|---|---|---|---|
| source-grounded | 围绕来源回答 (answering around the sources) | 接地式回答 ("grounded-style answer," a literal calque) | Emphasizes that answers can be traced back to the materials |
| hallucination | 无依据生成 / 错误编造 (unsupported generation / fabrication) | 幻觉 ("hallucination" left unexplained) | Explain it on first occurrence |
| citation | 引用标记 / 出处 (citation marker / provenance) | 证明 ("proof") | A citation does not mean the conclusion has been proven |
| context window | 可同时处理的上下文范围 (the span of context processed at once) | 记忆容量 ("memory capacity") | These are different concepts |
When asking ChatGPT to localize, explicitly forbid line-by-line literal translation:
Turn the following English material into an explainer for general readers in the target language. Do not translate line by line.
Requirements:
- Preserve the strength of each fact and its qualifying conditions
- Explain each technical term in one plain-language sentence on first occurrence
- Replace examples that only make sense in an English-speaking context with neutral scenarios the target audience can understand
- Do not add speculation about local market availability, laws, or product access
- Keep the evidence_id on every factual sentence
- Output "localized explainer + glossary + ambiguities requiring human confirmation"6. Settle on One Claim for the 90-Second Video
This case has plenty of evidence, but the video can carry only one claim:
A fluent AI answer does not mean the information is reliable; the point of checking sources is to confirm whether the conclusion is actually supported by the materials.
Design the structure around that claim:
00–08 s: A common misconception
08–20 s: The core conclusion
20–55 s: Three source checks
55–75 s: One counterexample
75–85 s: How to do it with free tools
85–90 s: A call to actionScript Prompt
You are an editor for knowledge-based short videos. Write a 90-second voiceover script based on the evidence table.
Audience: everyday users who frequently use AI for search and writing
Core claim: a fluent AI answer does not mean reliable information; you must confirm the citations support the conclusions
Rules:
- Use only information mapped to an evidence_id
- Keep internal markers like [E1] after key factual sentences
- Keep each sentence short and easy to speak aloud
- Explain three checking actions: look at the publisher, open the original, look at the qualifying conditions
- Include one example of "a citation that exists but does not support the conclusion"
- Do not disparage or glorify any tool
- End by asking viewers to verify one real answer themselves
- Output the voiceover script first, then the fact-mapping tableSample Finished Script
The AI gave a fluent answer and attached a few citations — is it safe to use? Not necessarily. Citations are not decoration, and they are not automatic proof that the conclusion is correct. They only tell you where to go back and check. [E1]
>
I usually look at three things. First, who published it. A product announcement is fine for describing that product's own features, but it is not an independent comparison. [E2]
>
Second, open the original and confirm the cited passage really supports the sentence before it. Sometimes the original only says "may improve," while the answer says "already solved." [E3]
>
Third, look at the conditions. Once the time frame, sample, and scenario in a study get stripped away, the conclusion can be inflated. [E4]
>
Free tools can work this way too: put a few reliable sources into one Notebook, ask it to organize only from those sources, then use another model to check for overgeneralization. The final step is still opening the original yourself.
>
Today, pick one AI answer that comes with citations and spot-check a single sentence. You will quickly see that what matters is not "whether there is a citation," but "whether the citation supports it."
Remove internal markers like [E1] from the published version, but keep the mapping table in your project files.
7. Visual Design: Turn the Abstract Idea of "Sources" into Visible Actions
This topic does not call for flashy character art. The most effective visuals are interface demonstrations, side-by-side comparisons, and highlights.
