Kindly Do the Needful: The Words That AI Writing Tools Delete First #
The most common sentence typed into a writing assistant is a small act of surrender. Make this more professional. A banker in Lagos types it before sending a memo to a colleague she has known for a decade. A project manager in Pune types it before a status update that already says exactly what she means. A civil servant in Singapore types it before a reply that is, by any reasonable standard, already polite. The model does what it is told. The meaning arrives intact. The email gets a reply. Nothing appears to have gone wrong.
Something has gone missing, and until recently nobody had measured it. In a paper presented at the 2026 CHI conference on human factors in computing systems, Satyam Kumar Navneet, an independent researcher based in Bihar, working with Joydeep Chandra and Yong Zhang of Tsinghua University in Beijing, gave the phenomenon a name: cultural ghosting. The term describes what happens when a model rewrites a sentence in Indian, Singaporean or Nigerian English and returns it stripped of the politeness conventions, idioms and small verbal signatures that tell a reader who is writing. The meaning survives. The voice does not.
The team also gave it a number. Across 22,350 model outputs, the average Identity Erasure Rate, their metric for how much of a text's cultural marking disappears in processing, came out at 10.26 per cent. Models varied enormously: the worst performer, Mistral-7B, erased markers at 20.5 per cent, while the best, Qwen3-8B, erased them at 3.5 per cent. Meanwhile semantic similarity between the input and the output stayed high, at a mean of 0.748 on a scale where one is identical. The models keep what the sentence means and quietly delete what it is. The researchers call this the Semantic Preservation Paradox, and it is the reason the problem has stayed invisible for so long. Nothing looks broken. The sentence still works. Only the person has been removed from it.
For what the researchers describe as more than 1.5 billion speakers of World Englishes, this is not a rounding error. It is the writing infrastructure of several continents quietly teaching itself that one particular way of being clear is the only legitimate one.
An Experiment That Put a Number on the Vanishing #
The study is deceptively simple in design. The team assembled 1,490 texts that had been written by real people in real registers: emails drawn from the Enron corpus, the Clinton Archive, the EmailSum dataset, plus posts from Twitter, now X, and news articles. Of those, 601 texts carried markers of Indian English, 261 carried markers of Singaporean English and 89 carried markers of Nigerian English, alongside 539 American English texts used as a baseline. Against these they ran a lexicon of 108 cultural markers: 52 Indian, 32 Singaporean and 24 Nigerian, words and phrases such as “kindly”, “do the needful”, “revert back”, “prepone”, “chope” and “Respected sir”, alongside Singaporean particles such as “lah”.
Then came the models. Five open-source instruction-tuned systems, Mistral-7B, Apertus-8B, DeepHat-7B, MiMo-7B and Qwen3-8B, were each asked to process the texts under three prompt conditions: a plain baseline instruction in the spirit of “make this more professional”, a neutral rewrite request, and an explicit instruction to preserve cultural voice. Every text through every model under every prompt produced 22,350 outputs, and within them 624 distinct marker instances to track.
The headline metric, the Identity Erasure Rate, is a ratio: cultural markers present in the original minus cultural markers present in the output, divided by what was there before. A rate of zero means the voice survived. A rate of one means it was completely erased. Averaged across everything the models produced, the rate was 0.1026. But averages conceal as much as they reveal here, and the honest way to read this study is to hold two numbers at once. Per text, the average loss is around a tenth. Per marker, the loss is far worse: under the plain baseline prompt, 1,657 of the 3,120 possible erasure events across five models actually happened, an instance-level rate of 53.1 per cent. Any given “kindly” that goes in is slightly more likely to vanish than to survive.
There is a reason the per-text number looks tamer than the per-marker number. The marked texts carry, on average, fewer than half a marker each, 0.42 instances per text, so most texts lose nothing simply because most texts had almost nothing to lose. The few markers that were there, though, were disproportionately deleted. Both statistics are true simultaneously, and the second is the one that describes a person's actual experience of opening a rewritten email and finding their own warmth gone.
What the Models Delete First #
Not all markers suffer equally, and the pattern in what gets erased is the paper's most human finding. The researchers sorted markers into three categories: pragmatic, syntactic and lexical. Pragmatic markers are the machinery of relationship, politeness conventions such as “kindly”, honorifics such as “Respected sir” or “my dear”, the deferential softeners that in Indian, Singaporean and Nigerian English do real social work. Syntactic markers are constructions such as “discuss about” or the reduplicated patterns that mark South Asian English. Lexical markers are the local vocabulary itself, words such as “prepone” or “chope”.
