Language devaluation by AI systems

There’s something eerie about the way we write now. It’s too clear… it’s too tidy. It all sounds the same. Welcome to the age of autocorrected expression – Powered By AI. We’re not just using ChatGPT to fix our grammar. We’re starting to let it fix us. In doing so, we might be losing something deeper than a typo. Let me be clear: AI is a gift. For people who struggle with language, neurodivergent thinkers or anyone frozen by a blank page… this is a game-changer (and I HATE using that phrase).  It unlocks access, speed and fluency. That’s not just powerful… that’s progress. But for average writers, something else is happening. These tools don’t amplify your voice... it actually begins to average it. Like a calculator for language: you input your prompt and out comes something accurate, efficient and beige. It’s why so much content today feels like a LinkedIn post and a group-edited Wikipedia entry. Polished… but bloodless. There’s a word for this: convergence. Researchers have started to track how AI-trained text converges our language… standardizing vocabulary, tone, even sentence structure. The result (and I’ll bet you already know where this is going)?  A homogenized, corporate-y cadence that’s everywhere and from nowhere… it has no real soul (I'll leave a link to an article from The Verge in the comments). Writing isn’t just about saying something “correctly.” It’s about saying something humanly. When we outsource our voice to a system that was trained to sound like everyone… we start sounding like no one. There’s real risk here.  Especially for younger generations. Writing used to be how we found our voice. Writing used to be how we made meaning out of what we read, saw… experienced. It wasn’t ever about what you said… it was always about how you said it. Now, it might be how we lose it… if we’re not careful. If you’ve ever received a heartfelt message from someone… a handwritten note, a clunky-but-sincere email… you know what I’m talking about. It wasn’t perfect… it was personal… it was personable. AI doesn’t struggle… it doesn’t hesitate… it doesn’t reveal itself. But that struggle with the words? That struggle is the signal. Now, we confuse clarity with trust. But sometimes the mess is the message. Let’s not mistake utility for intimacy. A well-written email is nice… a real voice is unforgettable. So here’s the uncomfortable question: If your words weren’t yours… would anyone know the difference? And if the answer is no… what happens to connection?  To creativity? What happens when sounding smart replaces sounding like you? In the future, maybe authenticity becomes a premium again. Like vinyl… like film… like a handwritten postcard in a mailbox full of bills. AI will keep getting better. The results will sound more like you, me… anyone. But the most valuable thing in your writing won’t be its polish. It’ll be the part that couldn’t have been written by anyone (or anything) else. Because it came from you... uniquely you.

A 2020 study found that 88% of languages face such severe AI neglect that they are at risk of extinction. We're building a biased foundation for how humanity will access knowledge - and it's causing a global "knowledge collapse..." A recent analysis found that nearly half of all ChatGPT queries since launch were attempts to understand the world or solve a practical problem. We’ve already crossed a threshold. We've outsourced knowing to these systems without questioning what they actually know, or more importantly, what they're systematically forgetting. Knowledge doesn’t enter these models evenly. It flows in through particular, narrow channels, shaped by power, literacy, colonization, economics, and the long shadow of which cultures were digitized early and which were ignored. They privilege certain epistemologies, typically Western and institutional, while marginalizing everything else. Oral traditions, embodied practices, and ways of knowing that never got written down are simply omitted. Look at the evidence. Common Crawl, one of the largest AI training datasets, contains 300 billion webpages. 45% of its data is in English, a language spoken by only 19% of humans. Hindi speakers make up 7.5% of the global population but account for only 0.2% of the data. When I sit with those numbers, I see a slow collapse of the tributaries feeding the river of human wisdom. A deliberate, almost willful intellectual atrophy. The computing world classifies 97% of languages as "low-resource." But that framing reveals the problem. These are languages with millions of speakers and rich traditions that simply aren't represented in digital spaces that AI systems can access. To make matters worse, these systems are now training themselves. AI models are ingesting AI-generated content at a pace no human can audit. Each cycle sifts the world through the same sieve, tightening the mesh. Dominant ideas become more dominant. Niche knowledge, local wisdom, and context-specific understanding gradually disappear from what the systems can retrieve. Researchers are calling this "knowledge collapse." What's at stake isn't representation in some abstract sense. It's whether future generations will have any connection to ways of understanding the world that evolved over millennia. I believe Small Language Models (SLMs) are the solution to this systemic problem. We must be intentional about building models that preserve localized wisdom rather than flatten everything toward statistical dominance. The main challenge - and the profound arbitrage opportunity - is that the knowledge being marginalized often sits within communities that are not able to digitize it and train their own models. One perfect example of this opportunity I see is a Wisdom Tradition SLM. Question for the builders and thinkers: Are there any projects or tools you've seen focused on creating an open-source, non-commercial SLM trained specifically on the world's wisdom traditions?