| Shot | Voiceover | Visual |
|---|---|---|
| 1 | A fluent answer plus a few citations — can you trust it outright? | A simulated answer with citation icons appearing |
| 2 | A citation is not automatic proof | A "Citation ≠ Proof" info card |
| 3 | Look at the publisher | Three card types: official announcement, research institution, personal post |
| 4 | Open the original | Cursor clicks the source; the original passage is highlighted |
| 5 | Look at the qualifying conditions | "may / specific sample / certain period" circled one by one |
| 6 | A failure example | "may improve" on the left, "already solved" on the right |
| 7 | The Notebook workflow | Three sources flowing into one Notebook |
| 8 | The final action | Spot-checking one sentence of an answer |
Prompt for generating a concept diagram:
Minimal editorial design, one AI answer connected to three source cards, the source cards visually labeled by category as "product description, research material, personal opinion," no specific brand marks, no readable text, white background, deep blue and light gray, clear information hierarchy, 16:9Prompt for the "citation is not proof" visual metaphor:
A bridge connecting "conclusion" and "source," with a magnifying-glass checkpoint in the middle of the bridge, minimal vector infographic, professional, restrained, no people, no brands, leave space for a title area, vertical 9:16If your free account has no image-generation quota available for now, use shapes, arrows, screenshots, and caption cards instead. The visual goal of knowledge content is to aid understanding, not to prove how many tools you used.
8. Bilingual Prompts: When Does Using English Actually Help?
There is no need to translate every prompt into English. Whether a model understands a task depends more on context, examples, and constraints. Consider adding English in these situations:
- You need to keep original terms stable and prevent translated names from drifting.
- You are processing English originals and want line-by-line correspondence in the output.
- You are generating an international visual style and need common photography or design vocabulary.
- The same content needs both an English and a local-language version.
A reliable bilingual task template:
Task / 任务：
Create a Chinese explainer based only on the provided English sources.
只根据所提供的英文来源，制作中文解释内容。
Audience / 受众：
General Chinese-speaking users with no technical background.
没有技术背景的中文普通用户。
Constraints / 限制：
- Preserve uncertainty and conditions.
- Keep technical terms in English in parentheses at first mention.
- Do not add region-specific claims.
- Map every factual sentence to a source ID.
- 保留不确定性与限定条件。
- 专业术语首次出现时保留英文括注。
- 不添加材料外的地区性判断。
- 每个事实句关联来源编号。9. When Free Quotas Are Limited, What Order Should Work Happen In?
Put the cheapest, easiest-to-verify tasks first, and the most expensive generation last.
Source screening
↓
Evidence table
↓
Structure and script
↓
Human confirmation
↓
Cover drafts
↓
A few key visuals
↓
EditingDo not generate images for ten unconfirmed topic ideas. Lock the script first, then produce one or two cover candidates and the essential shots. Free tools are best suited to "fewer, better" pieces of content, not to simulating a bulk content factory.
10. Pre-Publish Source Review Checklist
□ Does every key conclusion have a source_id?
□ Does each citation directly support the conclusion, rather than being merely topic-related?
□ Are qualifiers like "may," "under specific conditions," and "as of a certain date" preserved?
□ Are the source dates appropriate for the current content?
□ Has a product's self-description been mistaken for an independent conclusion?
□ Do multiple articles cite the same primary source?
□ Has the translation changed the strength of any fact?
□ Are simulated interfaces in the visuals clearly marked as illustrative?
□ Does the publishing copy state the time range of the materials?For fast-changing products, policies, prices, and personal information, re-check the latest official pages before publishing. Do not assume that what was verified yesterday still holds today.
