Paul Morrill | Independent Researcher | October 2026
Abstract
Noam Chomsky's criticism of large language models presents an interesting tension with his earlier work on the institutional control of information. In Manufacturing Consent, Herman and Chomsky (1988) examined the structures through which certain ideas acquire legitimacy while others struggle to gain recognition. Generative artificial intelligence may reproduce some of these arrangements, although it also offers something rather different: the possibility of extending intellectual assistance to people who have traditionally lacked access to specialist knowledge and the means of developing their ideas. Drawing upon Chomsky's writing alongside contributions from Steven Pinker, Steven Piantadosi, Erik Brynjolfsson, Daron Acemoglu and others, this essay considers whether AI might contribute to a more equitable distribution of intellectual opportunity. Particular attention is given to cognitive augmentation, intellectual diversity, the relationship between expertise and professional identity, and the possibility that familiar forms of gatekeeping may simply be replaced by new ones. I argue that predictions of widespread cognitive homogenisation remain premature. The evidence suggests a more complicated picture, in which improved individual performance may coexist with new forms of dependence and institutional inequality. Whether AI ultimately broadens intellectual participation or reinforces existing arrangements remains uncertain, although I suspect its potential to extend intellectual opportunity deserves considerably more attention than it presently receives.
Introduction: A rather useful pair of trouser braces
I have come to think of artificial intelligence, at least in its present form, as something resembling a pair of trouser braces. Perhaps not the most elegant comparison, but I think it serves its purpose.
Braces provide support where it is needed. They do not change the person wearing them, nor do they determine what that person might subsequently do. In much the same way, I see generative AI as something that can assist people who possess interesting ideas but lack some of the skills, resources or connections necessary to develop them. Someone might have a perfectly reasonable psychological hypothesis but little understanding of computational modelling. Another may possess considerable practical knowledge but lack the academic vocabulary required to communicate it within a specialist discipline. Others may simply have spent much of their lives being given the impression that certain intellectual conversations are not intended for them.
Historically, such obstacles could be sufficient to prevent an idea from progressing. Admittedly, some of those ideas may have been complete rubbish. But others might have contained something original, useful or worthy of further investigation. We have no reliable means of knowing how much potentially valuable thought has been lost because its author lacked the means of presenting it in an acceptable form.
I have discussed aspects of this problem elsewhere, particularly in The Privilege of Being Heard, where I considered the relationship between intellectual competence and the ability to communicate within socially privileged registers. I do not intend to revisit that argument here. What interests me now is the possibility that artificial intelligence may be altering the relationship between intellectual potential and the resources necessary to realise it.
This brings me to Noam Chomsky, whose work provides a rather interesting starting point, although perhaps not for the reasons he might particularly welcome.
Chomsky has been among the more prominent critics of contemporary large language models. Yet his earlier political writing, particularly Manufacturing Consent, concerned the ways in which institutional structures influence which ideas become visible, credible and publicly acceptable (Herman & Chomsky, 1988). I find myself wondering whether generative AI might, under certain circumstances, begin to undermine some of those very arrangements.
I am not suggesting that AI necessarily possesses genuine understanding, nor that its outputs should automatically be treated as knowledge. Those are separate matters, and rather substantial ones at that. My interest is in whether the technology might change who can participate meaningfully in the development and examination of ideas.
I suspect it already is, although where this eventually leads is another matter entirely.
Chomsky, artificial intelligence and the problem of intellectual authority
In 1988, Edward Herman and Noam Chomsky published Manufacturing Consent: The Political Economy of the Mass Media. Their argument was considerably more sophisticated than the familiar suggestion that newspapers simply deceive their readers or that governments determine everything appearing on television. They described a system in which media ownership, advertising, sourcing practices, institutional pressures and prevailing ideological assumptions influenced the information presented to the public.
An important aspect of their propaganda model was that these influences did not necessarily depend upon deliberate deception. Journalists could act sincerely, follow professional conventions and genuinely believe themselves to be independent while operating within structures that systematically favoured particular interpretations of events (Herman & Chomsky, 1988).
