Translation Ethics in the Age of Artificial Intelligence: New Questions

The classic principles of translation ethics – confidentiality, impartiality, competence, accuracy – are still in place. What has changed is the setting in which those principles are applied. This is exactly where the debate on artificial intelligence and translation ethics begins: how are the old principles to be applied to the new tools?

Contents

The aim of this article is not to give answers but to frame the questions properly. The field is changing fast; an answer that looks settled today may be void within a year. Being cautious does not mean admitting a lack of knowledge; it means behaving honestly in a debate that has not yet matured.

There are six headings below. Under each of them stand a few open questions. The answers your organisation gives to those questions also form the framework of your procurement policy.

Why a new debate?

A significant part of the literature written on translation technology was produced before large-scale language models became widespread. The work by Pietrzak and Kornacki, the main source for this series of articles, dates from 2021 and gives its attention to computer-assisted translation tools. It offers no direct assessment of the tools that appeared afterwards.

This is not a shortcoming but a difference in timing. The consequence, however, is this: the academic framework we have does not directly answer the ethical questions raised by the new tools. Care is needed when carrying the observations in the book across to the new context.

The roots of the debate, on the other hand, are older. Data on the use of machine translation has been collected for more than a decade; an industry survey dated 2012 reported that 55 per cent of freelance translators had tried machine translation at least once (DePalma 2012). The question, in other words, is not new; what has changed is the scale and the spread.

This historical continuity matters, because it shows that we do not have to build the debate from scratch. The headings of confidentiality, responsibility and rates have been discussed in the context of machine translation for years. What is new is that these questions are now asked in ordinary work rather than in exceptional cases.

Transparency: should it be disclosed?

This is the first and most concrete question. If an automated tool has been used in a translation process, does the client need to know about it?

The argument in favour is simple: the client has a right to know how the service they are buying was produced, and the tool used also determines whether the document leaves the organisation or not. The argument against runs as follows: no industry declares every tool it uses; what matters is the quality of the result.

In practice the industry appears to be moving towards a middle course: what is disclosed is not the name of the tool but the nature of the process. The distinction between “this is human translation” and “this is post-editing of machine translation” clarifies both pricing and expectations.

The question left open is this: should this disclosure be a voluntary good practice or an obligation written into the contract? For now the answer depends on what the parties agree.

Transparency also works in the other direction: the buying side stating its own expectation clearly. If the request that no automated tool be used is not written into the contract, the supplier cannot be expected to guess it. Uncertainty produces the same outcome on both sides; a wrong result comes about without anyone doing anything wrong.

Where does responsibility lie?

When a translation error caused harm, the chain of responsibility always ended in the same place: the party that delivered the text. Automated tools add new actors to that chain, and the question becomes more complicated.

Actor	Role	Question left open
Translator	Delivers the text	To what extent is accepting a suggestion their own decision?
Translation provider	Sets up the process, chooses the tool	Who owns the risk arising from the choice of tool?
Tool provider	Supplies the model and the interface	What undertaking does it give for the quality of the output?
Client	Sets the process and the timescale	What part does asking for the cheap and fast option play?

In practice today the answer largely stays in the first row: the party that delivers is responsible. This follows from the way contracts are written. How responsibility is to be distributed among the actors, on the other hand, has not yet settled into an established practice.

The distinction between input data and training data

Two different questions are frequently confused in the data debate. The first: is the text I send stored after it has been processed? The second: is the stored text used to improve the service?

The answers to these two questions may differ. One provider may store the text temporarily and not use it for improvement; another may say it stores nothing at all while keeping fragments of text in its logs. The questions therefore have to be asked separately and the answers obtained in writing.

Ethically, the real issue is the chain of consent. The people named in a text sent for translation – the witness in a petition, the patient in a report – have not consented to such use, and obtaining their consent is often impossible. The organisation that owns the text can decide on its own behalf; it cannot decide on behalf of the third parties named in the text.

This is a principle independent of technology, and the new tools have simply made it more visible.

Labour, rates and post-editing

The most contested effect of automated tools is on the rates side. As working by correcting machine output became widespread, the question of how that work should be priced arose.

The trouble is that the editing load is not constant. In a technical manual made up of formulaic sentences, machine output may be largely usable. In a marketing text where style does the carrying, or in a contract where the terminology is delicate, the same output can take longer to fix than translating from scratch.

Applying a fixed discount rate ignores this variability. The sounder approach is to measure the editing load on a small sample and tie the price to that measurement. This method gives predictability to buyer and supplier alike.