11. Twelve Prompts You Can Use Directly
1. Source Quality Grading
Classify the sources as "primary sources, reliable interpretations, opinion material, lead material," and state what each source can and cannot support.2. Shared Conclusions
Find conclusions supported by at least two independent sources. If multiple materials cite the same primary source, count them as one piece of evidence.3. Disagreement Map
List where the sources genuinely disagree, distinguishing "factual disagreement, terminology difference, different time frame, difference of opinion."4. Condition Recovery
Check whether the summary has dropped the original's time frame, sample, region, probability, or applicable scenarios. Restore missing conditions, but do not add new facts.5. Citation Support Check
Check each conclusion against its citation, sentence by sentence. Output supported, partially_supported, or unsupported, with a minimal explanation for each.6. Bilingual Glossary
Extract 15 key terms. For each, give the original English term, the recommended translation, a plain-language explanation for first occurrence, and translations to avoid.7. Localization of Examples
Find examples in the original that the target audience may not be familiar with. Keep the conclusions unchanged and replace the examples with neutral, understandable scenarios that introduce no new facts.8. The 90-Second Cut
There is plenty of evidence, but the video is only 90 seconds. Keep one claim, three pieces of evidence, and one limitation; move everything else to further reading.9. Adversarial Review
From a skeptic's perspective, point out the five sentences in the draft most likely to be challenged. Raise questions based only on the existing materials.10. Visual Explanation
Convert each abstract concept into an action that can be filmed, screen-recorded, or drawn as an infographic. Do not use purely decorative visuals.11. Time-Sensitivity Check
Flag facts in the draft that may change within 30 days, 90 days, or one year, and state which official page should be checked before publishing.12. Further Reading Card
Generate a further-reading card from the source table: one sentence of purpose per source, without inflating authority or adding commentary beyond the materials.12. Common Misconceptions
1. Global models are always more accurate
No general-purpose model stays accurate across all topics. Source quality, prompts, task type, and human verification matter more.
2. Having citations means everything is fine
A citation may be merely topic-related, or it may not support the specific conclusion at all. You must open the original.
3. English prompts are always better
For creative tasks in your own language, clear constraints in that language often matter more. English is useful for preserving terminology, aligning with the original text, or describing a specific visual language.
4. If multiple models agree, it must be true
Different models may have reused the same public information, or may make the same class of errors. Model agreement is only a lead, not a substitute for primary sources.
5. Free plans can't produce professional content
Free plans can support high-quality research and scripts, but they require stricter quota planning, fewer wasted generations, and acceptance that some features are restricted.
6. Translation equals localization
Line-by-line translation may preserve the words while losing the context. You need to handle terminology, background, examples, and tone.
7. The more international the visuals, the better
Visuals should serve audience understanding. A source-verification topic is best served by comparisons, annotations, and interface demonstrations — not irrelevant futuristic cityscapes.
8. One source can carry an entire video
A single source may have a stance and limited coverage. Seek independent support for core conclusions, and clearly flag what is only one party's claim.
13. FAQ
1. Do I have to use all three tools?
No. The minimal workflow is a Notebook for the materials plus one chat model for the writing. A second model mainly adds adversarial review and alternative phrasings.
2. Can I quote a Notebook answer directly?
Go back to the primary source to confirm. The Notebook's organized output is navigation, not a replacement for the original.
3. How do I handle paywalled material?
Only use content you obtained lawfully and have the right to process. Do not ask the model to fill in text you cannot see, and do not infer conclusions from headlines.
4. Can an English-language paper go straight into a video?
Yes, but state the research subject, time frame, sample, and limitations. A single study is not a universal fact.
5. How do I keep translated terminology consistent?
Build the glossary first and attach it to every subsequent prompt. Use the same translation across the script, subtitles, and cover.
6. What if the free image features aren't enough?
Prefer infographics, screen recordings, openly licensed assets, and simple shape animations you make yourself. Not every shot needs an AI image.
7. Can I produce the same piece in two languages at once?
Yes, but do not write one version first and back-translate it line by line. Work from the same evidence table and write separately for each audience.
8. Do I need to show all sources in the video?
Show the one to three most critical sources in the body, with material dates; the full source list can go in an article, the description, or companion materials, following each platform's rules.
9. How do I handle source updates?
Record the verification date in your project. For fast-changing information, reopen the official pages before publishing and update the script when necessary.
10. What kinds of content should not rely on free AI alone?
High-stakes medical, legal, or financial decisions, serious allegations, privacy-sensitive investigations, and anything requiring professional credentials should never rely solely on general-purpose AI tools.