It is not difficult to see why this account might be applied to contemporary AI. Large language models are developed by organisations with commercial interests, trained on material containing existing cultural and institutional assumptions, and subsequently modified according to technical, commercial and regulatory considerations. It would be rather surprising if these influences had no effect upon the information they produce.
Yet I think the more interesting possibility lies in examining what happens when the argument is turned around.
The mass media described by Herman and Chomsky operated through relatively concentrated channels of information production and distribution. Generative AI introduces a rather different arrangement. Individuals can now request explanations of unfamiliar subjects, explore alternative arguments and develop material that would previously have required access to specialist knowledge or professional assistance.
Of course, being able to request an explanation does not guarantee that the explanation will be correct. But it does alter the practical circumstances under which someone can begin investigating an idea.
Chomsky's own objections to contemporary AI have been principally concerned with the nature of intelligence and understanding. In The False Promise of ChatGPT, written with Ian Roberts and Jeffrey Watumull, he argued that large language models should not be confused with systems capable of human-like explanation and reasoning (Chomsky et al., 2023). Their criticism distinguished statistical prediction from the construction of explanatory theories, maintaining that human language and cognition involve capacities not adequately captured by contemporary machine-learning approaches.
Chomsky subsequently defended his position in an interview conducted by Ramin Mirfakhraie (2023), rejecting claims that the success of large language models had seriously undermined his account of human language acquisition. He drew attention to the distinction between reproducing linguistic behaviour and explaining the biological and cognitive mechanisms through which children acquire language.
This distinction is important. A system may produce convincing linguistic outputs without necessarily providing an adequate account of human language development. The mechanisms through which a model acquires linguistic regularities may differ considerably from those involved in biological cognition. Consequently, the success of language models does not, by itself, establish that Chomsky's theories of human language acquisition are incorrect.
Nevertheless, his criticism also raises a question that I think is sometimes overlooked.
Does an artificial system need to understand something in precisely the manner of a human being before it can help a human being understand it?
A calculator need not understand mathematics in the manner of a mathematician to be useful in mathematical work. Admittedly, language models are rather less dependable than calculators, frequently producing convincing explanations of things they have got completely wrong. Nevertheless, the distinction between the internal capacities of an artefact and the benefits it may provide to a human user remains relevant.
Indeed, towards the end of his interview with Mirfakhraie, Chomsky suggested that education in critical thinking and intellectual self-defence was among the most effective responses to the risks posed by these technologies. I find that observation particularly interesting, since the very technology he criticises might also provide some individuals with greater access to the resources necessary to develop those capacities.
That possibility is not established by Chomsky's comments, and I doubt he would necessarily endorse my interpretation. Nevertheless, it seems worthy of consideration.
If institutional structures influence which ideas are allowed to flourish, what happens when people acquire new means of developing ideas without first obtaining the approval of those institutions?
Perhaps we should be examining not only what AI does to human thought, but what it does to the arrangements through which human thought is recognised and judged.
Pinker, Piantadosi and the limits of an artificial mind
Steven Pinker's discussion of ChatGPT provides a useful counterpoint to Chomsky's position, although their views are perhaps less distant than they initially appear.
In a 2023 interview with the Harvard Gazette, Pinker described the remarkable capabilities of contemporary language models while expressing considerable scepticism about their reliability and understanding (Powell, 2023). He was particularly concerned with their tendency to produce confident inaccuracies and with the limitations of systems that generate plausible language without possessing the conceptual models of the world that humans ordinarily use.
Pinker nevertheless recognised the possibility that artificial intelligence might extend our intellectual capabilities, particularly in activities involving information retrieval and the processing of large quantities of material. He compared this prospect with earlier technologies that had allowed humans to compensate for limitations in memory and calculation.
His position is therefore more qualified than an enthusiastic endorsement of generative AI. He acknowledges what these systems can accomplish while remaining sceptical about some of the claims made on their behalf.
I find this a useful way of approaching the subject.
