
Introduction
Since we published our first CIOL AI Voices, CIOL Council and delegates at our webinars and Conference days have continued to underline the importance of enabling linguists to keep up to date with developments in AI with a regularly refreshed resource hub.
Here are some recent resources for linguists, which we will continue to add to and update regularly, alongside CIOL AI Voices 2025 and future CIOL Voices updates.
Why AI-generated text feels "flat" and underwhelming
A recent paper by Cory Massaro from CIOL HE partner University of Bristol, examines a question many language professionals will recognise: why does AI-generated writing often feel ‘flat’ compared to human writing, even when it is ostensibly accurate?
Using statistical analysis of millions of words, the study finds that AI/Large Language Models rely more heavily on repeated phrases and familiar patterns. Human writers, by contrast, tend to use a wider range of vocabulary and phrase combinations.
The author compared large collections of human-written and AI-generated texts. The results suggest that people introduce new ways of expressing ideas more frequently, while AI systems are more likely to return to the same phrases and constructions. This difference becomes particularly noticeable when looking at longer sequences of words, helping to explain why AI-generated text can sometimes feel repetitive or predictable despite appearing fluent and well structured.
The paper also argues that this is not simply a question of style. If an AI model draws repeatedly on a relatively limited set of linguistic patterns, it may also limit the range of ideas it is likely to express. Massaro concludes that the reduced variety found in current AI-generated text affects both readability and the richness of communication.
For linguists, translators and language practitioners, the research provides useful evidence that the differences between human and AI writing - and even more significantly in translations - are not just subjective impressions but can be measured, analysed and demonstrated.
When AI Builds itself - Anthropic
Anthropic’s recent paper on its Claude models documents a marked shift in AI development. It reports that Claude now generates more than 80% of the code merged into Anthropic’s production systems, with engineers increasingly moving from writing code to directing and reviewing it.
The paper also shows rapid gains in capability: the duration of tasks models can complete independently is doubling over time, and Claude is able to run increasingly complex experiments with limited human input. Across both engineering and research workflows, AI systems are taking on a growing share of implementation and execution.
The paper sets out “recursive self‑improvement” as a plausible next step, in which AI systems could design and build their own successors. While this is not yet realised, Anthropic notes that human roles are already narrowing towards goal‑setting and oversight.
For language professionals, the development is notable because it shifts more of the production and handling of language into systems operating with reduced human involvement, including in areas where meaning and interpretation are critical.
The authors highlight the need for wider attention to these implications. The full article is available here: When AI builds itself.
The Pope's Encyclical Letter on Safeguarding the Human Person in the time of AI
In his recent encyclical Magnifica Humanitas (May 2026), Pope Leo XIV places artificial intelligence firmly within a longer human story of language, communication and responsibility.
He recognises AI as “a valuable tool” that can extend human capacities, but insists it must be guided by ethical judgement, social purpose and respect for human dignity.
In a world where digital systems increasingly shape decision‑making processes and the collective imagination, he calls for clarity about who designs, governs and benefits from these technologies, and for renewed attention to truth, communication and education in the digital age.
Particularly relevant sections are paragraphs 4–6 (framing AI), paragraphs 99–100 (language and understanding), and paragraphs 132–137 (communication and truth in the digital age).
Paragraph 99 is among the most salient for working linguists:
It is not possible to provide a single, comprehensive definition of AI. What can be stated, however, is that we must avoid the misconception of equating this type of “intelligence” with that of human beings. These systems merely imitate certain functions of human intelligence. In doing so, they often surpass human intelligence in speed and computational capacity, offering tangible benefits across many fields. Yet this power remains entirely tied to data processing. So-called artificial intelligences do not undergo experiences, do not possess a body, do not feel joy or pain, do not mature through relationships and do not know from within what love, work, friendship or responsibility mean. Nor do they have a moral conscience, since they do not judge good and evil, grasp the ultimate meaning of situations, or bear responsibility for consequences. They may imitate language, behavior and analytical skills, or even simulate empathy and understanding, but they do not understand what they produce, for they lack the affective, relational and spiritual perspective through which human beings grow in wisdom. Even when these tools are described as capable of “learning,” their way of doing so is different from that of a human person. It is not the experience of those who allow themselves to be shaped by life and grow over time through choices, mistakes, forgiveness and fidelity. Rather, it is a form of statistical adaptation based on data and feedback, which can be very effective, but does not imply inner growth.
