Covert Machine Translation Use: Should the User Be Told?

At some point, reading a product description on a shopping site, you will have noticed that the text had been translated automatically. Most of the time you do not notice. That second situation is what is called covert machine translation: the user does not know that the text being read has come out of an engine. In its opposite, overt use, the user presses the button personally and knows what they are getting.

Contents

The distinction is not technical but communicative. The same engine may be producing output of the same quality. What changes is the user’s expectation and where responsibility sits. A reader who knows a text is automatic puts an odd turn of phrase down to the machine. A reader who does not know puts it down to the organisation.

In this article we take up the reasons for covert use, the feedback loops the large platforms have built and the practical consequences of the transparency decision. Let us give the answer at the outset: you are not obliged to tell the user, but there is a price to not telling them.

The Difference Between Covert Machine Translation and Overt Use

In overt use the user is aware that the text is machine output. They have pressed the language button on the page themselves, or there is a note in the interface saying so. Expectations are set accordingly: the user wants to form a quick idea and is tolerant of flaws.

In covert use there is no such signal. Large portals serve content directly according to the locale the user has chosen. The selection is mostly made by looking at the language of the operating system or the geographical location of the connection. The user never once questions how the text on the screen was produced.

The importance of this distinction emerges at the moment of error. In overt use a broken sentence reads as a limit of the engine. In covert use the same sentence reads as corporate carelessness. Covert use, in other words, ties quality risk directly to brand reputation.

Criterion	Overt use	Covert use
What the user knows	Knows the text is automatic	Does not know
Trigger	The user’s own choice	The system, the locale or the location
How an error is read	A limit of the engine	Carelessness by the organisation
Tolerance	High	Low
Typical setting	Public portals, archive pages	Large commerce and classified-ad platforms

Why Do the Large Portals Say Nothing?

The first reason is preserving the flow. Showing the user an extra notice during the purchase journey is a distraction and can lower the conversion rate. Platforms therefore design translation as an invisible infrastructure layer. The user sees a product and takes no interest in what the source language was.

The second reason is scale. On a platform holding millions of listings, marking up how every single item was produced is a serious administrative burden. On top of that, on the same page some fields may have been produced by hand and others automatically. A single label does not describe that complexity accurately.

The third reason is commercial perception. In some markets an automatic translation label can make the product or the seller look less trustworthy. The concern is understandable, but if the user is noticing the flaw anyway, the absence of a label does not protect trust, it merely removes the explanation.

Against that, covert use carries one further cost: the feedback channel closes. When the user does not know a text is automatic, they feel no need to report an error, because they take what is in front of them to be a deliberate choice by the organisation. Quality problems therefore never find their way back into the system and can stay live for years without being noticed.

The Fit-for-Purpose Test: The Help Pages Model

The most mature form of covert use is the model that measures after publication. Microsoft asked users of its automatically translated support and help pages whether the page had solved their problem. Where the answers were positive, that page was accepted as fit for purpose.

This approach changes the definition of quality. The criterion is no longer linguistic perfection but whether the user can get the job done. Even if a help article is written in somewhat rough Turkish, if the user can find the setting and solve the problem, that text counts as having served its purpose.

The strength of the model lies in its being measurable. Which page is not working is determined by data rather than guesswork. The budget for human correction is then directed not at random but at the pages that fail most often. We take up the differences between definitions of quality in detail in the quality articles in our knowledge centre.

The limit of the method is just as clear. The users who give feedback may not represent all users. Those who fail to understand the page and leave straight away never reach the survey at all. Feedback data therefore has to be read together with behavioural measures such as the abandonment rate.

Upgrading on a Demand Signal: The Commerce Platform Model

The second mature model follows demand. On a commerce platform, product descriptions are first published in automatic translation. When visits or purchase demand from a particular region rise, that page is flagged for human correction or for full localisation. The investment is made once interest has been proven.

This logic can be adapted for small-scale sites as well. Instead of localising the whole catalogue from the outset, a broad automatic coverage is opened up in the first stage. Language and country data from analytics tools shows which pages are to be promoted. The budget is then distributed on the basis of evidence.

The model has a side effect: the first impression is formed by automatic output. If demand is low to begin with, it becomes harder to tell whether the cause is the product or a poor translation. At the very least, therefore, category pages and the main entry points should be prepared by hand.

Setting the promotion threshold down in writing in advance also makes things easier. It can be defined as a rule, for instance, that a page will be sent for human correction once monthly visits or the add-to-basket rate in a given language pass a certain level. The decision is then not reopened every time and the budget stays predictable.

