Machine Translation Matters — but Without Human Oversight, It Can Create More Problems Than It Solves

What MT and HITL mean

Machine Translation (MT) refers to the automated translation of text from one language into another. Google Translate, DeepL and newer AI-powered language systems are familiar examples.

Human in the Loop (HITL) means keeping a qualified professional actively involved in the workflow. The translator does not merely correct spelling and grammar after the machine has finished. They assess the source text, its purpose, the intended audience and the possible consequences of each translation decision.

HITL should not be treated as a superficial final check. In a properly designed process, the human specialist:

• · Examines the document and its intended use before translation begins.
• · Establishes the appropriate terminology.
• · Compares the translated text with the original.
• · Identifies omissions, additions and shifts in meaning.
• · Verifies names, dates, amounts, reference numbers and institutional titles.
• · Adapts the text to the conventions of the target language and jurisdiction.
• · Decides whether the final translation is safe and suitable for its intended purpose.

A fluent translation is not necessarily an accurate one

Modern machine translation can produce remarkably natural sentences. This is valuable, but it can also create a false sense of security. One of the greatest risks is that an inaccurate translation may look polished, confident and entirely plausible. A reader who does not understand the source language may have no obvious reason to question it. Errors can therefore remain hidden until the document is examined by a public authority, a lawyer, a university, a bank, or another professional.

Machine translation systems may struggle with several critical areas.

Context-dependent terminology

The same word can carry different meanings in legal, medical, academic, technical and everyday contexts. A system may select a linguistically possible equivalent that is nevertheless wrong for the particular document.

Legal effect

Terms such as "shall", "may", "must", "subject to" and "without prejudice" cannot always be replaced mechanically. An apparently minor choice may alter whether a clause expresses an obligation, a possibility, a condition or a limitation.

Document-specific conventions

A birth certificate, criminal record, diploma, power of attorney and court decision are not simply collections of sentences. Each document type has its own terminology, structure and administrative function. A translation can sound natural while still failing to match the conventions expected by the receiving institution.

Names, numbers and dates

Automated systems may translate proper names, omit numbers, confuse date formats or alter reference codes. In an official document, a single misplaced character can cause delays, inconsistencies or rejection.

Unsupported additions

AI-based systems may occasionally produce information that is not present in the source, or present an ambiguous passage as though it were certain. The risk becomes greater when the original document is incomplete, poorly scanned or difficult to read.

Why "we can fix it later" is a costly assumption

It may seem efficient to generate a complete machine translation first and ask a professional to correct it afterwards. In practice, repairing a poor machine-generated translation can be more demanding than translating the document properly from the beginning.

The translator is not working from a blank page. They are presented with a finished-looking text containing an unknown number of visible and hidden errors. Every sentence may need to be checked against the source. The reviewer must determine: which parts can be trusted; what information has been omitted; whether the system has added unsupported meaning; where incorrect terminology has spread throughout the document; whether individual errors have affected the logic of later sections.

A mistranslated term rarely remains isolated. It may appear in headings, tables, definitions and related clauses. Once the wrong interpretation becomes embedded in the text, correcting it requires more than replacing a few words — the entire structure may need to be reconsidered.

This is why post-editing should not be confused with proofreading. Proofreading normally improves a translation that is already fundamentally sound. Genuine post-editing requires the reviewer to compare the target text with the source and verify the translation itself. If the source text is not examined carefully, the process is merely polishing the surface. An elegant sentence can still communicate the wrong meaning.

The anchoring effect of machine output

There is also a less visible problem: machine-generated wording can influence the human reviewer. When a complete translation is already on the screen, the reviewer may unconsciously treat it as the starting point. A phrase that would not have been chosen in a translation produced from scratch may appear acceptable simply because it is already there.

This is sometimes described as an anchoring effect. The machine's first suggestion shapes the reviewer's judgement. For low-risk content, this may be manageable. In high-stakes material, however, a plausible but inaccurate starting point can make errors harder to detect.

Human involvement must therefore be active and critical, not ceremonial. A translator who is expected to approve machine output quickly is not truly operating within a Human-in-the-Loop system. Human presence alone is insufficient; the professional must have the authority, time and access to the source material required to challenge and replace the machine's decisions.

HITL must begin before the final review

The most effective form of human oversight does not wait until errors have already spread through a document. A language professional involved from the beginning can:

• · Decide whether the material is suitable for machine translation.
• · Identify sensitive or high-risk sections.
• · Establish terminology and formatting rules.
• · Separate repetitive content from passages requiring direct human translation.
• · Prevent avoidable errors before they enter the workflow.
• · Apply the appropriate level of review to each document.

Not every text requires the same process. Machine translation can provide significant benefits for high-volume, repetitive and low-risk content. Other material may require full post-editing by a qualified specialist. Legal, official or personally consequential documents — such as those submitted to a foreign embassy or institution — may be better translated directly by a professional from the outset. The correct method depends on the document, its destination and the consequences of an error.

Who is responsible for the final translation?

A machine can generate a translation, but it cannot accept professional responsibility for it. It cannot ask the document holder to clarify an illegible sentence. It cannot explain why a particular legal equivalent was chosen. It cannot determine whether a foreign authority will interpret a title correctly. Most importantly, it cannot stand behind the final document.

This distinction is particularly important in certified and sworn translation. The role of a sworn translator is not to place a signature beneath unverified machine output. The translator must examine the source, verify the target text, and ensure that the translation is accurate and complete before accepting professional responsibility for it. Technology may support that work, but it cannot replace accountability.

This matters in every context where a translated document carries legal or official consequences: a residence application, a marriage certificate submitted to a foreign registry, a diploma presented to a credential evaluation body, a court decision used in international proceedings.

The best model: technology guided by expertise

Rejecting machine translation altogether would ignore its genuine benefits. Using it without meaningful human control would ignore its limitations. The strongest approach combines the efficiency of technology with the judgement of an experienced professional.

The objective should not be to produce a translation as quickly as possible and repair the consequences later. It should be to select the right process at the beginning and produce a reliable, fit-for-purpose translation the first time. Correcting a deeply flawed translation is not always cheaper or easier than starting again. In many cases, identifying and removing embedded errors requires more effort than creating a sound translation from the source.

Machine translation is not making professional translators irrelevant. It is making their most important contributions clearer: judgement, subject knowledge, contextual understanding and responsibility.