On the Production and Dissemination of Knowledge in the AI Era
Is AI (and AI Companies) Turning Knowledge into a Private Good?
Very short summary: This essay reflects on the nature of knowledge and the implication of its privatization by drawing a parallel between the production and dissemination of scientific knowledge through the academic publishing industry and the emergence of AI technologies. There are interesting commonalities, as in both cases “production factors” turn out to be freely used to produce further knowledge that, in contrast, is privatized. The negative implications for scientific knowledge are well-known. We are only beginning to understand the risks that AI technologies entail for general knowledge. Note: This essay has been written before the Mythos/Fable event.
Knowledge is a very peculiar “good” but also an essential one for any society. Peculiar because, contrary to most standard goods, once produced it is almost costless to disseminate. Essential because knowledge enters as an input in the production of virtually all products (including knowledge itself) and, more generally, is a necessary resource for any decision-making activity, both at the individual and the collective levels. In what follows, I draw a parallel between the way scientific knowledge has been produced and disseminated through the academic publishing industry and how AI technologies and companies will likely change the nature of knowledge in general. I identify a common property: the privatization of knowledge and the risks that go with it.
As an academic, I’m well-aware that I’m part of and directly contributing to a system of knowledge production and dissemination that, from the outside, seems hardly to make any sense at all. In this system, the producers of knowledge (academics), for the most part,[1] give their production (knowledge) for free to academic publishers who then disseminate this product in the form of journal articles. To work, this system needs “middle-(wo)men” whose job is to assess the product and decide whether or not it should be disseminated. These middle-(wo)men (referees, journal editors) are also academics, and they also provide this service for free. Of course, academics are paid, but by their universities, not by academic publishers. One implication is that there is a large disconnection between academics’ “marginal productivity” and their income. Economists will readily tell you that this is a problem.[2]
But this is not the biggest issue. Once turned into a material (a physical copy) or immaterial (a digital copy) product, knowledge is sold back to the very organizations that paid the producers, i.e., the universities. Universities pay very high subscriptions to academic journals, which in turn are used by academics to increase the stock of knowledge. In a way, that works – scientific knowledge has been drastically increasing over the past century. That works because, paradoxically, academics and universities are trapped in a typical social dilemma. We are all aware that there is something wrong with this system from an ethical point of view. Academic publishers make a lot of money[3] because their production costs are largely outsourced to universities and universities have no choice but to buy back the very product that they largely funded. As an academic, it is virtually impossible to commit not to participate in this system, unless you’re at a senior stage of your career where you don’t aim for another promotion or for a better job.
Part of the problem is that academics and universities have themselves outsourced a key function to academic journals, i.e., peer evaluation. Academics are judged by their publishing track-record, and since nobody has time to carefully read what others produce, you have no choice but to look at the list of their publications. We would all be better off if we don’t participate, but this is not credible, and academic publishers know that perfectly well.
Academic actors have tried to exit the dilemma by various means. It has proved partially effective, as witnessed by the fact that academic research is increasingly disseminated under various forms of open access. Readers (hence, universities) no longer have to pay to access knowledge. This “solution” consists, however, essentially in switching costs from the downstream of the production-dissemination process (those who read pay) to its upstream (those who produce pay), which makes it even more absurd. In the “open access era” of the academic publishing business, those who produce the good and most of the value (academics) have to pay several thousand dollars to acquire the right to publish in academic journals. Not only does that mean that publishers continue to outsource costs, but it also creates additional distortions. For instance, since their profits are now marginally affected by the size of the output (which was not the case under the traditional “closed access”), publishers are incentivized to increase the flow of accepted articles. Many academic editors will tell you stories about various forms of pressures they receive from their publishers to increase their acceptance rates.[4]
From an economic perspective, the academic publishing industry is a really interesting and non-standard case. Economists largely agree that knowledge is a non-rival good.[5] The fact that I “consume” a bit of knowledge doesn’t change anything about your ability to consume the same bit. My knowledge of economics, political philosophy, or American history is not rival with yours. In particular, non-rivalry implies that the cost of dissemination is very close to zero.[6] We can know the same things about the making of the Declaration of Independence or Rawls’s political turn without creating any additional cost to either of us. Non-rivalry implies that production and dissemination of knowledge are largely free from negative production and consumption externalities but, on the other hand, are very likely to generate positive externalities. Insofar as this is indeed the case, you want to encourage production and dissemination.