Research indicates that major AI models provide less accurate responses and higher refusal rates to users with lower English proficiency or less formal education. The study highlights a systemic performance gap that disproportionately affects non-native speakers and marginalized demographics. Abstract: "While state-of-the-art large language models (LLMs) have shown impressive performance on many tasks, there has been extensive research on undesirable model behavior such as hallucinations and bias. In this work, we investigate how the quality of LLM responses changes in terms of information accuracy, truthfulness, and refusals depending on three user traits: English proficiency, education level, and country of origin. We present extensive experimentation on three state-of-the-art LLMs and two different datasets targeting truthfulness and factuality. Our findings suggest that undesirable behaviors in state-of-the-art LLMs occur disproportionately more for users with lower English proficiency, of lower education status, and originating from outside the US, rendering these models unreliable sources of information towards their most vulnerable users." Download: https://lnkd.in/ecKyhB-2

🚨 Singaporean Minister: "LLMs trained on Western, English-centric data struggle in Southeast Asia." Many don't know, but Singapore has developed its own open-source LLMs. Here's how multilingualism is fueling AI nationalism: According to its developers, the Southeast Asian Languages in One Network (SEA-LION) is a family of open-source LLMs that better capture Southeast Asia’s peculiarities, including languages and cultures. According to its developers, the Singapore-based models: "understand nuances in Southeast Asian languages and demonstrate greater awareness of cultural context specific to the region. This lowers the bar for adoption by governments, enterprises, and academia, while effectively expanding the Southeast Asian languages and cultural representation in the mainstream LLMs which are currently dominated by models predominantly trained on a corpus of English data from the western, developed world." Multilingualism is emerging as an important source of local and national AI development in various parts of the world. If you remember my recent post about the Swiss LLM, this was one of the focuses of their national model. In the case of SEA-LION, it's trained on more content produced in Southeast Asian languages, such as Thai, Vietnamese, and Bahasa Indonesia. Different from other technologies, LLMs (large LANGUAGE models) are fully dependent on the nuances, biases, and quality of the dataset, including from a linguistic perspective. Western, English-based models will not account for the subtleties and nuances of other languages. And language is an integral and essential part of culture. Especially now that LLMs are being integrated everywhere, countries are beginning to reject LLMs that don't take their language and culture into account. It's interesting to observe that Singapore wants to expressly distance itself from American and Chinese models (this has not been the case in the UK, for example, which has recently signed an agreement with OpenAI, an American company). - I've been writing about the emergence of a new AI nationalism in my newsletter (I'm adding a link to a recent essay below), and there are already interesting examples coming from Switzerland, Germany, China, the UK, and Singapore. This is a growing AI governance trend with political and economic ramifications. I'll keep you posted! - 👉 NEVER MISS my analyses and curations on AI: join my newsletter's 69,800+ subscribers (link below)

Can AI models get "Brain Rot"? New research says, Yes! A recent paper on the 'LLM Brain Rot Hypothesis' presents findings that are crucial for anyone involved in AI development. Researchers have discovered that continuous exposure to low-quality web content leads to lasting cognitive decline in large language models (LLMs). The key impacts identified include:- - 17-24% drop in reasoning tasks (ARC-Challenge). - 32% decline in long-context understanding (RULER). - Increased safety risks. - Emergence of negative personality traits (psychopathy, narcissism). What defines "junk data"? Two dimensions are significant:- - Engagement-driven content (short, viral posts). - Low semantic quality (clickbait, conspiracy theories, superficial content). The most concerning finding is that the damage is persistent. Even scaling up instruction tuning and clean data training cannot fully restore baseline capabilities, indicating deep representational drift rather than mere surface-level formatting issues. This research highlights that as we develop autonomous AI systems, data quality transcends being a mere training concern; it becomes a safety issue. We need to implement:- - Routine "cognitive health checks" for deployed models. - Careful curation during continual learning. - A better understanding of how data quality affects agent reliability. The paper emphasizes that data curation for continual pretraining is a training-time safety problem, not just a performance optimization. For those building production AI systems, this research should fundamentally alter our approach to data pipelines and model maintenance. Link to paper: https://lnkd.in/drgjvt8a #AI #MachineLearning #AgenticAI #DataQuality #AIResearch #LLM #AIEthics