Recent AI Research, News & Articles for Linguists
Here Come the Student AI Resisters
In The Chronicle of Higher Education, Len Gutkin describes the emergence of “student AI resisters”, articulating a widespread concern that routine reliance on generative AI tools may weaken habits of thinking, writing and intellectual effort.
Students report both dependence and discomfort, describing a culture in which AI use is commonplace, often invisible, and only lightly monitored, yet increasingly regarded with embarrassment or concern. This tension, captured in the paradox of students who are “anti‑AI but use AI”, points to a deeper question about the purposes of education and the value of sustained engagement with language, ideas and argument.
Which countries are the worst and best in the public sector AI race?
France, Germany, and the United Kingdom are struggling to translate AI ambitions into practical tools for civil servants, according to the Public Sector AI Adoption Index 2026.
The study, published by Public First for the Center for Data Innovation with sponsorship from Google, reveals that while AI use is accelerating globally, European nations are taking a more cautious, risk-averse approach that has left many public servants without access to the technology reshaping government work elsewhere.
Edelman Trust Barometer 2025: Trust and AI at a Crossroads
This Edelman global survey (Brazil, China, Germany, UK, US) reveals deep divides in attitudes towards AI adoption, and confirms again (see Ipsos) that scepticism is strongest in Europe and Anglophone markets, including the UK.
Key points for linguists and CIOL members:
- Resistance Outweighs Enthusiasm: In the UK and US, rejection of AI use significantly exceeds acceptance. Over 40% resist increased adoption.
- Trust Deficit: Trust in AI lags behind trust in the wider 'tech sector' by 26 percentage points. Concerns about data use and data protection are major barriers.
- Economic Anxiety: Lower-income groups fear being left behind, but even high earners (notably in the USA) feel vulnerable. This mirrors concerns in language services about downward pressure on fees and job security.
- Workplace Impact: Positive experiences with AI at work can improve trust—but transparency and training are critical. 76% say they would embrace AI only if assured it is for productivity, not job cuts.
- Drivers of Adoption: Knowledge and trust are key. Those who feel informed are far more likely to accept AI, yet many linguists see poor performance in low-resource languages and lack of transparency as good reasons to resist AI.
- Trusted Voices: “Someone like me” and peers are the most trusted sources on AI, not CEOs or government leaders, suggesting that professional communities will shape attitudes more than corporate messaging.
Implications for language professionals: The report underscores that trust and reassurance on collective benefit are essential factors in AI adoption. For many linguists, this means evidence of quality, especially for minority and low-resource languages, fair rates and ethical use. Without these safeguards, AI scepticism will remain, albeit the very different perspectives in Brazil and China show that different countries see the balance of risks and possibilities of AI very differently.
European Language Council Report: AI for Translation and Interpreting
The European Language Council's Special Interest Group on AI in Translation and Interpreting has released a comprehensive reflection paper examining the impact of Large Language Models (LLMs) on the profession. The report, authored by 25 scholars from 19 universities across 14 countries, provides crucial insights for both practitioners and policymakers.
Recommendations:
The report calls for informed AI use in appropriate contexts only, while maintaining human expertise for sensitive settings (legal, medical, asylum proceedings). It emphasises that professional translators and interpreters remain essential for ensuring communication quality, cultural mediation, and protecting client rights. The authors advocate for comprehensive AI literacy training alongside reinforcement of core human skills - language proficiency, cultural competence, and critical thinking.