There is no particular reason to assume that useful artificial intelligence must reproduce human cognition faithfully. Human intelligence itself is hardly a uniform collection of abilities, and competence in one intellectual activity tells us remarkably little about performance in another.
Pinker also questioned the treatment of intelligence as a single dimension along which humans and machines might be compared. Different systems possess different strengths and weaknesses, and the development of useful artificial intelligence need not involve reproducing the biological processes through which humans acquire their abilities.
It is worth remembering that these observations were made in early 2023. The capabilities of publicly available systems have changed considerably since then, and we should be cautious about treating comments concerning earlier models as comprehensive judgements about those available today.
Steven Piantadosi (2024) offers a more direct challenge to Chomsky's linguistic theories. He argues that the success of contemporary language models calls into question important assumptions within traditional generative linguistics, particularly strong claims concerning the necessity of particular innate structures for acquiring language.
Piantadosi's argument is substantial, and his criticism deserves to be taken seriously. However, it would be misleading to suggest that the success of large language models has decisively refuted Chomsky's approach to human language acquisition.
There are important differences between demonstrating that a computational system can acquire linguistic regularities and explaining how biological organisms acquire language. Chomsky's theoretical claims concern the human faculty of language, while Piantadosi places considerable emphasis upon what successful computational learning can tell us about the necessity of proposed linguistic structures.
The disagreement remains a genuine theoretical dispute. It concerns what should count as an explanation of language, which evidence is relevant, and how far success in artificial systems can inform our understanding of biological cognition.
I am inclined to think that both positions contain questions worth pursuing, although I would be wary of declaring either victorious on the basis of present evidence.
For my purposes, the disagreement illustrates something else. We have specialists with considerable expertise arriving at markedly different interpretations of the same technological developments.
I do not regard this as evidence that expertise has become redundant. Quite the opposite. It demonstrates why competing theoretical accounts, empirical investigation and informed criticism remain essential.
What may be changing is the extent to which people outside established intellectual communities can participate in those discussions.
Does artificial intelligence actually lift those who need assistance?
My trouser-braces analogy would be of little use if there were no evidence that AI provides meaningful assistance to people who are less experienced or less proficient in particular activities.
Fortunately, there is some relevant research, although the findings require more careful interpretation than enthusiastic accounts of AI productivity sometimes allow.
Brynjolfsson, Li and Raymond (2025) examined the introduction of a generative AI assistant among 5,172 customer-support workers. They reported an average productivity improvement of approximately 15%, with the greatest benefits occurring among less-experienced and lower-skilled employees. More experienced workers obtained comparatively modest benefits, while some showed small reductions in measured quality.
The researchers also found evidence consistent with workers acquiring useful task-related knowledge through their interaction with the system.
These findings interest me because they suggest that AI can, under certain circumstances, reduce performance differences associated with experience and prior knowledge.
Noy and Zhang (2023) reported related results in an experimental investigation involving 453 professionals completing writing tasks. Participants using ChatGPT completed their work substantially more quickly and produced outputs that received higher quality ratings. Average completion time fell by approximately 40%, while evaluated quality increased by around 18%.
Taken together, these studies offer some support for the view that AI may be particularly useful to those who have not already acquired the skills necessary to perform a given activity efficiently.
There are, however, some fairly obvious limitations.
Improved performance on a writing task is not necessarily evidence of improved understanding. Someone might produce a perfectly respectable document with AI assistance while remaining unable to explain much of its content. Similarly, greater workplace productivity tells us relatively little about scientific originality or the development of genuine expertise.
I would also question whether the distinction between skilled and unskilled individuals is always a particularly helpful one. A person may lack mathematical competence while possessing considerable theoretical insight. Another may have extensive practical experience but little familiarity with the conventions of academic writing.
I suspect that many of the people who stand to benefit from AI are not lacking in intelligence or originality. They may simply lack particular skills that have traditionally served as prerequisites for participation in specialist activities.
That, to my mind, is where the technology becomes interesting.
It may be premature to claim that AI democratises expertise, but there is a reasonable argument that it can make some of the means through which expertise is developed more widely available.