The question left open runs deeper: which link in the chain will keep the gain from the increase in productivity? This is a matter of economic negotiation rather than a technical one, and it has not yet settled in the industry.

Automation anxiety and the future of the profession

The worry that technology will do away with the profession is not new in the translation field. The literature calls this state automation anxiety and notes that the anxiety often appears before any actual loss of work, and indeed independently of it (Vieira 2018).

The ethical significance of the anxiety is this: a professional who is unsure about the future of the profession may make different choices between quality and speed. In an environment where rates are under pressure, taking on more work means giving less time to each job. This is not an individual weakness but a structural outcome.

We do not have enough data to make a concrete prediction, and that is not the purpose of this article. The influence of the buying side, however, is beyond dispute: realistic timescales, predictable workflows and reasonable rates are the most direct determinants of quality. Your procurement policy is one party to the debate about the future of the profession.

Copyright and the rights in the source text

A translation is a product derived from a source work; it has a place of its own in copyright law. Automated tools have brought two new questions into this area.

The first is on the output side: who holds the rights in a text produced largely automatically? The second is on the input side: with what permissions were the texts used in developing the models used? Both questions are handled differently in different countries, and there is as yet no settled international practice.

For the party buying translation the practical upshot is narrow but clear: secure a contractual undertaking that the rights in the text you receive have been transferred to you and that it does not infringe the rights of third parties. This clause is the simplest way of protecting yourself against legal uncertainty.

The same clause should also cover the rights in the source text. If you do not hold the authority over the document you send for translation – if it is, say, a report published by another organisation – the right to translate it may depend on a separate permission. This is a matter independent of technology and one that has long been known.

The Türkiye context

Turkish is a language that is not as widely represented in large language resources as English or German. Its agglutinative structure and its relatively free word order create additional difficulty for automated systems. Using Turkish output without checking therefore carries a higher risk than in widely used language pairs.

The second local factor is document culture. In Türkiye a significant proportion of official procedures runs on notarised translation, and in that channel there is a natural person who makes the declaration. The presence of an automatically produced text in this channel does not remove the responsibility for that declaration; the person who signs is taken to have assumed responsibility for the accuracy of the text.

The third is the data side. Uploading documents containing personal data to tools with servers abroad raises questions about transfer abroad under Law No. 6698. The fact that a tool is popular, or offered free of charge, does not change that assessment.

What does this mean for you?

Uncertainty is not a reason for not deciding. Your organisation can put its own framework in place today; as the field develops, the framework is updated. To make a start, answering five questions is enough:

For which document types is the use of automated tools permitted, and for which is it prohibited?
Will there be a disclosure where they are used, and in what detail?
Does the contract state that input data will not be stored and will not be used for development?
How will post-editing of machine output be priced and how will quality be measured?
Are the rights in the delivered text and the undertaking not to infringe set down in writing?

Bringing these five answers together on a single page makes an adequate policy document for most organisations. The policy does not have to be perfect; it is enough for it to be written down and known.

We set our process up according to document classification, keep automated engines switched off for sensitive work, and state in the delivery note which method was used at which stage. For a framework suited to your organisation, get in touch with us; you can look through our other articles in the knowledge centre.

This article is for general information only and is not legal advice. For your specific situation, consult your legal adviser.

Frequently asked questions

Should the client be told that artificial intelligence was used in a translation?

The expectation spreading through the industry points towards disclosure. The soundest course is to write into the contract from the outset at which stages of the process automated tools will be used. Disclosure then ceases to be a matter for dispute.

If an error appears in a translation produced with an automated tool, who is responsible?

In practice, responsibility for the delivered text stays with the party that delivers it. A suggestion from the tool does not divide that responsibility. Organisations that want to clarify how responsibility is shared set it out separately in the contract.

How do I find out whether my text has been used in training a model?

The only way is to ask. Whether input data is stored and whether it is used to improve the service are separate questions. Ask for the answer to be written into the contract; general statements in the terms of use may not be clear enough.

Should the post-editing rate be lower than the translation rate?

There is no single established answer in the industry. The editing load varies a great deal with the type of text and the quality of the machine output; for some jobs it is quicker to translate from scratch. It is sounder to tie the rate to a measured editing load.

Sources

Pietrzak, P. & Kornacki, M. (2021). Using CAT Tools in Freelance Translation. Routledge, pp. 25, 40–46.
DePalma, D. A. (2012). Translation Productivity Benchmarks. Common Sense Advisory.
Vieira, L. N. (2018). Automation Anxiety and Translators. Translation Studies, 13(1).
Law No. 6698 on the Protection of Personal Data – mevzuat.gov.tr