Key Survey Findings (600 T&I professionals, 55 countries):
- 66% rarely or never use GenAI in their work
- 80% express concerns about output quality, with only 5% seeing quality as a benefit
- 62% believe GenAI threatens their livelihood
- 39% report increased mental effort when post-editing AI output
- 76% agree proper AI training is essential
Core Concerns Identified:
The report emphasises that LLMs generate text through probabilistic prediction rather than genuine understanding, creating "artificial eloquence" that may mask serious errors. Critical issues include:
- Deceptive fluency: Output appears convincing but may contain factual errors, inappropriate terminology, or cultural misunderstandings
- Linguistic inequality: Systems favour English and well-resourced languages, producing significantly poorer results for languages of lesser diffusion, including major migration languages
- Professional impact: Shift from creative translation work to correcting flawed machine output, leading to reduced job satisfaction and concerns about deskilling
- Legal and ethical risks: Unclear accountability for errors, data protection concerns, and potential copyright violations when using client materials
In Summary:
With this reflection paper, the authors discuss the use and impact of Large Language Models (LLMs) in multi- and interlingual communication and mediation. Their main goal is to raise awareness with users, both laypersons and professionals, as well as policymakers, concerning:
- How LLMs ‘communicate’, how this differs from human communication and what these differences imply when GenAI is used for T&I purposes (Section 1);
- What impact AI is having and is likely to have in the future, in T&I professions, in terms of communication quality, workflows, job content and job contentment (Section 2);
- What uninvited side-effects should be considered and duly weighed against the possibilities offered by AI (Section 3.);
- And whether and how AI should be included in T&I education and training, i.e. the formal education of future T&I professionals, as well as in-service training for current T&I professionals (Section 4).
The paper ends with a useful executive summary:
Microsoft AI Economy Institute - AI Diffusion Report
Microsoft's latest AI Diffusion Report clearly flags the significance and challenges posed by the 'languages gap' in AI. The report sets out the following:
As a general-purpose technology, AI stands on the shoulders of three others—electricity, connectivity, and computing. Its adoption is fastest where these foundations exist, and slowest where they do not. Nearly four billion people—half the world—still lack the basics needed to use AI.
AI adoption in the Global North is roughly twice that of the Global South, with the gap widening sharply in countries where GDP per capita falls below $20,000. In some Global North countries, more than half of the working-age population uses AI, while in parts of Sub-Saharan Africa and Asia, some of the least developed nations have adoption rates below 10%.
A new barrier is also emerging: language. Nations where low-resource languages dominate—like Malawi or Laos—show lower adoption even after adjusting for GDP and internet access.
Unlike earlier technologies, AI is built on data—particularly human language. The open web, where most of humanity’s knowledge is stored, serves as the single most important dataset for AI. Yet half of this content is in English, even when English is spoken natively only by just 5% of the world’s population. For low-resource languages, AI models are often less proficient—or even nonexistent.
This mismatch excludes millions: in Malawi, fewer than 4% speak English fluently, while the majority rely on Chichewa, Chitumbuka, or Chiyao, languages almost absent from the web and LLMs. The result is a systemic barrier: even with access to electricity, the digital infrastructure and digital skills, AI remains inaccessible when people cannot engage with it in their own language.

Data Sources: Common Crawl Foundation; UNESCO
The volume of open web content differs greatly between languages. Swahili, spoken by over 200 million people, has over 500 times less digital content than German, despite comparable numbers of speakers. Our analysis shows that low-resource language countries adopt AI at rates 20% lower than high-resource language countries—even under similar GDP and connectivity conditions. Model performance across languages, as shown in the study by Xuan et al, indicates that even state-of-the-art LLMs achieve around 80% accuracy in English but drop below 55% for some low-resource languages such as Yoruba (YO), one of Nigeria’s three major languages, spoken by more than 50 million people across Africa.
These gaps are not just technical; they determine who can access AI’s benefits in education, health, agriculture, finance, and public services. For some communities, like Māori speakers in New Zealand, bilingualism with English makes the issue one of cultural preservation. But for others—such as Hausa in Nigeria and Niger, or Guarani in Paraguay—unlocking AI in local languages is the only path to meaningful access.