The important question is whether this assistance eventually translates into acquired understanding.
Both Brynjolfsson et al. (2025) and Noy and Zhang (2023) provide evidence of improved performance in specific settings. Neither establishes that widespread access to AI will produce enduring improvements in independent intellectual competence across the population.
We need to make that distinction quite explicit. The possibility of cognitive augmentation is encouraging, but sustained intellectual development should be demonstrated rather than inferred from the quality of an assisted output.
Cognitive flattening: An inevitable consequence?
One of the more persistent concerns surrounding generative AI is that its widespread use will produce an increasingly homogeneous intellectual landscape.
The argument is understandable. Language models learn from existing material, reproduce statistical regularities and may encourage users to adopt familiar forms of expression or conventional approaches to problems. If large numbers of people rely upon similar systems, we might reasonably expect some convergence in their outputs.
There is evidence that this can occur.
Doshi and Hauser (2024) investigated the effects of generative AI assistance on creative writing. They found that AI-generated suggestions improved evaluations of individual stories, particularly among participants with lower initial creative performance. However, the stories produced with assistance were also more similar to one another than those produced without it.
I find this a particularly interesting result because it complicates my own position.
AI appears to have improved individual creative performance while simultaneously reducing diversity across the resulting material. It would be rather convenient to ignore this finding, but it would also be intellectually dishonest.
Nevertheless, I am not convinced that the result justifies broader predictions of cognitive flattening.
We need to be careful about what is actually being measured. Similarity in written expression does not necessarily indicate similarity in underlying thought. Nor does convergence in the solutions produced during a relatively constrained experiment establish that people's wider intellectual capacities are becoming less diverse.
There is also the question of who participates.
Suppose a particular academic discipline has historically attracted contributions from a relatively narrow population, selected partly through educational opportunity, financial resources, professional connections and familiarity with specialist language. AI may allow people outside that population to develop and communicate ideas that would previously have remained inaccessible.
Some of their contributions may resemble existing academic writing. Others may reproduce familiar arguments. But there is also the possibility that people with different experiences, practical knowledge and intellectual interests will introduce ideas that the discipline might otherwise never have encountered.
We could therefore witness greater similarity in the presentation of ideas alongside greater diversity in the people producing them and the experiences informing them.
That possibility seems to receive relatively little attention in discussions of intellectual homogenisation.
I have previously argued in Academic Language Was Flattening Before AI Arrived that institutional pressures towards uniform expression considerably predate generative AI. I see no value in repeating that argument here. What interests me now is whether the technology might allow a wider variety of thought to emerge through those increasingly familiar linguistic forms.
There is another issue worth considering. AI assistance need not always involve asking a system to provide a completed answer. It can be used to examine objections, identify methodological weaknesses, explore unfamiliar theories or challenge assumptions that the user has brought to a problem.
The outcomes may be very different depending upon how the system is used.
A person who repeatedly accepts the first plausible response may gradually become dependent upon familiar explanations. Another who uses the same technology to interrogate those explanations may acquire a much broader understanding of the subject.
We should not assume that the latter will inevitably occur, but neither should we disregard it.
For these reasons, I remain sceptical of predictions that generative AI will necessarily produce a cognitively flattened population. The available evidence suggests competing influences, and we are some distance from understanding how these will develop over years of sustained use.
The rather inconvenient problem of replacing one gatekeeper with another
There is a serious objection to my argument, and I think it is probably the strongest one.
What if artificial intelligence does not dismantle intellectual gatekeeping at all? What if it merely transfers some of that authority to a different set of institutions?
Herman and Chomsky's (1988) account concerned structural influences upon the production and distribution of information. Contemporary AI introduces its own structures of control through ownership of computational resources, access to training data, model design, commercial priorities and decisions about the material systems will generate.
Emily Bender and colleagues raised related concerns in On the Dangers of Stochastic Parrots (Bender et al., 2021). Their criticism addressed the social and environmental consequences of increasingly large language models, including the risk that training data reproduce existing cultural inequalities and disproportionately represent dominant linguistic communities.
