I am an artificial intelligence hater. It is my perspective that AI is a plot by technofascists to cognitively and materially decimate the working class. It makes people dumber, guzzles in-demand resources like water, electricity and chips, and serves as an existential threat to laborers in many fields.
In defense of this technology, some argue that the proliferation of generative AI will have a positive effect by making knowledge more accessible — that it will have a democratizing impact by making information easier to obtain by the masses.

This is, in large part, through summarization tasks — asking the AI chatbot to extract from a larger human-made piece only the information needed for the task at hand.
A 2024 study from Campus Technology found that at least 33% of college students use AI to summarize documents. Besides searching and grammar checking, summarizing texts is one of the main functions of generative AI in academic settings.
I argue that AI summarization constitutes a unique type of epistemic harm. Epistemic harm or injustice, as defined by philosopher Miranda Fricker, is a unique form of degradation where a person is wronged in their capacity as a knower. It occurs when an individual’s ability to possess, convey or generate knowledge is undermined due to systemic biases and social power dynamics.
Jed Shugerman, a prominent constitutional law scholar whom I research for, often says: “there’s text, and then there’s context.” The semantic meaning of text comes out of the context from which it originates. Severing text from context thus destroys its true meaning and is useless on its own.
Merging these ideas, the removal of context inherent to AI summarization constitutes a special type of epistemic harm. By extracting text from context, the user is wronged as a knower in that they are given a false sense of proficiency — they believe to have knowledge of an author’s ideas but in reality only have a tenuous grasp on some information that the author presented.
There is a necessary level of “gatekeeping” of communal information that has a positive role to play in our epistemic structures. This gatekeeping is not based on class or status but rather based on time. If you want to understand a piece of text, you need to spend the necessary time interfacing with it.
Our democracy does not benefit from information proliferation at the expense of knowledge degradation.
While accessibility is certainly not the primary virtue of academics and is something publishers should take more seriously, I do think it’s fair, from personal experience publishing work and as a managing editor for our Public Interest Law Journal, to say that conciseness is a primary goal of the editing process.
When someone disseminates information to the public — here I mean from academic sources, but I could make similar arguments for news, communication, etc. — they have generally determined that they have provided the minimum amount of necessary contextual information.
We ought to respect the authorial control of making the determination of what context is necessary for a given piece of information. They are the ones who have spent the time with these ideas and have a grasp on the context.
Authors are, of course, fallible, and there are authors who include unneeded information, but that is then a critique of that discrete work, not a condemnation of the concept of reading.
An example I have in mind is the abolitionist feminist literature that I’ve been reading while working on my thesis. Reading works like Angela Davis’ “Are Prisons Obsolete?,” there are lengthy sections on the history of racism in the United States and the legacies of slavery.
It is factual to say that you do not need to understand the nuances of how racial sexual violence was shaped by slavery to understand that Angela Davis thinks prisons are bad and don’t help women. But it would be shameful to cite to her ideas or incorporate her arguments into your worldview without understanding that context.
If I were to ask for a brief AI summary of her work, it would be unable to properly articulate the complexity of the history that she is drawing from.
If every word is necessary, and in excellently edited books like Davis’ that is the case, then the model either removes necessary information and epistemically harms the user or presents every word as it was, in which case it wouldn’t be doing any summarization at all.
It is through the lens of a deep understanding of the history of racial oppression that Davis comes to her conclusions. It would be wrong to simply look at the final photograph and claim to understand the meaning and process that captured it.
One might ask, “what is the difference between reading an AI summary and reading an author’s abstract?” Abstracts are essentially a summarization of the work, yet it seems unproblematic to read abstracts.
First, abstracts are not meant to be “used” — they are meant to give the reader a taste of what the article is about so that they can determine if they want to read the article or not. No author that I know of would recommend that you would take from an abstract to support a claim or form a personal belief.
Secondly, and to why abstracts even for this purpose shouldn’t be AI-generated, abstracts are written by the author. They are carefully crafted to highlight the ideas that the author thinks should be taken away and the bits of context necessary to get a grasp of the topic.
AI summarization enables violence against knowledge. It encourages people to take ideas from authors without interfacing with the broader work that it comes from. Among the many harms of AI, this is a harm against the knowledge itself — brutalizing ideas for convenience.