Recent AI Research, News & Articles for Linguists
China’s DeepSeek makes rare public comment, calls for AI ‘whistle-blower’ on job losses
Chinese artificial intelligence start-up DeepSeek reaffirmed the AI lab’s commitment to developing artificial general intelligence (AGI) despite its potentially “dangerous” impacts on society.
Chen Deli spoke during a panel discussion alongside the heads of five other companies collectively known as China’s “six little dragons” of AI at the World Internet Conference in Wuzhen, eastern Zhejiang province.
Chen said AI’s current limitations mean humans and machines are in a “honeymoon phase”, but warned that most jobs could eventually be automated.
“Humans will be completely freed from work in the end, which might sound good but will actually shake society to its core,” he said, urging AI companies to act as “whistle-blowers” by warning the public about the jobs that would be made redundant first.
A Call for Collaborative Intelligence: Why Human-Agent Systems Should Precede AI Autonomy
Recent improvements in Large Language Models (LLMs) have led many researchers to focus on building fully autonomous AI agents.
This position paper questions whether such an approach is the right path forward, as these autonomous systems still have problems with reliability, transparency, and understanding the actual requirements of humans.
The researchers argue that deploying fully autonomous LLM-based agents in complex real-world scenarios, at this stage of development, poses significant risks and limitations that could undermine both safety and effectiveness.
Rather than viewing autonomy as the primary measure of progress, they argue for a fundamental shift toward LLM-based Human-Agent Systems (LLM-HAS), where AI agents function as active teammates rather than independent operators. They advocate for the focused development and deployment of collaborative, supportive, ethical, and adaptive AI-human partnerships that enhance human capabilities while maintaining essential human oversight and judgment.
They say this approach does not represent a retreat from AI’s ambitious goals, but rather a redefinition of what constitutes advanced AI—measuring progress not by isolation, but by collaborative intelligence.

AI translation service launched for fiction writers and publishers prompts dismay among translators
The Guardian reports that GlobeScribe.ai has launched a new AI translation service for fiction writers and publishers, offering translations at $100 per book, per language. Prominent literary translators, however, argue that AI lacks the nuanced understanding required for cultural and contextual depth.
“The best literary translations offer more than simple accuracy, more than literal fidelity to the words making up the sentences,” said Polly Barton, writer and translator of works including the bestselling Butter by Asako Yuzuki from Japanese to English. “They are engaging with the context from which the book has come, and reproducing the pacing, atmosphere, emotional timbre, rhythm, and all the other, less superficially obvious factors that ultimately determine how fulfilling and rich the reading experience is.”
Ian Giles, Chair of the Society of Authors’ Translators Association said: “Suggesting that AI can match, or even surpass, the nuanced work of human translators on behalf of authors is flat-out wrong.”
How Are Interpreters Using AI in 2025?
The latest Slator survey finds that 55% of interpreters surveyed now report using AI tools for their interpreting work, compared to 76% of translators who reported doing so (but only 53 interpreters responded to this question, while 336 translators responded to it.)
When asked how they were using AI in their practice, 19% of the interpreters reported “looking up terminology during interpreting delivery” followed by “extracting named entities and key terms from reference materials,” which was reported by 8% of the surveyed interpreters.
Grammaticality representation in ChatGPT as compared to linguists and laypeople
Large language models (LLMs) have demonstrated exceptional performance across some linguistic tasks. However, it remains uncertain whether LLMs have developed human-like fine-grained grammatical intuition.
This large-scale investigation of ChatGPT’s grammatical intuition, builds upon a previous study that collected laypeople’s grammatical judgments on 148 linguistic phenomena that linguists judged to be grammatical, ungrammatical, or marginally grammatical (Sprouse et al., 2013).
Overall, the findings demonstrate convergence rates ranging from 73% to 95% between ChatGPT and linguists, with an overall average of 89%. These results are attributed to the differences in judgement and in language processing styles between humans and LLMs.
