Checking AI research: evaluating sources and tracing claims

Formulating a question that can be verified

Narrow down topic, time period and decision. Instead of “Which AI model is the best?” ask, for example, about models for a specific use case, a defined hardware framework and the status as of a specific date. Define whether you need an overview, a comparison or a recommendation.

Mention important exclusion criteria and the desired language. For current topics, explicitly ask for current research and sources. For a comparison, the answer should apply the same criteria to all candidates instead of listing different advantages for each. The LXI AI Chat supports work on questions, texts and research tasks.

Choosing sources by closeness to the claim

First check whether a source really supports the specific claim. For a product feature, the manufacturer's documentation is helpful; for a research result, the associated publication. A press release can reliably show what a company has announced, but is not automatically an independent assessment of its performance.

Open the linked pages. Compare publication date, change date if applicable, and the time of the event described. Several articles that all copy the same report are not independent confirmations of each other. Note the original evidence and, for disputed issues, add different perspectives.

Checking figures and comparisons for common conditions

For percentages, rankings and benchmarks, pay attention to test conditions, versions and measurement methods. A result may apply to a specific dataset without reflecting performance in your application. Do not compare values whose tasks, settings or evaluation methods differ substantially.

For your decision, a small test of your own is often more useful than a large table without context. Define typical tasks and evaluate the results according to predefined criteria. Separate published measurements from your own observations and from estimates. Missing values should remain recognisable as missing.

The research checklist before sharing

• Is the original question answered, including important limitations?
• Is there a directly relevant, accessible source for key claims?
• Is the information current enough for the specific decision?
• Are facts, interpretation and recommendation distinguished from each other?
• Are figures given with unit, time period and comparison conditions?
• Are contradictions explained instead of silently omitted?
• Do language and technical terms remain consistent throughout the report?
• Is it clear which questions remain open?

If a source cannot be found, ask for a correction with a verifiable source. A confident formulation is no reason to adopt an unsubstantiated claim. For decisions with major consequences, additional expert review is required.

Turning results into a traceable report

Structure the report with question, method, results, assessment and limitations. Place sources as close as possible to the respective claim and add a consistent bibliography. Record the research date so readers can assess timeliness. A summary should contain the same limitations as the main text.

A final fact check of quotes, figures and sources makes your report reliable. For work at universities, the respective rules on the use of AI apply. If you want to research with internal company documents, you will find the right starting point in the Private AI guide.