Under the baseline prompt, pragmatic markers were erased 71.5 per cent of the time, syntactic markers 56.3 per cent of the time, and lexical markers 37.1 per cent of the time. The politeness conventions, the stuff that carries how the writer feels about the reader, were 1.9 times more vulnerable than the local vocabulary. A model asked to tidy an email will let an odd word stand more often than it will let an odd courtesy stand.
That ordering deserves attention, because it inverts the way language is usually policed. When people complain about someone else's English, they complain about words: the “delve”, the “do the needful”, the unfamiliar noun. Words are visible. Pragmatics are not. When a manager in Chennai writes “Kindly do the needful at the earliest”, the courtesy is not decorative noise that a competent editor would strike out. It is the sentence performing a relationship, signalling deference, warmth and a shared understanding of how business gets done. A model that deletes it has not cleaned the sentence up. It has changed what the sentence is for. The researchers put the risk plainly: users seeking clarity may unwittingly receive what they call cultural whitewashing.
The Singaporean particle “lah” illustrates the other half of the problem. Particles like it carry stance and solidarity: affiliation, casualness, the sense that speaker and reader are on the same side of a shared language. Removing one does not damage any proposition in the sentence. It damages the reason the sentence felt like it came from a person. And precisely because no propositional damage occurs, no quality metric catches it. The email reads as fine. It reads as someone else.
The Paradox That Makes the Erasure Invisible #
The Semantic Preservation Paradox is the study's structural insight, and it explains why this harm has been so easy to overlook. The team measured semantic similarity between originals and outputs using sentence embeddings, and the mean score was 0.7482, high enough that every standard quality check would pass. Meaning in, meaning out. If your evaluation suite scores rewrites on clarity, on grammar, on fidelity to content, the rewritten email wins every time. The erasure is invisible to every instrument the industry uses to decide whether a writing tool is good.
There is a deeper caveat that the researchers themselves flag. Embedding models carry the biases of the data they were trained on, which skews towards Western norms of what a sentence should look like. The team checked its semantic metric against a multilingual embedding model and found a cross-correlation of 0.94, which is reassuring but not conclusive. A similarity score says two sentences resemble each other. It does not say whether the second sentence still carries the social meaning the first one carried, or whether a reader in Lagos or Colombo would still recognise the writer as one of their own. Form survived. Function was not measured, and in some cases could not be, because the instruments themselves cannot see it.
This is the uncomfortable shape of the problem. The tools are not malfunctioning. By the criteria they were built to, they are working beautifully, which is exactly what makes cultural ghosting a design failure rather than a bug. Nobody wrote a line of code that says delete the kindness. The deletion is an emergent property of what “professional” means to a model trained on the writing of the global technology industry's default culture, and that default has a geography.
A Word Became an Accusation #
The study did not appear in a vacuum. The politics of whose English counts as real English flared into public view in April 2024, when Paul Graham, the co-founder of Y Combinator, tweeted about a cold email he had received: “Someone sent me a cold email proposing a novel project. Then I noticed it used the word 'delve'.” The word, he said, was a sign that text was written by ChatGPT. The tweet drew millions of views and set off an argument that has never really ended.
The response from the Global South was immediate and bitter, and it was not about a word. It was about who gets to decide which words are legitimate. Nigerians pointed out that “delve” is ordinary formal English across West Africa, in professional emails and academic writing, as it is across much of Asia. The novelist and satirist Elnathan John captured the irony in the exchange that followed: people force-fed colonial languages, who mastered them, were now being told that no one uses words like “delve” in real life. The journalist Ruona Meyer described her brother being called out for using ChatGPT in postgraduate assignments, including one instance where he wrote a paragraph by hand in class, in front of a lecturer, and “the AI detector still said it was likely ChatGPT”. A student named Farida Adamu wrote that she “had to water down the writing in my thesis so the AI detector would stop flagging my work”, adding, “I swear I am not making this up”. The therapist Dedoyin Ajayi said someone assumed she had used ChatGPT purely because she used an unfamiliar word. The filmmaker Editi Effiòng argued that Nigerian English “has since evolved far above what the British handed over”. Nigeria's communications minister, Bosun Tijani, used the moment to ask whether the industry could “delve” into possible bias and lack of inclusion in AI datasets.
