Our research and editorial standards
How we research
and test AI.
ChinaAI helps readers use DeepSeek, Kimi and Qwen through guides based on official sources and recorded product tests. This page explains how to check our evidence, interpret test results and request a correction.
By China AI Chat Team · Updated 27 September 2026
Different claims need different evidence
Start with the claim you want to rely on. Is it a documented feature, an observed result or a calculated estimate? Each needs a different kind of support. A guide can include all three without every instruction having been tested.
Source-based guides. We use official documentation, model cards, pricing pages and product help centers to explain capabilities, access and costs. A provider’s performance claim remains a provider claim unless we independently test it. A source-review date does not mean that every feature was tested.
Hands-on observations. We describe actions performed in the named product and the outputs or files they produced. A successful preview, a downloaded archive and a working public website are separate checks.
Calculations. Cost examples identify the provider, model, rates, units and assumptions. An estimate based on token counts is not a measured bill. Cache savings, conversation history and retries can change the result.
Identify the service before comparing it
A consumer app, an official developer API, a third-party host and a downloadable model can have different features and prices. We distinguish the product, provider, model identifier, plan and region when those details affect the answer. A chat subscription does not automatically include API credit.
If an interface does not reveal the precise serving model, we say so. We do not infer it from a product announcement or treat an app’s name as a verified API version.
How we run and report practical tests
Our tests address specific tasks, such as extracting facts from a PDF or repairing a small program. The conditions and limits explain how closely a result matches your own use. These checks do not establish an overall model ranking. Limitations that affect the conclusion belong beside the result; supporting setup details belong in the test notes.
- Define the task before testing. Set the prompt, input and checkable requirements. In a direct comparison, use the same material and explain differences in settings or access.
- Record the conditions. Include the date, service, visible model or mode, account access, enabled tools and conversation context. Identify information the interface does not expose.
- Keep attempts traceable. Distinguish a first answer from a retry, regeneration or corrective follow-up. Keep the original result when reporting a correction. If a later prompt supplies the right answer, explain that assistance. Report failures that affect the conclusion; do not turn a selected success into a general success rate.
- Check the output. Use the check the task needs: verify arithmetic against the source, inspect generated files, run code checks or open a published link outside its editor. State which checks actually ran and in which application. A preview does not establish that an exported file works, and an assistant saying its code passed is not a substitute for a recorded test result.
- Limit the conclusion. State how many tasks or examples were tested and what remains untested. A short code check does not establish production readiness, and a single image edit does not establish general editing accuracy.
We use fictional or authorized test material. Published evidence must not expose credentials, private conversations or unrelated drafts. Generated illustrations are labelled, and display resizing or compression is distinguished from changes to the result being evaluated.
Examples you can inspect
These examples show how to read the evidence behind a result. Open the source material, compare it with the output, then check the scope of the conclusion.
- Check an answer against its source. Our DeepSeek PDF guide includes three small source PDFs, exact prompts and first answers. Readers can check the facts, page references and table calculation for themselves.
- Check citations separately from facts. In our Qwen PDF tests, the image-only brief produced the requested facts but two incorrect page references. The original answer and source pages make that distinction visible.
- Separate the first result from a correction. The Kimi Docs layout correction preserves the original document and a separately revised file, with before-and-after views in Kimi/WPS. That result does not establish desktop Word compatibility.
- Separate the assistant’s actions from independent checks. The Kimi Code repair includes the original project, a one-line change and independently run checks. Kimi’s own test commands failed in the recorded session; the independent checks are reported separately.
For comparisons between assistants, use our comparison directory. Each report has its own tasks, dates and account conditions. Results from different reports cannot be combined into a single overall ranking.
Recorded examples are demonstrations
The directory of official chat services also includes selected answers generated through Together AI and saved for display. Each identifies its model, prompt and generation date. These are recorded API examples, not live responses to visitor questions or representative benchmarks. The official apps may use different models and tools. ChinaAI does not currently offer live visitor chat.
AI assistance and human editorial review
We use AI tools to assist research, drafting, editing, coding and test execution. An AI-generated summary, citation or verdict is not evidence by itself: factual claims need a traceable source or an actual recorded observation.
AI tools help organise source material, check calculations and code, and improve explanations. Their output can still be wrong. Our editorial checks examine factual support, usefulness and clear English; AI-assisted checks do not count as a named human review.
China AI Chat Team is our editorial byline. Responsibility for the published content rests with the site’s publisher, identified on our About page. Ibrahim Ahmed and Ibrahim Abd Elhady contribute human editorial review. Their profiles explain their roles.
Read the review credit for the version it covers. A named credit applies to the version identified on that guide. An earlier-version credit does not mean a later edit has been reviewed by the same people. A revision requires fresh human review before it receives a current human review credit. Editorial review also does not mean that the reviewers personally performed every recorded test; test methods and evidence are documented separately.
Named review credits were first added on 26 September 2026. That date does not indicate a new product test or a review of later edits.
Review dates, updates and corrections
Tested identifies when the reported interaction took place. Sources checked identifies a documentation review. Updated marks a revision to that page; it does not by itself mean a new product test or a new human review. Linked guides retain their own evidence dates.
Model aliases, plans, prices and interface controls change. When current documentation conflicts with an older announcement or our observed interface, we explain the distinction if it affects the reader’s decision. We do not change a review date merely to make a page appear current.
To request a correction, use our contact page or email [email protected]. Include the page URL, the statement or step in question, and a supporting source or description of what changed. Do not send passwords, API keys or private conversation histories.
We review reports against the available evidence. Corrections that change a result, recommendation or important instruction should carry a dated note explaining the change; minor spelling and formatting fixes do not require a new test date.
27 September 2026: clarified how readers can inspect evidence and how human review credits apply to specific versions. No new product tests were performed for this revision.
Independence and commercial disclosure
ChinaAI is an independent publication, not an official DeepSeek, Moonshot AI or Alibaba service. Linking to a provider does not mean that provider endorses our conclusions.
Advertising, sponsorship, affiliate links or provider-supplied access must be disclosed when relevant. Compensation must not determine a test result or turn a commercial relationship into evidence of performance. A paid placement must be distinguishable from editorial content.
Before relying on a changing price, plan or feature, check the official source linked in the relevant guide. Our findings describe the evidence available on their stated dates.