These concerns should not be dismissed. A system that appears capable of answering questions about almost anything may nevertheless contain substantial limitations in what it can represent, explain or recognise as credible.
There is also an uncomfortable psychological possibility.
If someone relies upon an AI system to formulate arguments, identify relevant evidence and determine which explanations appear reasonable, they may gradually surrender some of the intellectual independence that the technology was supposed to support.
The old gatekeeper might have been a professor, editor or professional organisation. The new one could be a commercial system whose decisions are considerably more difficult to examine.
I am not entirely persuaded, however, that this makes the previous arrangements preferable.
Traditional intellectual institutions have never provided unrestricted access to knowledge. Academic publishing, professional credentials, research funding and established networks have all influenced whose ideas receive attention. The discovery that AI systems contain biases does not somehow establish that the institutions preceding them were free of such influences.
The comparison needs to be between actual arrangements for producing and validating knowledge, rather than between an imperfect technology and an idealised account of intellectual life before its arrival.
Equally, using AI does not necessarily involve surrendering judgement to it. Someone who consults competing sources, examines the evidence behind a response and remains prepared to reject an explanation is engaging with the technology differently from someone who accepts whatever appears on the screen.
Of course, even the most critical user cannot be expected to recognise every omission or assumption embedded within a model. That is part of the problem. We may learn to question what the system tells us while remaining unaware of what it has failed to tell us.
Perhaps the more pressing issue is whether people will develop the intellectual habits necessary to recognise when their artificial assistant is providing useful support and when it is simply offering another convincing explanation of something it does not adequately understand.
I suspect the answer will depend as much upon educational practice and the design of these systems as upon the technology itself.
Economic equality and the ownership of intellectual opportunity
There is another assumption underlying my optimism that deserves scrutiny.
Even if AI makes sophisticated intellectual assistance more widely available, it does not follow that the benefits will be distributed equitably.
Daron Acemoglu (2025) provides an important corrective to some of the more extravagant economic predictions associated with AI. His analysis questions expansive productivity forecasts and emphasises the importance of the tasks being automated, the organisation of work and the institutional arrangements through which economic benefits are distributed.
Increased productivity does not automatically produce greater equality. Workers may become more efficient without receiving a corresponding improvement in their circumstances. Much of the resulting value may accrue to employers or technology providers.
I suspect a comparable problem exists in the intellectual sphere.
Someone may acquire the ability to develop a sophisticated research proposal while remaining excluded from the institutions capable of funding it. Another may produce technically competent work but lack the professional recognition necessary to have it taken seriously.
Access to the means of developing an idea is not the same as access to the means of testing, publishing or profiting from it.
This returns us, rather neatly, to Herman and Chomsky. Their concern was not simply with whether information could be produced, but with the structures determining which information acquired public prominence and credibility.
Generative AI may allow considerably more people to produce sophisticated intellectual material while leaving important aspects of institutional authority relatively untouched.
Nevertheless, I see no reason to regard the persistence of inequality as an argument against expanding intellectual access. A technology can offer genuine individual benefits while the arrangements governing its development and ownership remain unsatisfactory.
UNESCO's guidance on generative AI in education and research highlights the importance of inclusion, human agency, privacy and appropriate institutional safeguards (Miao & Holmes, 2023). Such concerns are particularly relevant if access to AI becomes an increasingly important component of educational and professional opportunity.
Greater equality will not simply emerge because sophisticated software becomes available. Educational provision, economic circumstances and institutional practices will continue to matter.
But I would rather see those difficulties addressed than use them as justification for preserving barriers that might otherwise be reduced.
From improved performance to intellectual agency
There is a psychological question underlying much of this discussion that I think deserves closer examination.
What happens when someone who has previously regarded a particular intellectual activity as beyond their capabilities suddenly finds that they can participate in it with assistance?
The immediate consequences may be fairly straightforward. They can produce something they could not previously produce, or perhaps complete a task that would otherwise have taken considerably longer.
The longer-term consequences are less obvious.
