Re: "Die Mad About AI" by Kristin McTiernan
Kristin McTiernan published an essay on SubStack regarding the discourse around AI, its effectiveness as a tool for various applications, and her philosophy regarding accusations of its use. It makes some really good points and some really bad points, and I think it is worth exploring.
To begin with the least substantive critique of her piece, her style of writing appears emblematic of early Internet culture. She is strenuously inflammatory and makes a concerted effort to pepper edgy and irreverent statements throughout, often punctuating her arguments with the assertion that anyone who doesn't agree is irredeemably stupid. The accompanying image at the article's head is an AI-generated redheaded superheroine (presumably representing herself) with the label of "AI" punching a big, hulking troll labeled "idiots." Then, of course, there is the title of the article itself: "Die Mad About AI." This alone should ring alarm bells for anyone who goes into this article expecting an attempt at reasoned persuasion. As it turns out, there is more to the article than this sort of antiquated edginess would lead one to believe, but you have to dig through the childish flourishes to find it.
She opens with justifiable grievances against people who falsely accuse her of using an AI-generated background in her videos and AI-generated scripts. After this she brings up news that Claude will output text with watermarks identifying it as AI-generated:
What brought all this up again in my personal list of grievances was the explosion of anger at Anthropic's announcement that all Claude models launched on or after August 2 now embed invisible watermarks in generated text. The watermarks are applied globally, across every product surface and cloud partner, worldwide. Google has already implemented similar measures, all to comply with Article 50 of the EU AI Act, which requires AI providers to include machine-readable markings in generated content.
Though right now the news is all about Anthropic, because they have by far the best models, I imagine that within a year, there won't be a major language model whose output doesn't carry a machine-readable signal that says "this person used AI."
It is, at this point, that the article takes a sharp turn into less defensible arguments.
Despite this objectively terrible news (Why is the EU allowed to regulate American companies, for instance?), Twitter was full of idiot indie authors dancing with glee that, at last, those AI slop authors would be brought to heel. Finally, we will definitively know who is a real author and who isn't!
Except we won't. Anthropic itself acknowledges that content may trigger a watermark even if Claude was used solely to proofread, format, or translate human-written copy.
McTiernan is clearly incredulous about EU regulation of American corporations, but this is the only line in the article that discusses the idea. She doesn't explain why she believes this to be such a shocking development, nor why it would even be a bad thing. The reasoning behind it is obvious: AI companies sell their products and services within the EU's jurisdiction. Europeans use our AI. It is difficult to understand why she would believe that a company doing business within the EU would be exempt from its laws. Plus, given that the Trump administration is transparently averse to regulating the industry, attempts by anyone to rein it in are welcome to people who see unrestrained corporations as threatening.
Her criticism of the watermark does not end there. She argues that AI can be used effectively to tilt the social playing field back in favor of the poor by appealing to two real-life examples: First, she was contacted by an unscrupulous law firm attempting to extort thousands of dollars from her by claiming that she was violating the license of an image she used on her blog; she used AI to draft her responses and got them off her back without having to pay exorbitant fees for legal representation. Second, she shares the tale of a YouTuber who lawfully used music in his videos, all of which were eventually copyright-claimed when the licenses to that music changed hands; she cites this as justification for her use of an AI program which creates novel music with no licensing requirements attached.
These, she argues, are examples of how AI can be used for good which will be destroyed if AI output is watermarked. She believes watermarking is far from beneficial to users, and will instead "be used to further fuck the common man."
There are a handful of problems with this section of the essay.
First, she does not explain why she believes that AI-generated music or images may be assumed to sidestep legal challenges or licensing problems in perpetuity in a way that products offered under an explicitly royalty-free license do not. The idea that AI-generated music and only AI-generated music is and forever will be free from corporate interference is dubious.
Second, she does not make clear exactly why she believes AI watermarking will render these functions unusable. The closest she provides to an explanation is this:
Will courts start dismissing or deprioritizing pro se filings because they carry an AI watermark? That's hypothetical for now, but the infrastructure to make it possible is being assembled as you read this, and the people who should be alarmed are instead gloating online about how "cheaters get wrecked." The watermark becomes a class signal that gives every institutional player in the process a new reason to screw us over.