These anecdotes describe the front end of a measurable problem. In 2023, before the delve affair, a Stanford team led by Weixin Liang and James Zou tested seven widely used GPT detectors on essays written by non-native English speakers, mostly Chinese, preparing for the TOEFL test, and on essays by American eighth-graders. The detectors were near-perfect on the American essays. On the TOEFL essays they misclassified more than half as machine-written, an average false positive rate of 61.3 per cent. All seven detectors unanimously flagged 19.8 per cent of the non-native essays, and at least one flagged 97.8 per cent of them. The mechanism is structural: detectors reward low-perplexity text, and people writing in a second language tend towards more formal, more predictable phrasing. The formality that marks a Global South professional as educated marks them, to a detector, as a machine. When the Stanford researchers passed the same TOEFL essays through ChatGPT with a prompt to make them sound more native, the false positive rate collapsed to 11.6 per cent. The machine can launder you out of suspicion, for a price it does not advertise.
Put the two findings side by side and the pincer closes. Write in your own English, and the detector calls you a fraud. Send your writing through a model for polishing, and the model deletes your English. Both routes terminate at the same destination: a standardised voice that reads as nobody in particular.
The Theory That AI English Is Borrowed English Does Not Survive Contact #
The delve affair produced a seductive counter-narrative, and it deserves an honest burial. In April 2024, the Guardian's technology newsletter TechScape, written by Alex Hern, proposed that ChatGPT's fondness for formal words like “delve” might come from its training pipeline: the reinforcement learning work that aligns these models is heavily outsourced to annotators in Nigeria, Kenya and elsewhere, so perhaps the models had absorbed the formal English of the workers who taught them manners. If true, it would have been a story of quiet poetic justice, the Global South's English becoming the machine's default.
The evidence does not support it. Simon Willison, writing at the time, noted the theory was speculative. And a systematic study published at COLING 2025 by Tom S. Juzek and Zina B. Ward of Florida State University tested it properly. They analysed 26.7 million PubMed abstracts, tracking 21 “focal words” whose frequency spiked after ChatGPT's release, “delves” alone rose roughly 6,700 per cent between 2020 and 2024. They then checked every plausible source: the training corpora, the International Corpus of English, the possibility that Nigerian or Indian English usage explained the overrepresentation. They found no evidence for any of it. What they did find was evidence consistent with reinforcement learning from human feedback: the aligned chat versions of models were measurably less surprised by AI-generated text than their base versions were. The machines did not learn “delve” from Lagos. They learned the preferences of their alignment processes, and then the world started mocking Nigerians for sounding like the machines.
Kingsley Ugwuanyi of the University of Nigeria, Nsukka, analysing the affair in a 2026 paper on AI and standard language ideology, adds a corpus detail that twists the knife: in global web data, “delve” is most frequent not in Nigerian English but in the Englishes of Malaysia, Hong Kong and India. The word that got a Nigerian professional baselessly accused of machine authorship was, if anything, an Asian English word. The direction of travel, Ugwuanyi argues, is not that the periphery's English is conquering the machine. It is that AI systems act as ideological actors, reproducing at every level of the pipeline, from training data to evaluation benchmarks to public commentary, a standard language ideology that privileges Inner Circle norms, the English of Britain and the United States, and treats everything else as deviation. Non-dominant Englishes get the worst of both discourses: their formal registers are now conflated with machine speech, while their local vocabularies are marked as errors to be corrected.
Reading Dialect as Defect #
Cultural ghosting is the polite face of a problem with an ugly one, and the uglier one has been measured too. In August 2024, a team led by Valentin Hofmann of the Allen Institute for AI, with Pratyusha Ria Kalluri of Stanford, Dan Jurafsky of Stanford and Sharese King of the University of Chicago, published a study in Nature that adapted a sociolinguistic technique called matched guise probing. The idea, borrowed from decades of dialect research, is elegant: present a model with texts in African American English and texts in Standardised American English, never mention race, and measure what the model infers about the writers.
The results were worse than the stereotype literature. Models associated African American English speakers with negative traits, including “ignorant”, “lazy” and “stupid”, at levels exceeding the most negative human stereotypes ever experimentally recorded, closer to attitudes measured in the 1930s than to anything a modern survey would find. When asked directly about African Americans, the same models, especially those refined with human feedback, produced admirable traits. The overt face had been scrubbed clean, while the covert associations underneath had hardened. In downstream tests the models assigned African American English speakers less prestigious jobs, were more likely to convict them in a hypothetical murder trial, 68.7 per cent versus 62.1 per cent, and were more likely to sentence them to death, 27.7 per cent versus 22.8 per cent. The larger and more carefully aligned the model, the more overtly polite it became about race and the more negative its dialect-linked associations grew.