The relevant question is how production and dissemination must be organized, acknowledging that the incentive issue doesn’t disappear just because the good is non-rivalrous. The academic publishing system has evolved toward an institutional arrangement where knowledge is essentially a club good: you have to pay an entry fee (a subscription) and then you can virtually consume as much as you want – a bit like streamed music. The open access era slightly alters the logic. There is no longer any exclusion in use (everyone can consume for free), but access to the journals for producers is costly. Academic publishers have turned the publishing market into a two-sided one, very similar to what newspapers did to bring advertisers and readers together, with the exception that here producers and consumers are mostly the same.[7]
“Prometheus Bound,” Peter Paul Rubens (1612)
Let’s switch from academic publishing to the AI industry and consider the following basic facts. First, LLMs have been trained on a gigantic amount of data, most of it accessed without paying anything to “producers.” Among this data is the scientific knowledge produced by academics and disseminated by academic publishers. Second, LLMs not only parrot existing data, they can also create new ones. In other words, they not only disseminate preexisting knowledge, they can also potentially create new knowledge, building on the preexisting one. Third, while initially mostly free, access to this newly created knowledge has become over the past months increasingly expensive. While subscription fees haven’t necessarily increased, all AI companies have significantly reduced the number of “tokens” you can use for a given cost. Finally, it is highly likely that this trend will continue. As a matter of strategy, AI companies have been keeping prices artificially low to attract and capture new users. Once users, especially businesses, strongly depend for their performance on AI products, they will have no choice but to consent to pay higher prices – this is a classic instance of the “hold-up problem.”[8]
These facts induce a strong analogy with academic publishing. Like academic publishers, AI companies have been relying on a product whose production cost is entirely outsourced. Like academic publishers in the “closed-access era,” they charge users without remunerating producers. Like academics with respect to the academic publishing system, users are increasingly trapped in a social dilemma and cannot credibly commit to stop using AI products, even as their prices keep increasing. There are some differences, though: AI companies clearly create additional value as LLMs and other products generate new knowledge. In turn, AI companies’ costs are much higher than those of academic publishers. We know that the academic publishing business is highly profitable; this remains to be established for AI companies.
However, these differences are not essential. In the same way that academic publishers have turned scientific knowledge into a club good, AI companies could restrict access to knowledge in general. This creates significant risks that should not be underestimated. A first one is to reinforce economic inequalities and diverging growth trajectories between countries. If you have tried various paid subscriptions from the same AI company, you have probably noticed that there is a huge difference between the cheapest one (typically around $25/month or $250/year) and the typical “max” one (around $200/month). Not everyone can afford the latter, let alone the even more expensive versions. The point is not that this is not worth the price – it probably is. The problem is that on our imperfectly competitive markets, many agents face liquidity and wealth constraints that limit their capacity to access resources, which in turn creates inefficiencies, on top of creating fairness concerns.[9]
The second risk is that access to knowledge could be intentionally restricted for reasons other than affordability. Security considerations, as illustrated by Anthropic’s decision not to publicly release its new model Mythos, are the most relevant candidates. But in a multipolar and conflict-ridden world, what is “security” for some means insecurity for others. This is, of course, not new. Regular citizens ignore many things that are kept secret by the highest state officials, and that is surely for the best. However, in a democracy, what is kept secret and what is allowed to become public should be determined based on principles that most agree with, not by private companies – even well-intentioned ones – and even less at the discretion of a state with authoritarian proclivities. Moreover, we are talking about knowledge in general, not bits of highly specific information. What is at stake is people’s access to knowledge and knowledge-producing tools that are directly relevant and valuable for their daily private and professional lives.
