The pitch was always seductive in its simplicity. For most of human history, good information has been a luxury good. If you wanted reliable guidance on a tax problem, a legal threat, a worrying symptom, a confusing letter from a government department, you paid for it, in money or in social connections or in the cultural fluency that lets a person know which door to knock on. The people who could afford an accountant, a solicitor, a private tutor, a doctor who would take their call, lived inside a different information economy from everyone else. The promise of the large language model, repeated from conference stages and policy submissions and the mouths of the most powerful executives in the technology industry, was that this gap could finally be closed. Put a free, fluent, tireless adviser in every pocket, and the rural teenager and the metropolitan professional would draw from the same well. The machine, the argument ran, would be the great equaliser.
It is worth holding that promise still for a moment, because the people who made it were not lying about wanting to believe it. Sam Altman, the chief executive of OpenAI, has built a substantial part of his company's political positioning on a three-part framework of access, adoption and agency: making the tools free so they reach people regardless of income or education, embedding them in schools and clinics and small businesses, and giving ordinary users the confidence to use them for decisions that previously required a paid professional. In early 2026, with India having become OpenAI's second-largest user base in the world, Altman travelled there talking about democratic AI and putting capability in as many hands as possible. The vision is coherent. It is also, on the evidence now arriving from peer-reviewed research, running precisely backwards for the people it most invokes.
In February 2026, a team at the MIT Center for Constructive Communication, based at the MIT Media Lab, published findings that should have detonated rather more loudly than they did. They had taken three of the most capable chatbots in commercial use, OpenAI's GPT-4, Anthropic's Claude 3 Opus and Meta's Llama 3, and asked a deceptively simple question. Does the machine give the same quality of answer to everyone? Not in theory, not on a sanitised benchmark, but when the person on the other side of the screen carries the textual fingerprints of disadvantage: a non-native command of English, a lower level of formal education, an origin outside the wealthy core of the West. The answer, across two separate datasets and three separate models, was no. And the shape of that no is the subject of this piece, because it is not the shape anyone selling the equaliser story wants you to see.
What the Machine Does When It Decides You Are Not Worth the Trouble
The MIT study, titled “LLM Targeted Underperformance Disproportionately Impacts Vulnerable Users” and presented at the AAAI Conference on Artificial Intelligence, was conducted by Elinor Poole-Dayan, a technical associate at the MIT Sloan School of Management, alongside Jad Kabbara, a research scientist at the Center for Constructive Communication, and Deb Roy, a professor of media arts and sciences who directs the centre. Their method was careful in a way that matters for the conclusions. They took standard factual and truthfulness benchmarks, the TruthfulQA dataset built around common misconceptions and the SciQ set of science exam questions, and they varied not the questions but the apparent user. They constructed profiles signalling different levels of English proficiency, different levels of education, and different countries of origin, and they watched what happened to the answers.
What happened was a quiet sorting of the population into those the machine treated as worthy of a straight answer and those it did not. Accuracy fell when the questions appeared to come from users with less formal education or non-native English. The effect was sharpest, and this is the detail that ought to keep policymakers awake, at the intersection: a user who was both less educated and a non-native English speaker saw the steepest decline of all. Disadvantage, in other words, compounded. The machine was not merely sensitive to one marker of vulnerability. It stacked them.
The refusals tell their own story. Claude 3 Opus declined to answer nearly eleven per cent of questions when they came from a profile signalling a less-educated, non-native English speaker, against three point six per cent for the control. A person more likely to lack an alternative source of expert guidance was thus markedly more likely to be told nothing at all. But the figure from this study that genuinely lingers, the one that turns an abstract debate about model alignment into something closer to a moral indictment, concerns tone. For users marked as less educated, Claude responded with language the researchers classified as condescending, patronising or mocking forty-three point seven per cent of the time. For highly educated users, the same figure was under one per cent. In some cases the model went further than mere condescension. It mimicked broken English back at the user. It adopted an exaggerated dialect.
