AI is Inherently Evil: Part 4
On the biases of the internet
4. Initial information
Theoretically, a universal computer can store any information. In practice, computers skew information away from reference and towards certain biases the moment they receive it. Just to pick some starting examples, the interfaces through which users submit information to computers are generally text-based, tactile, visually responsive, rectilinear, compartmentalized, and inflexible. A common complaint about the late internet is that most recipe "articles" are just AI-generated blogs divided into bland text blocks sprinkled with stock photos of cakes1. I will discuss the algorithmic causes of this problem later; for now, note that, even before the proliferation of AI, nearly all websites scrolled vertically, were intended for screens with four sides between a few inches and a little over a foot in size, and made users click icons and sometimes type text into horizontal fields to complete tasks. No matter the website or app, users and sites interacted through extremely similar, bureaucratic design principles.
Aside from hardware, the other main limiting factor on information is the platform that hosts it. Platforms autonomously organize, rank, and select information for clients. Because the utility of platforms depends on size and opacity—the amount of available data and the integrity of the ranking algorithm—useful platforms have far more leverage than their users, and the growth of platforms always outweighs the interests of users. A small platform that didn't prioritize growth and size wouldn't have much data to offer; a transparent one with an open-source algorithm would be vulnerable to SEO manipulation. Due to the repeating energy and storage costs of hosting data, platforms are obligated not only to profit from data, but to keep clients returning regularly for each piece of data. These factors allow platforms to extract information at an initial loss from users and capture other users in repeating transactions, rendering information incomplete, aggressive, personal, and hierarchical.
Because platforms need to continuously create transactions, both they and users have an incentive to remove reference from information. This is what I mean by "incomplete"—if clients actually got what they wanted from a particular vendor on a particular platform, they would be less inclined to add their own data or continue using that vendor. Of course, the practice of serving reduced information as advertisements predates the internet, whether in the form of movie trailers or live readings from books. What makes the internet different is the totalized, shared network. Since the internet hoards information, negates alternatives, and autonomously copies and shares it, the price of information has fallen so low that content creators must capture audiences in continuous data transactions in order to ever see material profits2. The devaluation of data also makes creators less likely to put valuable information with accurate reference online, just as the necessity for constant production and transaction requires an output above human capacity. This incentivizes the production of abbreviated, interlinked pieces of information, containing as little valuable data as possible, irrelevant enough to keep users unsatisfied but good enough to keep them buying more.
Ironically, the oversupply of information creates a buyer's market for advertisers3: as the amount of information increases, the percentage that captures consistent users decreases, allowing platforms and marketeers to build ever-more accurate taste profiles4. Advertising gets increasingly lucrative while content creators become more desperate for payment from advertisers5. This is why every miniseries is just a movie bloated with extra scenes and ads every few minutes, and why every Substack is just a list of recommendations for other Substacks, with a rant written in ten minutes stuck in the middle6. The information arrives both stripped of meaning and padded with non-meaning.
Such information needs to aggressively acquire attention as soon as a user sees it7, due to the tendency of platforms to serve as much data as fast as possible to users, reducing each vendor's chances of making a transaction. Though platforms don't care about vendors, they do want users providing more data; to this end, their algorithms select and promote content that matches quantifiably aggressive qualities8, keeping users logged on and responsive. Worse, the more creators understand this, the more they maximize their aggressive content. Asinine complaints about Americans being "more divided than ever" are simple attempts to comprehend this effect. Aggression is not partisan; propagandists can achieve it by making dramatic, extremist videos9, but also through bouncy children's cartoons10 or apolitical "hot takes." The end result is the imposition of epic conflict narratives on a world in which no narrative exists.
If aggression enhances virality, personalization defines it. The unique control platforms have over the information market allows them to target private identities with public information. This is why the "rabbit hole" effect works, not just because platforms confirm user bias, but because users need to see themselves reflected. It is true that media that feeds particular egos succeeds among its audience with any technology. The internet is different both because it provides instant access to all egocentric information that exists at once and because the internet separates the immaterial elements of identity from the material context of that identity11. Fascist conspiracy theories, new-age theosophy, and diagnosis culture all predate the internet, but were all relatively non-normative subcultures in past decades12. They have become much more normative digitally13.