Finally, as for academic publishing, we cannot ignore the legitimacy and fairness issues. The nature of the rents earned by academic publishers raises a fairness issue. These rents exist because publishers are exploiting a social dilemma that puts academics and academic institutions in the practical impossibility of forcing publishers to pay for the marginal value they are creating while they must use (and pay for) publishers’ services. If the AI business turns out to be profitable, the rents earned will have similar features – except that at some point LLMs and other AI tools will be trained on AI-generated knowledge. The market power of academic publishers and AI companies is an economic phenomenon. But because the production and dissemination of (scientific) knowledge is also politically relevant, this also triggers a legitimacy issue: what justifies granting so much power to private actors?
As I noted above for academic publishing, the fact that knowledge is a non-rival good doesn’t mean that the incentive problem goes away. Non-rivalry can perfectly well go with (price-based) exclusion and, indeed, the very production of the good (knowledge) is likely to require it. Many academics argue that we should simply get rid of private academic publishers and reorganize the production and dissemination of scientific knowledge by turning it into a public good. While the transition to this institutional arrangement is unclear (again, academics are trapped in a social dilemma), the solution may eventually work just because the value-added of academic publishers is unclear and because academics are, de facto, already incentivized.[10]
Nothing like this solution exists in the case of general knowledge and AI companies. If we grant the claim that AI technologies will trigger social benefits that largely surpass their social costs (a claim that can of course be disputed), then it is desirable that these technologies are effectively developed. In all likelihood, this is something that can only be done where incentives are derived from market-based competition and, therefore, one form or another of exclusion. This would not be a problem, were it possible to make AI companies pay for the knowledge on which they train their models. Since it is clearly difficult, an ex-post tax on their profits, possibly used to fund a basic income, seems to be justified. This doesn’t directly address the first two risks identified above, but may lessen the fairness concern.
[1] I put aside the case of knowledge produced and disseminated through books, which in any case represents a small and decreasing fraction of the knowledge produced by academics.
[2] The problem is partially mitigated by the fact that some universities or business schools pay their professors when they publish an article, on top of their regular income. Also, of course, hiring and promotions depend on publishing track-records, which generates monetary incentives to publish, at the margin.
[3] Many studies indicate that the major academic publishers’ (Elsevier, Springer Nature, Wiley, Taylor and Francis) operating margins are between 25% and 40%, which is more than in the pharma industry or the average S&P 500 company.
[4] Such stories are behind the resignations en masse of the editorial boards of major philosophy outlets like The Journal of Political Philosophy (which is now terminated) and Philosophy and Public Affairs over the past couple of years.
[5] This is the core idea behind modern “endogenous” growth theories like Nobel prize-winner Paul Romer’s.
[6] Not absolutely zero. Academic publishers do have production costs and, at least with regard to physical copies of articles, their marginal cost of production is positive.
[7] Another difference is that the value-added of academic publishers is probably fairly low compared to newspapers.
[8] The hold-up problem is the most severe when assets are highly specific. As there are several AI companies, a user can still in theory switch from one to the other, which should limit the hold-up risk. However, the market is oligopolistic (there are few producers), AI companies are increasingly specializing their products, and switching from one set of products to another is likely to be costly. Finally, there is no substitute for AI products, and users cannot produce them themselves.
[9] For instance, wealth constraints are a well-known cause of inefficiencies in the credit market.
[10] I would add that the value-added of academic publishers is even lower with AI. Editing and proof-reading articles are tasks that can now be largely automated by the writers themselves. What publishers chiefly provide are platforms to manage submitted articles and reputation, nothing that universities (public or private) cannot provide themselves in principle.