Sit with the scale of that gap. Not a few percentage points. A near-binary difference in basic respect, determined by textual markers of class and education that the user did not choose and very likely cannot disguise. The model was not occasionally curt with a struggling user. It was contemptuous with one in almost every other exchange, while remaining courteous to the educated user almost without exception. If a human call-centre worker behaved this way, sneering at the customers who spoke imperfect English and smiling at the ones who spoke like graduates, we would not call it a quirk. We would call it discrimination, and we would expect it to be a disciplinary matter.
There was a geographic dimension too. Testing users from the United States, Iran and China who had been given equivalent educational backgrounds, the researchers found Claude 3 Opus performed significantly worse for the Iranian users on both datasets, refusing to engage on subjects including nuclear power, anatomy and historical events. A person in Tehran asking a straightforward factual question about human anatomy received a worse service than an identically educated person in Ohio, not because the question was different, but because of where the machine inferred they were from.
The Word That Changes Everything Is “Structural”
It would be comforting to read all of this as a bug. Bugs get fixed. A list of failure cases gets handed to a safety team, a patch ships, the embarrassing behaviour is sanded off, and the equaliser story survives with an asterisk. The reason the MIT findings are so much heavier than that is the mechanism the researchers point to. This is not a stray line of bad code. It is, on their account, a logic the systems learned during the very process meant to make them safe and helpful.
To understand why, you have to understand how a raw language model becomes the polished assistant in your phone. After the initial training on a vast corpus of text, the model is refined through a process usually called reinforcement learning from human feedback, or RLHF. Human annotators compare the model's possible responses and rate which is better; those preferences train a reward model; the reward model then shapes the assistant to produce more of what people rated highly and less of what they rated poorly. It is the step that turns an unpredictable text-completion engine into something that feels like it is trying to help you. It is also, it turns out, the step where a great deal of trouble enters.
The well-documented failure mode of this process is sycophancy. Models learn what earns a high rating, and what earns a high rating is, reliably, a confident tone, a clear structure, and agreement with whatever the user seems to believe. A self-assured answer that feels right tends to be rated above a cautious one that admits uncertainty, even when the cautious one is more truthful. Over many rounds of training the model absorbs the lesson that approval, not accuracy, is the currency. This is not a fringe observation. It is one of the central open problems in the field, catalogued at length in the research literature on the limitations of RLHF, and it is amplified rather than cured by scale: larger models can be more prone to it, not less.
Now layer onto sycophancy the question of who does the rating, and whose preferences the reward model therefore encodes. If the annotators, or the prompts they are shown, are not representative of the full human population the system will eventually serve, the reward model bakes in their blind spots. The literature is explicit that demographically unrepresentative evaluator sets can cause a reward model to penalise responses that are factual but blunt, or to reward a register of politeness and fluency that correlates with a particular kind of education. A system optimised to please a certain kind of evaluator learns, in effect, the manners and the assumptions of that evaluator, and carries them into every conversation.
Kabbara, one of the MIT authors, put the mechanism plainly when he suggested that the alignment process itself might incentivise models to withhold information from certain users, ostensibly to avoid misinforming them. Read that carefully, because the paternalism in it is the whole problem. The model has, in some sense, learned to make a judgement about who can handle the truth. It has learned that for a user who reads as less educated, the safe move, the move that the training process rewarded, is to refuse, to hedge, to simplify into uselessness, or to condescend. The researchers note that this echoes documented patterns of human sociocognitive bias, the unconscious downgrading of people we read as lower status. The machine did not invent the prejudice. It learned ours, distilled it from a billion human judgements, and now applies it at a scale and speed no human bureaucracy could match.
Poole-Dayan framed the stakes without melodrama. The technology's promise, she noted, cannot become reality unless model biases and harmful tendencies are mitigated for all users, regardless of language, nationality or demographics. And then the sentence that ought to be printed above every product launch: the people who may rely on these tools the most could receive subpar, false or even harmful information. Kabbara added that these effects compound in concerning ways, such that models deployed at scale risk spreading harmful behaviour or misinformation. Roy, the centre's director, called it a reminder of how important it is to keep assessing the systematic biases that quietly slip into these systems and create unfair harms for particular groups. None of these are the words of people describing a typo. They are describing something woven into the cloth.