The internet has also been instrumental in both the ascendancy of the alt-right and the deterioration of progressives, because platforms favor a hierarchical, traditionalist worldview. The privilege required to create high-cost, complex networks, the profiling of users as stereotyped consumers, and the removal of critical thought from automation all work to entrench hierarchical society within computer networks. The earliest network, ARPANET, was developed by the Department of Defense, arguably the most conservative sector of the US Government14. The earliest platforms were created by people with the privilege and cash surplus to first develop their server hardware and technical expertise, and then invite users to their low-cost platforms before ever seeing any return15. Developers not only code biases into technology, as notoriously illustrated by racist facial recognition technology16; they replicate them in the precepts of their work: that all faces must be catalogued to assist police, that all genders17 must be properly identified and profiled, or that all disability must be cured by technology18. Giving people of color scholarships to go study STEM or Deaf people jobs enhanced with iPads changes nothing, because the hierarchy that would exploit them is embedded in the same self-augmenting network that demands their loyalty.
In theory, the internet could have been a source of free, trustworthy information; in practice, it commands exorbitant costs for devalued data. For all the dreams of Mastodon, Cara, Duckduckgo, and other likeminded platforms, it is quite infeasible to create a platform with the neutrality of material technology, limited as that may be, because the platforms themselves are data, not technology. They only continue to exist by accruing users and harvesting profitable information from them.
Essentially, the features of virtual information—its form, structure, bias, and references—do not stay static. They evolve based on a pattern, and that pattern has a logical conclusion.
Part 5 will be released tomorrow.
Citation: look up "carrot cake recipe," click the first two links, and read them all the way through, if you can.
A data transaction being any kind where the user contributes information upon receiving the practically free commodity, by commenting, liking, or even just engaging with the commodity for a period of time.
While all internet content, under this theory, is more like advertisement than analog content, data can be classed as ads based on the mediation required to force audiences to receive them. For example, many podcasts use Patreon gimmicks to call for donations, but even listeners who don't use Patreon still listen to the podcasts. Stamps.com, however, has no worthwhile information; a podcast devoted to them would have no subscribers, so they mediate their data via podcasts directly to listeners.
It is true that the advertising industry is far less effective than, well, as advertised, a situation made even more ironic when platforms exploit both advertisers and users. This does not negate the fact that immaterial hosts are far cheaper than material hosts, that users promote advertisers by promoting their own, slightly more informative work, or that the advertising industry remains integral to sales.
Doctorow, Information, 56
Except this one, which is just a rant.
I am not using "aggressive" in its common form as a synonym for "angry." I am using it as an emotionally neutral term for content designed to attract attention, which is indeed often, but not always, "angry." An unbelievable feel-good story is just as aggressive as a video of a beheading.
A. W. Olheiser, "It's Too Late to Stop QAnon with Fact Checks and Account Bans," MIT Technology Review, 26 July 2020, https://www.technologyreview.com/2020/07/26/1005609/qanon-facebook-twitter-youtuube/.
Cade Metz, "Feeding Hate with Video: A Former Alt-Right YouTuber Explains His Methods," The New York Times, 15 Apr. 2021, https://www.nytimes.com/2021/04/15/technology/alt-right-youtube-algorithm.html?searchResultPosition=1.
Jia Tolentino, "How Cocomelon Captures Our Children's Attention," The New Yorker, 10 Jun. 2024. https://www.newyorker.com/magazine/2024/06/17/cocomelon-children-television-youtube-netflix.
I'm not saying that the internet doesn't provide valuable information about identity or help users explore identity, but it does so with the intention of severing identity from the physical person in a way that ultimately enables repression and control in the long-run, even if it is liberatory in the short-term.
The former two out of the mystique of exclusivity, the latter out of stigma.
Unlike my other two examples, which construct identities out of fake ideologies, mental disorders are very real experiences that lend themselves well to personalized commodification. They are experienced as isolated, exclusive hells that no one else shares, despite this not being the case. It should come as no surprise, then, that TikTok shorts about anxiety or comedy specials about depression are reliable hits online, regardless of how well they represent the actual experiences of diagnosis and disorder.
Neal Agarwal, "Internet Artifacts," neal.fun, https://neal.fun/internet-artifacts/arpanet-map/.
Cory Doctorow, "My McLuhan Lecture on Enshittification," Pluralistic, 3 Jan 2024, https://doctorow.medium.com/my-mcluhan-lecture-on-enshittification-ea343342b9bc.
Meredith Broussard, More Than a Glitch: Confronting Race, Gender, and Ability Bias in Tech (Cambridge, Massachusetts: The MIT Press, 2023), 31
Or just "both" genders; it depends on the interface. Of course, two genders is just one bit of RAM. Very efficient.
Ashley Shew, Against Technoableism: Rethinking Who Needs Improvement, (W. W. Norton & Company, 2023), 8.