Crowd Post-editing: Bringing the User into the Process

The third approach invites the user to make corrections. In this practice, known as crowd post-editing, community members do not translate the text from scratch but correct the engine output (Tatsumi et al. 2012). Mitchell et al. (2014) define it as a community practice of editing rather than translating directly.

Microsoft’s community translation framework is one of the known examples of this approach (Aikawa et al. 2012). Participants were able to suggest corrections on automatically translated pages. A similar arrangement was used in the translation services of the large search engines too. The shared idea is that the person best placed to spot an error is the reader.

For the approach to work, a willing community is needed. In organisations without a loyal body of users, participation generally stays low. Reviewing the suggestions that come in also takes separate effort; corrections accepted without review can break consistency.

The Transparency Decision: What Does Labelling Change?

The labelling decision is a question of balance. Putting up an automatic translation notice lowers expectations and explains the flaws that may appear. The same notice can leave some users with the impression that the content is second class. The right answer varies with the type of content.

A practical criterion is this: if misunderstanding the text can harm the user, covert use should be avoided. In health, legal, safety and financial content, publishing raw engine output without saying so is not a defensible choice.

Labelling also opens a feedback channel. When the user knows a text is automatic, reporting an error becomes easier. Over time that produces free quality data, which feeds the correction priorities. Transparency is not only an ethical choice, it is a functional instrument.

The form of the notice matters as well. A small warning hidden away at the very bottom of the page and a short statement visible at the start of the content do not have the same effect. The aim is not to alarm the user but to set expectations in the right place. A single sentence, plain and free of blame, is enough in most cases.

The Türkiye context

In Türkiye, covert use is seen most often on marketplaces and classified-ad sites. Product descriptions coming from abroad are mostly turned into Turkish automatically. The odd product names that result can create a problem of trust in the seller on the consumer’s part.

The structure of Turkish enlarges this risk. Because the language is agglutinative, a wrong inflection stands out immediately in short product titles. And when units of measurement, size charts and material names coming from English are not converted into their local equivalents, they create outright uncertainty for the buyer.

Consumer legislation calls for care too. Information such as the essential characteristics of the product, the warranty terms and the right of withdrawal is expected to be presented accurately and understandably. If an automatic translation error corrupts that information, responsibility remains with the seller. This article is for general information only and is not legal advice.

What does this mean for you?

Ruling out automatic translation altogether is not realistic at most content volumes. What can be done is to turn covert use into a conscious decision. The sequence below may help you frame that decision.

1. Divide your content into three groups by the cost of an error: critical, medium and low risk.
2. Do not publish automatic output in the critical group; do not put anything live before it has been corrected.
3. Add a short automatic translation notice to the pages in the medium group and place an error reporting link there.
4. Set up measurement in the low-risk group; build a promotion list from the abandonment rate and the feedback data.
5. Send the promoted pages for human correction in three-month cycles and measure the result.

Once this loop is in place, automatic translation stops being a source of risk and turns into a managed layer. If you are not sure which group to put your content in, you can get in touch with us with sample pages, and we can work out the risk classification together.

Frequently asked questions

What does covert machine translation mean?

Covert use is the term for situations in which the user does not know that the text being read is an automatic translation. In overt use the user presses a button and knows the output has come from a machine. The difference lies in where expectation and responsibility sit.

Why do platforms prefer covert use?

The reason is usually that the flow should not be broken. Offering the user an extra choice can lower the conversion rate. And when the content volume is very high, marking up how each page was produced brings an administrative burden.

How does the fit-for-purpose criterion work?

Feedback is gathered from users after the content has been published. Users are asked, for example, whether a help page solved their problem. If the majority can solve their problem, the page counts as fit for purpose; if they cannot, it is queued for human correction.

Should I state that I use automatic translation?

There is no obligation, but saying so works in your favour in most cases. It sets user expectations correctly, makes error reporting easier and shows good faith in any dispute that may arise. In legal and medical content, not publishing automatic output at all is the safest route.

References

• Jiménez-Crespo, M. A. (2024). Localization in Translation. Routledge, pp. 246-247.
• Tatsumi, M. et al. (2012). How good is crowd post-editing? AMTA Workshop on Post-editing Technology and Practice.
• Mitchell, L. et al. (2014). Community post-editing of machine-translated user-generated content. Post-editing of Machine Translation.
• Aikawa, T. et al. (2012). The impact of crowdsourcing post-editing with the collaborative translation framework. JapTAL Proceedings.