What the researchers mapped
Somia Qudah, Nader Al-Refai and Abrar AlAbed-Alhaq published a study in Discover Artificial Intelligence that formalises something the profession has handled by instinct: Human–AI Interaction in Translation, shortened to HAIT. The study treats translation work as two separate workflows rather than one.
Ecosystem A is the pipeline: a CAT tool or machine translation engine sits in the middle, and generative AI is attached to it through terminology lookup, post-editing suggestions or segment-level prompting. Ecosystem B is the chat window, where a translator pastes paragraphs into a general conversational model and edits what comes back. Both exist in working practice. They demand different habits, and the authors argue they should be taught differently.
The eight HAIT requirements were then matched against the 21 competences of DigComp 3.0, the European Digital Competence Framework used across EU education and training. That produced 168 candidate relationships for the authors to evaluate one by one.
DigComp 3.0 is the fifth version of the framework. The first version was published in 2013, with updates in 2016, 2017 and 2022. It is developed by the European Commission's Joint Research Centre (JRC) together with the Directorate-General for Employment, Social Affairs and Inclusion (DG EMPL), and it is organised into five competence areas: Information search, evaluation and management; Communication and collaboration; Content creation; Safety, wellbeing and responsible use; and Problem solving and critical thinking. The 21 competences sit inside those five areas, which is the structure the authors checked the HAIT requirements against.
37 validated links, five missing competences
Of the 168 candidates, 37 relationships across 16 competences were formally accepted. That is 22% of the candidate set and 76% of the framework. Which means the generic European framework covers roughly three quarters of the ground, and five DigComp competences found no validated link to translator work at all.
The remaining requirements were classified as translation-specific residues: skills the DigComp structure does not carry, and which a translation curriculum has to supply on its own. Terminology management under time pressure is one obvious candidate. Deciding when a fluent LLM output is factually wrong is another, and DigComp was not written to describe it.
For anyone who builds training material, the useful part is the direction of the result. DigComp is not useless for translators. It is incomplete, and the study states where.
Four teaching levels, from first year to CPD
The requirements are ordered into a four-level pedagogical progression that runs from undergraduate training up to continuing professional development. A first-year student does not need the same HAIT competences as a freelancer who has been post-editing for a decade, and the progression separates the two rather than issuing one flat checklist.
Validation came from a course-embedded mixed-methods pilot with 42 second-year translation students. That is a small sample, and the authors present it as their own relevance check rather than a measurement of learning outcomes. Read the numbers accordingly.
Europe changed the syllabus first
The timing is awkward in a useful way. The European Master's in Translation network maintains a competence framework for university translation programmes across Europe. The HAIT mapping is best read as an audit of how a general digital framework lines up with that training, not as a claim about any single recent update.
That matters for accreditation. A programme can point at a published framework and a published mapping, and show which competences it covers and which residues it fills locally. In practice, the translation departments most likely to use this are the ones inside EMT-affiliated programmes, where the network's competence framework is the reference document.
What Article 50 changes for the output
The AI Act's transparency rules under Article 50 became enforceable in August 2026, after the transitional period ended. Providers of generative systems must mark synthetic output in a machine-readable format, and deployers publishing certain synthetic content must disclose it. For a translation desk the practical question is who counts as the deployer when a sentence passes through an LLM, gets edited by a human, and then ships to a client. That is a policy question for the agency, not something the HAIT study answers.
Data residency is the other European constraint. Ecosystem B sends client text to whichever endpoint the translator picked, and a US-hosted API with no EU processing option is a contractual problem for public-sector and legal clients. Ecosystem A keeps the text inside a tool the agency already vetted. This is one reason the pipeline model survives despite being less convenient.
Why the profession is already half-automated
Two figures frame the whole discussion. Around 50% of current professional translation projects incorporate machine translation, and more than a third of professional workflows actively use AI. The study is not describing an emerging practice. It is describing the ordinary working conditions of the people it studied.
The limitation is the sample size and the scope: 42 students in one pilot, eight requirements, one competence framework. The 131 candidate relationships that were rejected are not proof that DigComp is wrong for translators. They are a list of places where the framework describes digital competence in general and the translation desk in particular needs something else.
Do the 37 validated links apply to translators outside Europe?
Not directly. The mapping is built against DigComp 3.0 because European translation programmes and the EMT network use that framework as their reference. Curricula in other regions work from different competence models, so the specific links would need to be re-checked rather than imported.
What exactly is a translation-specific residue?
Any HAIT requirement that found no validated DigComp competence to attach to. The study's position is that these have to be covered by dedicated translation coursework, because a general digital skills framework does not describe them.
Does the study say which ecosystem produces better translations?
No. It maps the skills each workflow requires and orders them into a teaching progression. Quality comparison between CAT-integrated generative AI and direct chat interfaces is outside what the pilot measured.