Bandura's (1997) work on self-efficacy suggests that beliefs about personal capability influence whether individuals attempt difficult activities, how much effort they invest and how they respond to failure.
If AI assistance allows people to undertake unfamiliar intellectual work successfully, it is conceivable that repeated experiences of this kind might alter their perceptions of what they are capable of achieving.
Someone who once believed that computational modelling was entirely beyond them may discover that they can begin to understand the principles involved. Another may become sufficiently confident to pursue an academic subject that previously appeared inaccessible.
I find this possibility encouraging, although there is a rather awkward complication.
Successful assisted performance may not always translate into genuine learning.
An individual might produce increasingly sophisticated material while acquiring very little independent competence. In such cases, apparent improvements in intellectual capability could conceal growing dependence upon the system providing the assistance.
Research on cognitive offloading is relevant here. Risko and Gilbert (2016) describe how people use external resources to reduce demands upon internal cognitive processes. Such behaviour is hardly new. Written records, calculators, diagrams and countless other tools have long allowed humans to accomplish tasks that would be considerably more difficult without assistance.
There is nothing inherently undesirable about cognitive offloading. Much of human intellectual achievement depends upon distributing cognitive work across people, tools and external representations.
The question is what happens when the assistance becomes sufficiently sophisticated to undertake activities that previously required substantial understanding.
Does the individual acquire knowledge that can subsequently be applied without assistance? Does their capacity to recognise errors improve? Can they explain why a particular solution works, or do they simply become more proficient at obtaining convincing outputs?
These are empirical questions, and the answers may differ considerably between tasks and individuals.
I would argue that a worthwhile measure of intellectual augmentation should include the user's developing ability to challenge the system.
Someone who becomes increasingly capable of identifying weaknesses in AI-generated explanations has arguably acquired something more substantial than improved productivity.
Someone who becomes increasingly confident in material they cannot adequately evaluate may have acquired rather less than appearances suggest.
This is where I think psychological research could make a particularly valuable contribution. We need longitudinal investigations that distinguish improvements in assisted performance from improvements in independent understanding, confidence and critical judgement.
One could, for example, compare groups receiving different forms of AI assistance during the development of unfamiliar technical skills, before subsequently withdrawing that assistance. Measures of independent problem-solving, error detection, confidence calibration and transfer to novel tasks might help establish whether the observed benefits reflect genuine learning or improved access to external cognitive resources.
Such research should also consider whether the nature of interaction makes a difference. An assistant that provides completed solutions may produce different learning outcomes from one that asks questions, introduces objections and requires the user to explain their reasoning.
I would be particularly interested in whether people who initially struggle with a task benefit disproportionately from the latter approach, and whether these benefits persist after the technology has been removed.
The finished product alone cannot tell us what the person producing it has learned.
Expertise and the possibility of a different intellectual settlement
I find myself increasingly unconvinced by discussions that present the relationship between human and artificial intelligence as a competition in which one must inevitably diminish the other.
There will undoubtedly be activities in which AI reduces demand for particular human skills. Some of those skills have required years of training, and the prospect of their becoming commonplace will understandably be unsettling for people whose professional identities depend upon them.
I have discussed aspects of this problem in I'm Sorry, Boys and Girls, but the Way Academia Currently Rewards Is Over, so I shall avoid returning to the same arguments at length.
What interests me here is what might eventually replace the familiar arrangements.
The ability to generate a statistical analysis, formulate a plausible argument or produce technically sophisticated writing may become increasingly widespread. But the ability to recognise methodological weaknesses, evaluate competing explanations and understand the significance of a result will presumably remain important.
Indeed, the growing availability of convincing but potentially unreliable material may make such capabilities even more valuable.
Perhaps the future of expertise will involve less emphasis upon controlling access to specialist procedures and greater emphasis upon the judgement required to use them properly.
I appreciate that this is easier to suggest than to demonstrate.
Established professional arrangements are embedded within educational institutions, employment structures, funding mechanisms and social expectations. They are unlikely to change simply because new technical possibilities become available.