Cultural ghosting sits on the same continuum. Both phenomena share a single root assumption: that one variety of English is the neutral substance of communication and every other variety is a deviation to be managed. In Hofmann's study the deviation is punished with suspicion. In the CHI study it is punished with deletion. The erasure is the gentler sentence, and in some ways the more corrosive one, because it arrives dressed as help.
One Variety to Rule the Inboxes #
The scale question is whether these deletions aggregate into something a civilisation would notice. The field of World Englishes, founded on the work of the linguist Braj Kachru and his model of three circles, an inner circle where English is the native language, an outer circle where it is an institutional inheritance of colonisation, and an expanding circle where it is a foreign language, has spent decades insisting that outer circle Englishes are not broken English but complete systems with their own grammars, pragmatics and literatures. Indian English, Singaporean English and Nigerian English are not trying and failing to be British English. They are doing different things with the same raw material, and their speakers number in the hundreds of millions. Ethnologue counts roughly 1.53 billion speakers of English worldwide, of whom about three-quarters, more than 1.1 billion, speak it as a second language. The native speakers of the standard that models default to are a minority of the English-speaking world.
What happens when that majority routes its writing through tools defaulting to the minority's variety? A 2026 study in Nature Human Behaviour by a University of Southern California team led by Zhivar Sourati and Morteza Dehghani offers a population-scale answer. Across three studies, seven datasets and more than 880,000 texts, they found that LLM rewriting reduced the variance of writing complexity by between 21 and 50 per cent, while meaning stayed intact, 87 per cent of original and rewritten pairs scoring above 0.95 on similarity. It is the Semantic Preservation Paradox again, scaled from sentences to entire communities: content preserved, diversity compressed. Classifiers trying to infer authors' personal traits, their personality, morality, age, political leaning, from their writing were around six percentage points less accurate after rewriting, because the markers those traits lived in had been smoothed away. The authors note the consequences reach into anywhere writing is used as evidence of a person: hiring, psychological assessment, cultural preservation.
This is what distinguishes cultural ghosting from ordinary language change. Languages have always borrowed, shifted and levelled, but they did it through the distributed, self-interested choices of their speakers. A politeness convention survives in Indian English because generations of Indian writers found it useful and kept reaching for it. When the reaching is delegated, when “kindly” is deleted by a tool before a human reader ever sees the sentence, the feedback loop that keeps a variety alive is cut. The writer's choice never enters the record. The linguist Christian Mair has described AI's effect on English as a standardisation paradox: the same systems that homogenise written English also expose more people to more of its varieties than ever before. The Ugwuanyi analysis is less optimistic about which force wins in a workplace setting, because exposure is passive and standardisation is wired into the defaults of the tools people are required to use.
The Fix Turns Out to Be a Sentence #
The briefest grounds for hope in this literature is how cheap the mitigation turns out to be. In the CHI study, the third prompt condition simply told the models, in effect, to preserve the cultural and regional voice of the writer while improving clarity. The Identity Erasure Rate fell from 11.56 per cent under the baseline prompt to 8.21 per cent, a 29 per cent reduction. The neutral prompt condition achieved a 19 per cent reduction with blander instructions. Nothing was paid in clarity. Semantic similarity did not fall, it rose slightly, from 0.7421 to 0.7548, because outputs that stay closer to the original text drift less in meaning. The researchers' own conclusion is blunt: the result refutes the assumed trade-off between authenticity and clarity. The trade-off was never real. Nobody had checked.
A parallel study reaches the same destination by a different road. At the 2026 C3NLP workshop, Soumedhik Bharati, Shibam Mandal, Swarup Kr Ghosh and Sayani Mondal quantified what South Asian English loses in Llama 3.3 70B under the two most common writing-assistant operations. Standard grammar correction retained only 26 per cent of dialect markers; “formalisation”, the make-this-professional operation, retained 14 per cent. In the majority of lexical cases the model did not merely delete the local word, it substituted a Standard American paraphrase, a process the authors call Americanisation, which occurred in 56.2 per cent of correction cases and 59.4 per cent of formalisation cases. The fix, again, was a sentence. A dialect-aware prompt raised marker retention to 92 per cent and cut Americanisation to 6.2 per cent.