As an example, half of the items exhibiting the greatest human-ChatGPT differences were related to the use of reflexive pronouns. In general, reflexives were judged to be very ungrammatical by ChatGPT, even for the cases that are grammatical to native English speakers. It is likely that when human participants saw sentences with reflexive pronouns, they tended to think of possible discourse contexts that could justify the use of the reflexives. By contrast, ChatGPT’s grammaticality judgment of the reflexive pronouns was based simply on the distribution of the model’s training data.
Google Calls for Rethink of Single-Metric AI Translation Evaluation
A new study by researchers from Google and Imperial College London challenges a core assumption in AI translation evaluation: that a single metric can capture both semantic accuracy and naturalness of translations. The researchers conclude:
“Single-score summaries do not and cannot give the complete picture of a system’s true performance,”
They observed that systems with the best automatic scores — based on neural metrics — did not receive the highest scores from human raters. “This and related phenomena motivated us to reexamine translation evaluation practices,” they added. The researchers argue that translation quality is fundamentally two-dimensional, encompassing both accuracy (also known as fidelity or adequacy) and naturalness (also known as intelligibility or fluency).
In this paper, they “mathematically prove and empirically demonstrate” that these two goals are in inherent tradeoff — and optimising for one often degrades the other, a point echoed in a recent Slator article that found separating accuracy and fluency improves AI translation evaluation.
How Mature is AI Interpreting?
AI speech technologies are rapidly advancing, potentially significantly impacting some parts of the traditionally human-led interpreting market.
Nearly half of respondents in a Slator weekly poll believe AI interpreting can be effective for niche scenarios, although almost a third believe that it is still highly experimental. However, the sample size was small (31 respondents) and the highly variable performance of AI by language was not part of the framing of the question.
To understand the future of AI, take a look at the failings of Google Translate
This article discusses the limitations of Google Translate, and how they highlight the broader challenges faced by AI.
Despite significant advancements, machine translation still struggles with idioms, place names, and technical terms, indicating that AI has a long way to go before achieving true linguistic and contextual understanding. The final paragraphs conclude: "...the forseeable future for LLMs is one in which they are excellent at a few tasks, mediocre in others, and unreliable elsewhere. We will use them where the risks are low, while they may harm unsuspecting users in high-risk settings."
The author asks: "The urgent question: is this really the future we want to build?"
Meta’s BOUQuET Brings Linguistic Diversity to AI Translation Evaluation
Meta unveiled BOUQuET, a comprehensive dataset and benchmarking initiative aimed at improving multilingual machine translation (MT) evaluation. The researchers noted that existing datasets and benchmarks often fall short due to their focus on English, narrow range of registers, reliance on automated data extraction and limited language coverage. These constraints hinder the ability to fairly evaluate translation quality across diverse linguistic contexts.
BOUQuET addresses these gaps by originating content in seven non-English languages — French, German, Hindi, Indonesian, Mandarin Chinese, Russian, and Spanish — before translating into English.
Experts Weigh In on DeepSeek AI Translation Quality
Chinese AI company DeepSeek launched two open-source large language models in January 2025 - DeepSeek-R1 and DeepSeek-V3 - generating significant attention for their ChatGPT-level performance at lower costs. Based in Hangzhou, the company achieved this through efficient training methods including improved Mixture of Experts (MoE) technique, allowing the models to be trained on lower-end hardware while maintaining high performance.
User feedback indicates strong translation capabilities across multiple languages, with particular excellence in Chinese-English translation. Users reported superior performance compared to other models in languages including Serbian, Spanish, Turkish, Czech, Hungarian, Punjabi, and Malayalam. Business adoption has proven cost-effective, with one CEO reporting 50x cost reduction compared to Google Translate, though some users still prefer alternatives like Claude or Copilot.
