AI eloquence tempts institutions to abandon human translators, hiding errors that alter meaning

Google Translate alone handles a trillion words per month, while direct sales of independent machine-translation software reached about $1.4 billion last year. With the arrival of advanced models such as DeepL, Microsoft, Claude, ChatGPT and Gemini, institutions face an escalating economic temptation to eliminate human cost.Multiplying expenses a hundredfold to audit the remaining share of slip-ups may seem unjustified in internal documents, but that economic logic collapses entirely when words are tied to life-changing decisions.

John Zimmermann, founder of the Open Knowledge Foundation (Switzerland), a non-profit organization, explains this shift precisely; his institution has become the largest contributor to translating Wikipedia articles, with more than ten thousand pieces rendered into Spanish, Portuguese and English. That massive task would have cost millions of dollars via the traditional route. Two or three years ago, human translators were re-phrasing large portions of model drafts; today the output flows with a superficial smoothness that leaves little room for quick doubts, prompting many sectors to consider removing the human element from the workflow entirely.

Field testing, however, reveals a troubling side to this artificial eloquence. In a study conducted by the Swiss organization with Wikimedia funding that covered 119 articles, ten language pairs and more than a thousand hours of translator work, 27 percent of the machine-generated text underwent human correction before acceptance, and three quarters of those edits addressed errors produced by the models themselves, while 5-6 percent of the errors altered the original meaning of the text entirely.The danger of the new generation of algorithms lies in its ability to craft serious errors in natural language that appears highly coherent and persuasive.As Carol Perouiter, a specialist in medical translation, points out, flawed sentences no longer jump out at the reader with clumsy grammar as they once did; instead they now appear in a natural, cohesive style that is difficult to detect.

This gap becomes a critical liability in sensitive sectors; a procedural error in machine translation of asylum applications in the United States has already delayed the release of detainees. Legal expert Ariella Baltiye recounts that automated systems rendered the plaintiff’s status in court documents as merely an applicant and confused nouns with verbs, describing a family situation as a new marriage rather than an existing one. In the medical field, a drug dosage was once multiplied a thousandfold when the algorithm read the Greek letter denoting micrograms as milligrams. François Evon, head of machine-translation research at the French National Centre for Scientific Research (CNRS), notes that these systems rely on statistical probabilities rather than understanding; this mechanism can mistake shoe polishing for the name of a Polish language, or fail to distinguish a corner kick in a football match, making absolute reliability unattainable with their current architecture.

This reality demands a radical reassessment of automation plans for institutions and firms across the Gulf, Egypt and the Arab Levant. With a flood of massive infrastructure tenders, cross-border investment partnerships and multilingual legal disputes, the rush to cut translation budgets by relying exclusively on machines becomes a costly contractual trap. The initial financial saving on drafting bills turns into severe legal risk when binding documents contain a hidden distortion in liability clauses or payment terms. The needed shift in the regional business environment does not lie in abandoning the tools, but in redefining the human reviewer’s role: moving them from an editor who smooths clumsy phrasing to an expert inspector who tracks meaning distortions hidden behind deceptive artificial eloquence.

The next pivotal stage will not be achieved when models translate longer texts faster, but when software gains the ability to doubt and notify us that it is not confident in its outputs.And until the synchronous review environments proposed by researchers mature, where one model translates, a second critiques it and a third refines it, blind reliance on machines in contract editors and health-care protocols remains a uncalculated gamble that puts signatories’ reputations and professional futures at stake.