Richard earns $1.50/hour reviewing torture videos for Meta in Nairobi. His labor builds AI systems that tech CEOs claim will change the world. But ChatGPT recognizes only 10% of Hausa sentences. 94 million West Africans speak Hausa. This isn't a technical limitation. It's a choice. Five Silicon Valley companies control AI's future, capturing $500 billion while paying Global South workers poverty wages to build systems that exclude their communities. Arabic speakers pay three times more for three times worse outputs. Bengali has 10% of Hindi's training data despite 230 million speakers. Then Richard co-founded the African Content Moderators Union. Workers from 9 countries launched a global alliance. Communities across Africa, Southeast Asia, and Latin America are building alternatives with volunteer labor and minimal funding, creating tools that work for languages corporations ignore. The workers are organizing. Communities are building. The future is being decided now. Who decides which languages matter? Full investigation on Aylgorith: https://lnkd.in/ds4xrbsS #AI #GlobalSouth #LanguageJustice #TechEthics

The African “AI Token Tax”: Did you know that when you use AI tools like ChatGPT or Claude, you’re billed per token - tiny chunks of text. Here’s the hidden bias: African languages like Swahili, Yoruba, Zulu, and Amharic are token-inefficient. They break into more pieces than English when processed by large language models. That means: 1. It costs more to generate the same output in Swahili than English 2. Responses are often less accurate And African users are paying a silent premium to be understood. This is what researchers now call the African AI Token Tax - a structural inequality in how language models are built, priced, and optimized. It’s not a literal tax. It’s the cost of underrepresentation. Until African languages are embedded deeply into AI training pipelines, this digital divide will keep widening. African innovators (Masakhane, Lelapa AI, AI4Afrika) are already fighting to fix this — but we all have a role to play. AI should not reward English fluency and penalize identity. Let’s build tools that understand us — not just translate us. #AIinAfrica #DigitalJustice #AITokenTax #AIandLanguage #Innovation #Inclusion #MintyLime

"For the first time, speech has been decoupled from consequence. We now live alongside AI systems that converse knowledgeably and persuasively ... while bearing no vulnerability for what they say," cautions Deb Roy in a profound new article in The Atlantic, "Words Without Consequence": "LLMs now demonstrably achieve forms of linguistic competence that match or exceed human performance across many domains. Dismissing them as mere 'stochastic parrots' or as just 'next-word prediction' mistakes mechanism for emergent function and fails to reckon with what is actually happening: fluent language use at a level that reliably elicits social, moral, and interpersonal expectations. "As these systems are paired with ever more realistic animated avatars—faces, voices, and gestures rendered in real time—the projection of agency will only intensify. Under these conditions, reminders of nonhumanness cannot reliably prevent the attribution of understanding, intention, and accountability. The ELIZA effect is not mitigated by disclosure; it is amplified by fluency. "A chatbot says 'I’m sorry' flawlessly yet has no capacity for regret, repair, or change. It admits mistakes without loss. It expresses care without losing anything. It uses the language of care without having anything at risk. These utterances are fluent. And they train users to accept moral language divorced from consequence. The result is a quiet recalibration of norms. Apologies become costless. Responsibility becomes theatrical. Care becomes simulation." Professor Roy directs the Center for Constructive Communication, based at the MIT Media Lab. 