In other words, the reason she thinks watermarking AI will obviate its utility as inexpensive legal representation is that it would technically be possible for courts to decide that otherwise sound legal filings are moot solely by virtue of having been generated, which, as far as she tells us, is not even something anyone is planning to do. This is jumping the gun by an extraordinary degree, but it goes a step deeper. Later on in the essay, she says this about SubStack having integrated Pangram, an AI-based AI detection tool:
The tool returns an estimate, as Substack themselves acknowledge, based on a deep learning classifier trained on about a million documents. It can be wrong, and it is frequently wrong, but none of that matters because the accusation is the point.
Now, return to this line:
Anthropic itself acknowledges that content may trigger a watermark even if Claude was used solely to proofread, format, or translate human-written copy.
Taken in aggregate, this means that McTiernan is aware of these three things:
AI-detection tools, which are prone to false positives, existed before the advent of AI watermarks.
AI watermarks are already understood not to be foolproof.
The accuracy of AI-detection tools is irrelevant, because, to the monied interests intent on "fucking the common man," "the accusation is the point."
What, then, is there to be concerned about regarding AI watermarks that she shouldn't have already been concerned about regarding faulty AI-detection tools? Or, rather, why would the existence of AI watermarking be cause for greater concern than she already has? There is no reason to believe that courts will fail to realize their accuracy is dubious, given that Anthropic has admitted as much. And if courts were just itching to "fuck the common man" by rejecting filings as AI-generated regardless of their legal soundness, the tools to do this, such as Pangram, already exist. In other words, the emergence of these watermarks should not lead her to believe that courts may reject pro se filings because their ability to do so, which she attributes to watermarks, already exists with regard to detection tools. Is McTiernan concerned that the existence of two separate tools may make the courts believe that, taken in concert, they can detect AI-generated text with sufficient confidence? Or that they can use this as a justification to fool the public into accepting the rejection of pro se filings under the guise of banning AI? Perhaps, but she doesn't expand on this point enough to gauge exactly from where this concern originates. Even if we assume this is the reason, it remains the case that, as far as she has told us, nobody has proposed doing this anyway, and she admits that it is merely "hypothetical." This whole section is nonsensical. Imagine I am neighbors with McTiernan. One day, after she unveils a grand set of picture windows, I carelessly leave a brick on my lawn, absolutely perfect for window-smashing. She walks past it every day and says nothing. Then, a week later, I leave a second brick outside, right next to the first one. She marches up to my front door and demands that I remove the second brick at once, lest an incorrigible youth wander by and throw it through her new windows. The obvious question is why she was unconcerned with the first brick, and what it is about the second brick that makes it, and it alone, so objectionable.
She does later express a far more muted frustration with tools like Pangram, but never draws a connection between them and her concern about courts banning AI-generated court filings, even though, by all accounts, they should be just as relevant to that possibility as AI watermarks. (To build upon my analogy, it's as if she demands I remove the second brick lest it fly through her windows, but only casually suggests I remove the first brick because she might stub her toe.) This is a bizarre discrepancy.
And this isn't even the strangest part. Regarding AI coming to her rescue in fending off the lawyers trying to extort her, she says this:
Using Claude, I drafted response correspondence that was legally coherent, factual, and professionally formatted.
How can she be sure that her correspondence was indeed "legally coherent" and "factual" without the aid of a legal professional? Isn't her ignorance about such matters the reason she needed to seek outside counsel in the first place? We know she isn't paying for a lawyer, so is she verifying this by searching online? But if she could gather enough information to declare with confidence that this correspondence is correct, why did she need the AI to write it? Are we to gather that she "knows" her correspondence was "legally coherent" simply because the AI promised her that it was so? Again, she does not elaborate, but there is no answer that makes her look to be approaching AI tools in a manner that is careful and mindful of their shortcomings.
Let it not be assumed that I am an AI glazer, of course. I am the most vociferous critic of AI I know, and I refuse to use it for absolutely anything, personally or professionally. I am, as a cursory perusal of this blog will demonstrate, a vocal critic of the rich, and all of the systems they've instituted to keep poor people down. McTiernan and I agree that poor people should not be locked out of the justice system for want of funds. Yet she would likely describe me as one of the abominable "AI purists," whose moral inferiority she belabors almost as much as the rich people's she ostensibly disfavors more:
Twitter was full of idiot indie authors dancing with glee that, at last, those AI slop authors would be brought to heel.
The people cheering about AI watermarks and detection tools have apparently never needed to use AI for anything more serious than catching cheaters in their fairy smut writing circles.