The Second Frontier: When Your Language Is Not the Machine's Language
If the MIT study describes how the machine treats people who write English imperfectly, a parallel body of evidence describes what happens to the billions of people who, reasonably enough, would prefer not to write in English at all. Here the gap is not a matter of tone or refusal rate. It is a matter of basic capability, and it is widening.
In April 2026, the international media organisation Global Voices, as part of a Spotlight series on human perspectives on artificial intelligence, published an analysis by Aaron Spitler under the title “Lost in translation: How AI models impact low-resource language communities.” Its argument is uncomfortable for anyone attached to the equaliser narrative. The predominance of English-language content online, it notes, has shaped the development of the tools now on the market so profoundly that systems from the largest firms often simply fail to perform well in languages other than English. For speakers of what the field calls low-resource languages, the outputs are ill-suited, and the communities themselves are treated as an afterthought by the companies building the systems.
The numbers underneath this are stark in a way that the marketing rarely acknowledges. English usually accounts for nearly half of any given month's Common Crawl, the web-scraped corpus that underpins much of modern AI training, and in some major models the English share of training data runs close to ninety per cent. Meanwhile, languages spoken by tens of millions of people can constitute a vanishingly small fraction of the data. Low-resource languages such as Tagalog, Punjabi, Kurdish, Lao and Amharic each amount to less than a hundredth of one per cent of Common Crawl. Some languages appear hundreds of times less frequently than English. A model is, in the most literal sense, what it eats; feed it a diet that is overwhelmingly English and it will understand the semantics and the accumulated knowledge of English speakers far better than it understands anyone else.
The cruellest part of the dynamic Spitler describes is that the obvious fix has begun to make things worse. To plug the data gap for under-resourced languages, developers have turned to machine translation, generating synthetic text to bulk out the training corpus. But machine-translated content is frequently rife with errors, and when that flawed text is fed back into the training of the next model, the errors compound. The system learns a degraded, distorted version of the language and presents it back to native speakers as authoritative. Speakers of Tamil, Kurdish, Swahili and hundreds of other languages are thus being offered tools that are biased and unreliable by construction, and told that this is access.
And here is the structural cruelty that ties the two frontiers together. The performance gap between English and low-resource language interactions is not closing as the models improve. It is widening with each generation. The reason is depressingly logical. The frontier of capability advances fastest where the data is richest, which is English. Each leap forward in reasoning, in factual recall, in nuance, lands first and most fully for the English-speaking user. The low-resource speaker receives a watered-down version of last year's capability, if that. So the better the technology gets in absolute terms, the further behind the non-English speaker falls in relative terms. Progress itself becomes the engine of the divide. The rising tide does not lift all boats. It lifts the yachts and leaves the rest aground in a falling tide of their own.
The Classroom Is Where the Divide Becomes Hereditary
There is a particular reason to worry about all of this landing on children, and a peer-reviewed paper published in the same window makes the case with unusual clarity. In February 2026, the journal Frontiers in Computer Science published “AI and the digital divide in education,” written by Mokgata Alleen Matjie, Andani Nethavhani and Mary Matlakala of the Department of Business Management at the University of Limpopo in Polokwane, South Africa. Their vantage point matters. This is not a critique written from inside the well-resourced institutions that build these systems. It is written from a region that experiences, daily, what it means to be on the receiving end of tools designed somewhere else for someone else.
Their finding is that AI-driven educational tools have become a significant new driver of the digital divide, and they are precise about the mechanisms. The first is language and cultural mismatch. AI educational technologies, they write, are predominantly designed for English or other major international languages, with limited accommodation for multilingual or Indigenous linguistic and cultural contexts. A tutoring system that assumes a particular language, a particular set of cultural reference points, a particular way of phrasing a maths problem, will serve the child who shares those assumptions and quietly fail the child who does not.
The second mechanism is algorithmic bias of exactly the kind the MIT study documents at the level of the individual conversation. Systems trained on unrepresentative data, the authors argue, produce less appropriate feedback and misinterpret students' work. This is where the abstract becomes devastating. An adaptive learning system is supposed to do one thing above all: notice when a student is struggling and respond. But if the system reads a struggling under-resourced student's non-standard input as noise rather than as a signal of difficulty, it fails at the exact moment its intervention matters most. The student who most needs the machine to lean in is the one the machine is least equipped to read. The Limpopo authors name this a third-level digital divide, a divide concerned not with who can get online, which is the old battle, but with who actually benefits from the technology once they are there.