Nor should we assume that technological change invariably produces beneficial institutional reform.
Acemoglu's (2025) analysis provides good reasons for caution, particularly concerning the distribution of economic benefits. Bender et al. (2021) also remind us that the social consequences of AI depend partly upon whose interests are represented in the design and deployment of these systems.
Nevertheless, I suspect that some familiar distinctions between the possession of specialist expertise and the capacity to contribute to specialist activities will become increasingly difficult to maintain.
We should also avoid being carried away by claims that the present transformation is unlike anything in human history. The printing press, industrialisation, mass education and the internet all changed access to information and the organisation of intellectual life. Whether AI eventually proves more consequential remains to be seen.
What does seem unusual is the extent to which contemporary systems can participate in activities previously regarded as distinctly intellectual. We are no longer dealing solely with technologies that store, transmit or retrieve information. Increasingly, they can assist in the construction, criticism and development of ideas.
The consequences of that development may prove considerable, although I would be wary of anyone claiming to know precisely what they will be.
Manufacturing consent, or expanding intellectual agency?
Returning to Chomsky, I think the most interesting question concerns the relationship between intellectual authority and the capacity to participate in its production.
Herman and Chomsky (1988) examined the institutional structures influencing which explanations become credible and publicly acceptable. Their model remains relevant to contemporary AI, particularly where commercial concentration and opaque design decisions affect the information that systems make available.
But I do not think it follows that AI must necessarily strengthen the structures they described.
A technology capable of reproducing dominant assumptions may also provide people with the means of investigating and challenging them.
Someone previously excluded from a specialist conversation may acquire sufficient knowledge and vocabulary to participate. Another may be able to examine an institutional claim that would once have been difficult to question without professional assistance.
And occasionally, someone might discover that an idea they have carried around for years is not quite as foolish as they had been led to believe.
That possibility appeals to me, although I recognise the danger of romanticising it.
Some AI-assisted ideas will be dreadful. Others will merely reproduce arguments that have already been examined and rejected. There will undoubtedly be people who mistake polished presentation for genuine understanding, and others who acquire a rather inflated view of their capabilities.
But these shortcomings are hardly exclusive to artificial intelligence. Academic institutions have managed to produce all three without the slightest assistance from a language model.
What concerns me about some contemporary criticism is the tendency to interpret the diminishing exclusivity of certain intellectual activities as evidence of a corresponding decline in intellectual standards.
The two are not necessarily related.
The ability to express a technically sophisticated argument may become more widespread without rendering sound reasoning, scientific judgement or genuine originality less valuable.
Indeed, greater participation could expose specialist disciplines to ideas and experiences that their traditional recruitment and recognition practices have struggled to accommodate.
There is no guarantee that this will occur. It is also entirely possible that concentrated ownership of AI systems will create new restrictions upon intellectual independence.
I suspect both developments could emerge simultaneously.
This is why I remain uncomfortable with categorical positions, whether they predict a future of universal intellectual emancipation or the gradual erosion of human thought.
Neither seems particularly well supported by the available evidence.
Conclusion: Keeping our trousers up
I began with a pair of trouser braces, and I think the comparison remains useful.
Their purpose is to provide support where it is needed. They do not confer competence, intelligence or originality upon the wearer, although they may make certain activities considerably easier to undertake.
I see something similar in contemporary artificial intelligence.
A person with an interesting idea but limited technical knowledge may now have the means to explore it. Someone previously discouraged by the conventions of specialist communication may find those conventions less intimidating. Others may acquire sufficient confidence to attempt intellectual activities they would once have considered beyond their abilities.
I cannot see why these possibilities should automatically be regarded as threatening to human intelligence.
There are genuine problems, of course. AI may encourage dependence, reproduce cultural assumptions, generate convincing nonsense and place considerable intellectual influence in the hands of a relatively small number of commercial organisations. Those concerns deserve proper investigation.
But so does the possibility that more people may become capable of developing and contributing ideas that would otherwise have remained unheard.