SAFE AI founding members join SlatorPod to discuss AI and Interpreting
Katharine Allen, Director of Language Industry Learning at Boostlingo, and Dr. Bill Rivers, Principal at WP Rivers & Associates, join SlatorPod to discuss the challenges and opportunities AI brings to interpreting. Both are founding members of the Interpreting SAFE AI Task Force, which aims to guide the responsible use of AI in language services (for more on SAFE AI see below).
Allen describes AI as a double-edged sword — capable of expanding multilingual access but limited in its ability to handle the nuanced human dialogue essential in fields like healthcare. She emphasizes the ongoing shift toward a hybrid model, where human interpreters collaborate with AI tools.
Google Finds ‘Refusal to Translate’ Most Common Form of LLM Verbosity
Researchers from Google have identified 'verbosity' as a key challenge in evaluating large language models (LLMs) for machine translation (MT). Verbosity refers to instances where LLMs provide the reasoning behind their translation choices, offer multiple translations, or refuse to translate certain content. This behavior contrasts with traditional MT systems, which are optimized for producing a single translation.
The study found that verbosity varies across models, with Google's Gemini-1.5-Pro being the most verbose and OpenAI’s GPT-4 among the least verbose. The most common form of verbosity was refusal to translate, notably seen in Claude-3.5.
A major concern is that current translation evaluation frameworks do not account for verbose behaviors, often penalizing models that exhibit verbosity, which can distort performance rankings. The researchers suggest two possible solutions: modifying LLM outputs to fit standardised evaluation metrics or updating evaluation frameworks to better accommodate varied responses. However, these solutions may not fully address verbosity-induced errors or reward 'useful' verbosity, highlighting the need for more nuanced evaluation methods.
AI Pioneer Thinks AI Is Dumber Than a Cat
Yann LeCun helped give birth to today’s artificial-intelligence boom. But he thinks many experts are exaggerating its power and peril, and he wants people to know it.
“We are used to the idea that people or entities that can express themselves, or manipulate language, are smart—but that’s not true,” says LeCun.
“You can manipulate language and not be smart, and that’s basically what LLMs are demonstrating.”
Today’s models are really just predicting the next word in a text, he says. But they’re so good at this that they fool us. And because of their enormous memory capacity, they can seem to be reasoning, when in fact they’re merely regurgitating information they’ve already been trained on.
Google Aligns LLM Translation with Human Translation Processes
Google researchers have developed a new multi-step process to improve translation quality in large language models (LLMs) by mimicking human translation workflows. This approach involves four stages: pre-translation research, drafting, refining, and proofreading and aims to enhance accuracy and context in translations. Tested across ten languages, this method outperformed traditional translation techniques, especially in translations where context is crucial.
How the Media Covers the 'AI vs Translators' Debate
Slator reports that translation is one of the most-referenced professions in 2024's media coverage of AI’s potential impact. Generally, according to Slator, media opinions on AI and translation can be grouped into one of four categories: "Humans Are Still Superior (For Now)" with c25% of articles, "Translators Are In Danger", accounting for c40% of articles, "AI + Humans = Optimal Translation" was the message of around 20% of articles and finally c15% of articles talked about translation but without mentioning translators at all. A mixed picture at best.
Ipsos - The Ipsos AI Monitor
Among a wealth of other data, the Ipsos AI Monitor 2025 confirmed findings from over the last three years that attitudes to AI differ significantly around the world.
Respondents from English-speaking countries are far more 'nervous' about AI than in Asian countries where respondents are far more 'excited'. In EU countries, respondents are both less 'excited' and less 'nervous', perhaps reflecting the increasingly more structured regulatory environment for AI in the EU.
However, AI sentiment appears to be warming globally, particularly in countries that were once the most skeptical. Among the 26 nations surveyed by Ipsos in both 2022 and 2024, 18 saw an increase in the proportion of people who believe AI products and services offer more benefits than drawbacks.
Given these attitude differences, the use and impact of AI in translation, interpreting and language learning may unfold very differently in different countries and regions around the world in the years to come.
Read our pre 2025 AI Updates here