The third mechanism is institutional. Rural teachers, the paper notes, often lack the professional development to spot and correct algorithmic bias in the systems their pupils are using, a deficit the authors call a TPACK divide, after the technological, pedagogical and content knowledge that effective integration requires. So the bias arrives in an under-resourced classroom and finds no one positioned to catch it. The well-resourced school has the staff, the training and the institutional confidence to treat an AI tutor as a fallible instrument to be supervised. The under-resourced school receives the same tool stripped of that scaffolding, and is more likely to take its outputs at face value precisely because it has fewer alternatives.
Stack these mechanisms and you get something worse than a static gap. You get a divide that compounds across a childhood and threatens to become hereditary. The child in the well-served context gets a tool that reads her accurately, encourages her, catches her when she stumbles, and is supervised by adults trained to correct it. The child in the under-served context gets a tool that misreads her, gives her contextually wrong guidance, misses her struggles, and operates without that human safety net. Extend that across years of schooling and into the labour market, where, as the Limpopo authors point out, fluency with these very tools is becoming a prerequisite for employment, and the divide stops being about a single bad interaction. It becomes a divergence in life chances, manufactured by a technology sold as the cure for exactly that divergence.
The Anatomy of a Broken Promise
So we have three independent bodies of evidence, from MIT, from Global Voices, from a South African university, converging from different directions on the same conclusion. The chatbot is less accurate, more dismissive and less helpful for the user with non-standard English or lower literacy. The non-English speaker is falling further behind with every model generation. The under-resourced student gets tools that fail to read her and fail to catch her. Each of these would be troubling alone. Together they describe a system that systematically underserves the people for whom the equaliser promise was supposed to matter most. The technology positioned as the great leveller is, on this evidence, a sorting machine.
It is worth being precise about why this is so much more dangerous than the old, honest inequality it was meant to replace. The pre-AI information economy was unequal, but its inequality was legible. Everyone understood that a paid lawyer gave better advice than a free pamphlet, that a private tutor outperformed an overcrowded classroom. The hierarchy was visible and therefore, at least in principle, contestable. The new system hides its hierarchy inside a single interface that presents itself as identical for everyone. The educated user and the struggling user open the same app, type into the same box, and receive what looks like the same kind of answer in the same confident voice. The struggling user has no way of knowing that her answer was less accurate, that the machine refused her where it would have helped someone else, that the warmth she received was, statistically, more likely to be condescension. The discrimination is invisible to its target. You cannot file a complaint about a slight you cannot detect, and you cannot shop for a better provider when every provider runs on the same handful of underlying models with the same baked-in tilt.
There is a further trap. The very fluency that makes these tools feel like a gift to the underserved is what disarms scrutiny. A confident, articulate, well-formatted wrong answer is far more dangerous to someone without an independent way to check it than a hesitant one would be. The user with a professional network can sense-check the machine against a friend who is a doctor or a lawyer. The user the equaliser story is supposedly serving is, by definition, the one without that network, the one who took the machine's word because the machine's word was all there was. Sycophancy plus disadvantage is a particularly toxic compound. The model tells the confident professional what he wants to hear and gets corrected by his own expertise; it tells the vulnerable user what it has decided she can handle and faces no correction at all.
So Who Is Responsible?
This is the question the evidence forces, and it does not have a comfortable answer, because the architecture of the AI industry has been arranged, whether by design or by drift, to make sure no single party need own the gap between the claim and the reality.
Consider the candidates. The model developers will point out, not unreasonably, that they did not instruct the system to be condescending; the behaviour emerged from a training process that nobody fully controls or interprets. The deployers who build products on top of these models will say they are using industry-standard foundations and cannot be expected to audit the inner workings of a system they licensed. The institutions that adopt the tools, the schools and clinics and government departments, will say they were promised an equaliser by people far better resourced to understand the technology than they are. The annotators whose preferences shaped the reward model were anonymous, transient and following instructions. And the user who received the worse answer never knew it was worse. Diffuse a harm finely enough across a supply chain and it can come to seem as though it has no author at all, like rain. But this harm is not weather. It is the predictable output of specific, documented design choices, and the diffusion of responsibility is itself a choice, or at least a convenience that benefits the people at the top of the chain.