The CHI team also tested two technical interventions beyond prompting. Constrained decoding, a mechanism that penalises the model for dropping listed markers during generation, reduced the Identity Erasure Rate by 47 per cent while keeping semantic similarity within 2 per cent of baseline, though only on a subset of 200 texts. Contrastive reranking, which scores candidate rewrites for marker retention and picks the best, reduced erasure by 31 per cent. The engineering exists. The prompting result requires, literally, one sentence of instruction.
The closing line of the CHI paper is worth quoting exactly, because it reframes the industry's obligation: “Cultural ghosting is not inevitable. It is a design choice we can unmake.” That framing shifts the question from whether models can respect non-standard English to whether their makers will make respect the default. The study's results make the answer awkward for the vendors. A preservation prompt improves the output by the field's own quality metrics. The technology to deploy it is a sentence. What is missing is the decision to treat a billion non-standard writers' voices as worth one.
There is a catch, and it is a policy catch rather than a technical one. Every mitigation demonstrated so far is opt-in: the preservation prompt must be requested by the user, the dialect-aware instruction must be typed by someone who already knows it exists and already believes their English needs defending. The person most likely to be ghosted is the person least likely to know the term for it, and the least equipped to negotiate with a tool about their own identity. An identity-preserving feature that users must discover and activate is, for most of the affected population, indistinguishable from no feature at all. Defaults are the whole game.
What the Study Cannot Yet Prove #
The honest reading of the CHI paper requires sitting with what it does not establish, and the authors are unusually direct about this. The 108-marker lexicon is a finite subset, a sampling of the surfaces of three varieties that cannot capture their full internal diversity, let alone the varieties it does not include. There was no formal validation with human speakers of Indian, Singaporean or Nigerian English. The team used proxy validation instead: an instruction-tuned model judged whether outputs preserved markers, agreeing with the automatic metrics 89 per cent of the time, and the marker detection method was checked against existing annotated corpora with 91 per cent alignment. Those are reasonable proxies. They are not testimony from the people whose voices are at stake.
The study also covers only open-source models, a limitation the authors attribute to funding constraints, and the proprietary systems that actually dominate professional writing, ChatGPT, Gemini, Claude, were not tested. The authors hypothesise similar behaviour, given that all these systems are trained towards similar standards of “professional” English, but that remains a hypothesis. The Nigerian English sample is small, 89 texts against 601 Indian. No per-variety results are published, so the study cannot yet say whether Singaporean markers fare better than Indian ones. The statistical effect of the preservation prompt is real but modest in size, a small effect on the standard scale, and the strongest technical result, constrained decoding, was measured on 200 texts. And the semantic metric that underpins the paradox itself leans on embeddings that encode Western norms, a circularity the team acknowledges and partially checked.
None of this collapses the finding, because the finding does not stand alone. An independent team found 14 per cent retention and majority Americanisation for South Asian English under formalisation prompts. The Hofmann group found dialect triggering measurable penalties in judgement. The USC team found diversity compression across 880,000 texts. Four separate methodologies, four separate literatures, one convergent direction: models treat the one billion-plus speakers of outer circle Englishes as errors awaiting correction, at the sentence level, at the judgement level and at the population level. What the CHI study adds is precision, a name, a rate, a paradox and a demonstrated fix. What it waits for is replication with human raters, proprietary models and per-variety breakdowns.
Clarity Was Never the Reason #
So back to the question a busy professional in Lagos, Delhi or Singapore is entitled to ask. If the tools erase my voice, and the fix costs a sentence, why is the sentence not in the product? The research suggests the answer is not technical. The evidence that preservation degrades clarity does not exist, and the CHI study's own numbers point the other way. What exists instead is a definition of quality that was never neutral to begin with, one in which “clear” has quietly meant “clear the way we write”, for so long that the equation feels like common sense.
The stakes are not sentimental. A variety of English stays alive the way any living thing does, through use that reproduces itself, and the pragmatic markers the models delete first are precisely the ones that carry the variety's social life. “Kindly” in a Chennai email and “lah” in a Singaporean one are not obstacles between two minds. They are part of the mind being expressed, evidence of where the writer stands and what they assume about the reader. Strip them out and the writing becomes more legible to a benchmark in California and less legible, as writing, to everyone who shares the writer's world. Multiply the strip-out across every inbox, every status update, every customer reply that passes through a polishing model, and the process starts to look less like an editing convenience and more like the largest standardisation programme ever run, conducted without a vote, a policy or a signature.