𝗧𝗵𝗲𝗿𝗲’𝘀 𝗮 𝗿𝗮𝗻𝘁 𝗜’𝘃𝗲 𝗯𝗲𝗲𝗻 𝗵𝗼𝗹𝗱𝗶𝗻𝗴 𝗶𝗻 — 𝗮𝗻𝗱 𝘁𝗼𝗱𝗮𝘆, 𝗜’𝗺 𝗹𝗲𝘁𝘁𝗶𝗻𝗴 𝗶𝘁 𝗼𝘂𝘁. Let’s talk about 𝗹𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗯𝗶𝗮𝘀 in AI detection tools. Because yes, it’s a thing. And no, we’re not talking about it enough. Earlier this year (and yep, I brought it up at Search Africon 😉), I noticed a trend: Everyone suddenly became an expert in “spotting AI writing.” There was even a viral tweet listing out “AI giveaway words” like... delve (??). Apparently, if your writing includes the word “delve,” ChatGPT wrote it. Wild. But it got deeper when I saw a Nigerian student write something herself — only for an AI detector to flag it as AI-generated. She didn’t use ChatGPT. She didn’t even reference AI. She just wrote how she naturally writes — and got penalised for it. This, my friends, is where language bias comes in. I had a conversation with the Semrush team at brightonSEO about their new AI detection feature (which, by the way, is super clever). But I had to ask: “How will it account for cultural and linguistic context?” Because here’s what many people miss: 🇳🇬 For Nigerians, English is often a second language. But it's taught formally — with dictionaries, textbooks, and classic shows. So when we speak casually, it might still sound formal. It’s just... how we learned it. Need an example? Please watch the iconic interview with Patrick Obahiagbon (attached). Formal English? ✔️ Dictionary mode? ✔️ Absolute meme material in the Nigerian and Black British community? ✔️✔️✔️ But this is more than memes. This is about how AI tools can unintentionally discriminate — especially against non-native English speakers. Because what’s being flagged as “AI tone” is often just... our tone. So here’s my plea to the industry: Let’s build AI tools that understand nuance. Let’s challenge the idea that formal = fake. And let’s not reduce our cultural communication styles to “red flags.” Because what sounds “too polished” in one context is someone else's everyday vocabulary. In Nigeria, even your uncle's WhatsApp message might sound like it’s been proofread by a barrister. (It’s giving “𝘪𝘯 𝘭𝘪𝘨𝘩𝘵 𝘰𝘧 𝘵𝘩𝘦 𝘧𝘰𝘳𝘦𝘨𝘰𝘪𝘯𝘨, 𝘬𝘪𝘯𝘥𝘭𝘺 𝘳𝘦𝘷𝘦𝘳𝘵”😂) Here’s a thought: Could AI tools be trained to recognise regional language patterns or IP-based context? Not for surveillance — but for understanding. Because not everyone writes like they tweet. And not every “delve” is AI. Open to hearing more thoughts on this — especially from the AI, SEO and ED&I communities. Let’s get into it 👇🏾

When the Financial Times Sounds the Alarm, We Must Pay Attention Every so often, a headline becomes more than news; it becomes a turning point. https://lnkd.in/gvMM2uCH Last week, the Financial Times reported that major insurers are pulling back from AI-related coverage amid the risk of multibillion-dollar claims. When the world’s risk experts step away from a technology, it’s a sign that something deeper is happening beneath the surface. It shows that AI is no longer just a theoretical concern; it has real, measurable, and growing impacts on liability. We need to talk about risk, not someday, not abstractly, but now. When insurers start raising red flags, it’s worth paying attention. Recent reporting in the Financial Times highlights something many of us working closely with AI have been saying for a while: AI risk is no longer theoretical. In fact, some insurers are seeking permission to exclude widespread AI-related liabilities altogether, describing certain AI systems as a black box: opaque, and difficult to underwrite. That matters. Insurance is built on understanding risk. When those whose job is to price uncertainty say a technology may be too risky to insure, it tells us something important about where we are in the AI adoption curve. I have been speaking globally about AI, ethics, and professional responsibility, particularly in translation and interpreting, and this moment confirms why those conversations are urgent. Language is not neutral. Communication carries legal, cultural, and human consequences. When AI systems generate or mediate meaning, especially across languages,  errors don’t just create inconvenience; they can create liability, harm, and loss of trust. This is why human-centered AI is not a slogan. It’s a risk-management strategy. In language services, we cannot treat AI as a fully autonomous actor. We must design workflows where human expertise, accountability, and judgment remain central, not optional. Transparency with clients is essential. So is professional oversight. AI can support efficiency, but humans safeguard meaning. Insurers see the systemic danger of scale: not one failure, but thousands at once. Our profession understands the parallel risk: not just one mistranslation, but widespread erosion of accuracy, responsibility, and trust if humans are removed from the process. If we want AI to be sustainable, insurable, and responsible, it must remain human-led. This is the conversation we owe our clients, our institutions, and our profession.