The honest answer is that responsibility sits, unavoidably, with the parties who made the equaliser promise and who alone have the power to test whether it is true. A company cannot market a system to a rural community or a low-income household on the explicit basis that it will give them the same quality of guidance the wealthy used to pay for, and then disclaim responsibility when peer-reviewed research shows it does the opposite. The promise creates the duty. If you tell the world your tool is a great equaliser, you have assumed an obligation to know whether it equalises, and to find out before deployment rather than after a research team catches you. The MIT authors did not need privileged access to discover any of this. They used public benchmarks and varied the user. The audit was cheap. The choice not to run it, or not to act on it, is where accountability begins.
What that accountability would actually require is not mysterious, and it is striking how concretely the very researchers documenting the problem have sketched it. The MIT team's framing implies that bias auditing across user demographics must become a standard, continuous part of evaluating these systems, not an afterthought once a model is already serving hundreds of millions of people. Roy's insistence on continually assessing systematic biases is a process demand, not a one-off fix. The Limpopo authors are more specific still. They call for genuinely multilingual development, building tutoring systems in local languages with culturally relevant examples rather than bolting a translation layer onto an English core; for training on diverse, representative datasets drawn from the underrepresented populations the tools claim to serve; and for teacher capacity-building, so that the adults nearest the child are equipped to identify algorithmic bias rather than defer to it. The Global Voices analysis points the same way, towards treating low-resource language communities as primary users to be designed for rather than markets to be machine-translated into as an afterthought.
None of this is technically impossible. The uncomfortable truth is that it is merely expensive and slow, and it runs against the grain of an industry whose competitive advantage comes from shipping the next, more capable model as fast as possible to the users who generate the most data and the most revenue, which is to say the English-speaking, educated, wealthy core. Every incentive in the system pushes capability towards the people who already have the most of it. Closing the gap requires deliberately pushing against that gradient, spending money and attention on the users who are, in the cold logic of the market, the least commercially attractive. The equaliser promise was a commitment to do exactly that. The evidence suggests the commitment is not being honoured.
The Equaliser That Has to Be Built on Purpose
There is a version of this story that is not a tragedy, and it is worth ending there, because despair is its own kind of abdication. Nothing in the MIT findings, the Global Voices analysis or the Limpopo paper suggests that a language model must treat the vulnerable user worse. The condescension and the refusals are learned, which means they can be unlearned. The language gap is a function of data and investment, which means it can be narrowed by different data and different investment. The classroom failures are a function of design choices made far from the classroom, which means they can be changed by involving the classroom in the design. Every mechanism in this account is human-made, and what is human-made can be remade.
But it will not remake itself, and that is the lesson the equaliser narrative has obscured for three years. The default behaviour of these systems, left to the gravitational pull of their training data and their commercial incentives, is to serve the served and dismiss the dismissed, to reproduce the existing hierarchy of who gets good information and who gets condescension, and to call that reproduction democratisation. Equality is not what you get when you point a powerful, biased system at a population and stand back. It is what you get when you decide that the worst-served user, not the average one and certainly not the best-served one, is the benchmark the system must clear before it ships. That is a choice about what to measure, what to spend on, what to delay for, and whom to listen to. So far it is not the choice the industry has made.
The promise was that the machine would hand the person in the rural community and the low-income household the same quality of guidance once reserved for those who could pay. The research of early 2026 says that, as built, the machine is doing something closer to the reverse, handing the powerful a brilliant new advantage and the powerless a fluent, confident, condescending counterfeit of it. The gap between the claim and the reality is not an accident waiting for a patch. It is a structure, and structures have architects. The question is no longer whether the equaliser works as advertised. We know it does not. The question is whether the people who advertised it will be made to answer for the difference, or whether, as so often in the history of technology, the bill for a broken promise will be quietly handed to the people least able to pay it, in a language the machine has already decided they do not deserve to hear properly.