Closing: The Habit Worth Keeping Is "Sources First"
The value of free global tools is not making a passage sound more sophisticated, but helping you work across more languages and materials while keeping a path back to the original sources.
Once you get used to building the source table first, then the evidence table, and only then writing the script, your content naturally becomes more restrained: you know what is certain, what needs qualifiers, and what should be cut. This workflow keeps working even when you switch models.
The next article moves to the global professional route. The focus is no longer adding a few subscriptions, but splitting research, editing, visuals, audio, and automation into distinct roles, backed by systems for versioning, budgets, licensing, and review.
Illustration Production Checklist (20 items)
All 20 illustrations below are already embedded in the body as SVG illustrations (interface-style images are workflow mockups, not real product interfaces). For formal publication, you may replace "operation screenshot" mockups with real screenshots, hiding accounts, keys, private data, and unauthorized content.
| No. | Visual | Format | Suggested position | Production notes |
|---|---|---|---|---|
| 01 | Article cover | Existing SVG | Opening | Emphasize "source-driven" |
| 02 | Sources-first workflow | Existing SVG | After the introduction | Collect, cross-check, localize, publish |
| 03 | Tool division of labor | Existing SVG | Tools section | Ideation, search, source library, visuals, editing |
| 04 | Pyramid of four source types | Infographic, embedded as 03-source-pyramid.svg | Section 2 | Primary sources first, lead material last |
| 05 | Source table | Table screenshot, embedded as 03-source-table.svg | Section 2 | ID, type, purpose, and evidence eligibility |
| 06 | Notebook material space | Operation screenshot, embedded as 03-notebook-ui.svg | Section 3 | Three sources and the Q&A panel |
| 07 | Evidence table | Table screenshot, embedded as 03-evidence-table.svg | Section 3 | Conclusions, sources, conditions, and disagreements |
| 08 | Does the citation support the conclusion | Before/after comparison, embedded as 03-citation-compare.svg | Section 4 | Topic-related citation vs. direct support |
| 09 | Source disagreement map | Relationship diagram, embedded as 03-disagreement-map.svg | Section 4 | Fact, terminology, time, and opinion disagreements |
| 10 | Bilingual glossary | Info card, embedded as 03-glossary-table.svg | Section 5 | English, translation, explanation, and disallowed renderings |
| 11 | Line-by-line translation vs. localization | Side-by-side comparison, embedded as 03-localization-compare.svg | Section 5 | Show context and example adjustments |
| 12 | 90-second structure | Timeline, embedded as 03-90s-structure.svg | Section 6 | Misconception, conclusion, three checks, and action |
| 13 | Key line of the script | Quote card, embedded as 03-key-quote.svg | Section 6 | "A citation is not automatic proof" |
| 14 | Eight-shot table | Table screenshot, embedded as 03-shot-table.svg | Section 7 | One explanatory visual per voiceover line |
| 15 | Source-bridge visual metaphor | AI-generated image, embedded as 03-source-bridge.svg | Section 7 | A checkpoint between conclusion and source |
| 16 | Bilingual prompt template | Prompt card, embedded as 03-bilingual-prompt.svg | Section 8 | Task, audience, and constraints side by side in two languages |
| 17 | Order of free-quota usage | Funnel diagram, embedded as 03-quota-funnel.svg | Section 9 | Text confirmation first, visual generation later |
| 18 | Source review checklist | Checkbox card, embedded as 03-source-checklist.svg | Section 10 | Nine pre-publish checks |
| 19 | Changes from the source-driven approach | Existing SVG | After Section 9 | Shifts in transparency, bilingual work, and verification |
| 20 | Retrospective loop | Existing SVG | Before the closing | Update sources, rules, and the next article |
Source Verification Note
The product capabilities discussed in this article were verified against the following official materials: OpenAI ChatGPT Plans, Google's official Gemini pages and help documentation, and the official NotebookLM / Gemini Notebook pages and help documentation. Free plans, model access, quotas, and product names may change; rely on your actual account pages and the latest terms of service when in doubt.