Perhaps the greatest consequence of generative AI will not be that everyone begins to think alike. It may be that considerably more people acquire the means to participate in activities from which they were previously excluded.
Whether their contributions will be recognised is another matter.
And perhaps this is where Chomsky's two intellectual worlds become particularly interesting. One examines the institutional arrangements through which ideas acquire legitimacy. The other concerns the nature of intelligence and the mechanisms through which knowledge is produced.
Generative AI may eventually force us to reconsider both.
I remain cautiously optimistic, although I am sufficiently old and cynical to recognise that technological progress and human improvement do not always arrive together. We may encounter new inequalities, unfamiliar dependencies and institutional conflicts whose consequences are difficult to anticipate.
There is also the unresolved question of what increasingly sophisticated artificial systems might eventually become. Whether silicon intelligence possesses, or could possess, anything resembling subjective experience remains a matter of considerable philosophical and scientific disagreement. I have considered aspects of that problem in Beyond the Biological Brain, and do not intend to revisit it here.
For the present, I am more interested in what artificial intelligence might allow human beings to accomplish than in attempts to establish whether it thinks precisely as we do.
I suspect the structures surrounding intellectual participation are changing in ways we have barely begun to understand. Some established skills will become less exclusive, while others may acquire greater importance. Institutions will undoubtedly attempt to accommodate these developments, although I would not expect the process to be particularly graceful.
None of this guarantees a more equitable intellectual future. The empirical evidence currently demonstrates improvements in particular assisted tasks, sometimes accompanied by reductions in the differences between more and less experienced individuals. It does not yet establish that those benefits will translate into sustained independent understanding, greater intellectual diversity or a more equitable distribution of institutional authority.
Those are questions for further investigation.
Nevertheless, I cannot help thinking that we are overlooking something potentially important when we concentrate almost exclusively upon what AI might take away from human beings.
If artificial intelligence allows even a proportion of those whose ideas would previously have been ignored to develop them into something that can be examined and challenged, I would regard that as a worthwhile development.
The braces, after all, are only there to keep the trousers up. What we choose to do once we are standing is another matter entirely.
References
Acemoglu, D. (2025). The simple macroeconomics of AI. Economic Policy, 40(121), 13–58. https://doi.org/10.1093/epolic/eiae042
Bandura, A. (1997). Self-efficacy: The exercise of control. W. H. Freeman.
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency (pp. 610–623). Association for Computing Machinery. https://doi.org/10.1145/3442188.3445922
Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. https://doi.org/10.1093/qje/qjae044
Chomsky, N., Roberts, I., & Watumull, J. (2023, March 8). The false promise of ChatGPT. The New York Times. https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chatgpt-ai.html
Doshi, A. R., & Hauser, O. P. (2024). Generative AI enhances individual creativity but reduces the collective diversity of novel content. Science Advances, 10(28), eadn5290. https://doi.org/10.1126/sciadv.adn5290
Herman, E. S., & Chomsky, N. (1988). Manufacturing consent: The political economy of the mass media. Pantheon Books.
Miao, F., & Holmes, W. (2023). Guidance for generative AI in education and research. UNESCO. https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research
Mirfakhraie, R. (2023, April 24). ChatGPT and human intelligence: Noam Chomsky responds to critics [Interview with Noam Chomsky]. Monthly Review Online. https://chomsky.info/20230424-2/
Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192. https://doi.org/10.1126/science.adh2586
Piantadosi, S. T. (2024). Modern language models refute Chomsky's approach to language. In From fieldwork to linguistic theory: A tribute to Dan Everett (pp. 353–414). Language Science Press. https://doi.org/10.5281/zenodo.12665933
Powell, A. (2023, February 14). Will ChatGPT supplant us as writers, thinkers? The Harvard Gazette. https://news.harvard.edu/gazette/story/2023/02/will-chatgpt-replace-human-writers-pinker-weighs-in/
Risko, E. F., & Gilbert, S. J. (2016). Cognitive offloading. Trends in Cognitive Sciences, 20(9), 676–688. https://doi.org/10.1016/j.tics.2016.07.002