The optimistic reading is that this is the rare AI problem with a cheap solution and a measurable win. One prompt condition, one sentence of instruction, and erasure falls by nearly a third while semantic quality edges up. One generation constraint, and it falls by half. A vendor genuinely committed to the global character of its user base could ship identity preservation as a default in a release cycle, not a research programme. The pessimistic reading is that defaults do not change without pressure, and the people best placed to apply pressure are the people least likely to be believed when they complain, in any English, that the machine has started making them sound like somebody else.
References and Sources #
- Navneet, Satyam Kumar; Chandra, Joydeep; Zhang, Yong. “When AI Writes, Whose Voice Remains? Quantifying Cultural Marker Erasure Across World English Varieties in Large Language Models.” Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems. DOI: 10.1145/3772363.3799085. https://dl.acm.org/doi/10.1145/3772363.3799085 and full text athttps://arxiv.org/html/2602.22145
- Ethnologue. “What is the most spoken language?” https://www.ethnologue.com/insights/most-spoken-language/
- Jaiyeola, Temitayo. “Online uproar over Nigerian English flagged as ChatGPT-ish.” BusinessDay Nigeria, 10 April 2024. https://businessday.ng/technology/article/online-uproar-over-nigerian-english-flagged-as-chatgpt-ish/
- Liang, Weixin; Yuksekgonul, Mert; Mao, Yining; Wu, Eric; Zou, James. “GPT detectors are biased against non-native English writers.” Patterns 4(7), 100779, July 2023. DOI: 10.1016/j.patter.2023.100779. https://www.cell.com/patterns/fulltext/S2666-3899(23)00130-7
- Hern, Alex. “TechScape: How cheap, outsourced labour in Africa is shaping AI English.” The Guardian, 16 April 2024. https://www.theguardian.com/technology/2024/apr/16/techscape-ai-gadgest-humane-ai-pin-chatgpt
- Willison, Simon. “How cheap, outsourced labour in Africa is shaping AI English.” simonwillison.net, 18 April 2024. https://simonwillison.net/2024/Apr/18/delve/
- Juzek, Tom S.; Ward, Zina B. “Why Does ChatGPT 'Delve' So Much? Exploring the Sources of Lexical Overrepresentation in Large Language Models.” Proceedings of the 31st International Conference on Computational Linguistics (COLING 2025), 6397-6411. https://aclanthology.org/2025.coling-main.426/
- Ugwuanyi, Kingsley. “AI systems and the reproduction of (standard) language ideologies in World Englishes.” Linguistics Vanguard, 2026. DOI: 10.1515/lingvan-2025-0245. https://arxiv.org/abs/2607.28528
- Hofmann, Valentin; Kalluri, Pratyusha Ria; Jurafsky, Dan; King, Sharese. “AI generates covertly racist decisions about people based on their dialect.” Nature 633, 147-154, 28 August 2024. DOI: 10.1038/s41586-024-07856-5. https://www.nature.com/articles/s41586-024-07856-5
- Kachru, Braj B. “Standards, Codification and Sociolinguistic Realism: The English Language in the Outer Circle.” In Quirk, Randolph; Widdowson, Henry G. (eds.), English in the World: Teaching and Learning the Language and Literatures. Cambridge University Press, 1985.
- Sourati, Zhivar; Karimi-Malekabadi, Farzan; Ozcan, Meltem; McDaniel, Colin; Ziabari, Alireza; Trager, Jackson; Tak, Ala N.; Chen, Meng; Morstatter, Fred; Dehghani, Morteza. “The shrinking landscape of linguistic diversity in the age of large language models.” Nature Human Behaviour 10, 1925-1942, 24 August 2026. DOI: 10.1038/s41562-026-02550-0. https://www.nature.com/articles/s41562-026-02550-0
- Bharati, Soumedhik; Mandal, Shibam; Ghosh, Swarup Kr; Mondal, Sayani. “The American Palimpsest: Quantifying South Asian English Dialect Erasure in LLMs.” Proceedings of the 4th Workshop on Cross-Cultural Considerations in NLP (C3NLP 2026), 108-118. DOI: 10.18653/v1/2026.c3nlp-1.8. https://aclanthology.org/2026.c3nlp-1.8/
Tim Green UK-based Systems Theorist & Independent Technology Writer
Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.
His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.
**ORCID:** [0009-0002-0156-9795](https://orcid.org/0009-0002-0156-9795)
**Email:** [tim@smarterarticles.co.uk](mailto:tim@smarterarticles.co.uk)
Listen to the free weekly [SmarterArticles Podcast](https://smarterarticles.fm)