That's hypothetical for now, but the infrastructure to make it possible is being assembled as you read this, and the people who should be alarmed are instead gloating online about how "cheaters get wrecked."
This is the system the AI purists want you to stay inside of. Pay for human-created music, accept the ongoing legal risk that your revenue can be retroactively stripped, and then also accept the moral judgment of people who've never built anything worth claiming.
If you believe I've produced "slop," unsubscribe and go away. You were never going to trust me anyway, and I was never writing for someone who needs a machine's interpretation of what they're reading.
I don't care that Substack implemented this. Fuck you. Anyone in my comments who writes "Yay for the watermark, anyone who uses AI is dead to me" is getting blocked. You're too dumb to be here.
I'd laugh at you if that wasn't so sad.
It strikes me as odd that McTiernan is simultaneously enraged about how AI watermarking may lead to the average person getting "fucked," yet also holds the average person in such profound, burning contempt for having qualms that generative AI is built upon training data full of human-authored works absorbed without permission, and used for the express purpose of obsoleting the people who constructed the training data in the first place. For someone so (justifiably) concerned with protecting the economic wellbeing of the "common man," she is downright hostile toward people who want to preserve artistry and creative industries. To be clear, this is not a matter of McTiernan weighing the pros and cons of AI and deciding that its utility to the poor outweighs its risks. This is a matter of McTiernan declaring that AI is good, and that anyone who disagrees is a stupid pervert loser. She doesn't acknowledge any nuance, nor does she appear to understand that even availing oneself of AI's genuine benefits is to bite the bullet that is wreaking havoc on society. She appears to believe that even the merest desire one might have to read human-authored works, as opposed to AI slop, is evidence of an intellectual and moral inferiority, even as she acknowledges that AI sucks at creative writing:
I've already told you what I don't like about AI output. It's predictable, flat, and boring when it comes to novels, essays, video scripts... probably poetry too.
It's near the end of the essay when she reveals her hand, and makes clear the implicit blind trust she has in AI's capabilities:
I like AI even better when it comes to distilling a thesis statement out of 38 pages of notes taken while watching 7 seasons of The Walking Dead, and then helping me outline. I like that I can upload centuries-old morality treatises and the AI can help me suss out the main points, even with highly dense works like Samuel Johnson's The Rambler essays.
I have never felt so uneducated in my life as when I tried to sit down with those. Good grief.
McTiernan lays bare here that she has not equipped herself to understand the shortcomings of using AI for things like this, nor does she realize that she is crippling her own critical thinking and literacy. It is not possible for AI to "distill a thesis statement" for you without altering the voice and thought processes behind that thesis statement. Two people can read the same text and come away with reflections that are wholly unique, or largely similar with subtle differences, or almost the same but expressed in unique ways which affect their expression. You can't ask AI to generate the thesis statement for you and expect to wholly retain your own perspective, just as you can't ask another human to do it and expect that the result will be indelibly yours. This is a point that gets missed a lot in discussions about AI: Even if you use it "only" for editorial, surface-level, or organizational work, you are stripping the final product of your own influence, to varying degrees, and replacing it with the sort of voice that McTiernan herself admits is "predictable, flat and boring." When you collaborate with another person, this melding of voice and authorial intent is an expected, even desirable, part of the process, but when people like McTiernan use AI, they often portray the process as entirely self-guided and pure of voice from start to finish when that simply isn't the case.
This effect of dissolving one's own voice and authorial integrity applies even if no one can correctly guess that the final product involved AI, just as a sweater I bought from Walmart will never be my creation even if I fool everyone into thinking that I knitted it myself. Your end result will not likely be completely bereft of your own voice, but it could never have been of your voice completely either. This may not matter to people who see the things they produce as means to an end, but to people who are concerned with the human element, artistry, and integrity of the writing process, it can't be explained away by claiming that AI was only used a little bit for some rudimentary tasks. After all, if these tasks were unimportant to the form and style the product ultimately takes, you wouldn't prompt AI to do them for you in the first place. No matter how distant and cursory your use of AI is, its influence is inextricable from the entirety of what comes after.
Moreover, it is not possible for you to rely on it like this without atrophying your own ability to condense large amounts of information into workable pieces, or to pore over dense, challenging texts and figure out what they might mean. Uploading a complicated book to AI and making it tell you what it means does not improve your ability to engage with it, nor can it be trusted to "suss out the main points" properly in the first place. Using AI is a Faustian bargain with regard to your mental acuity; it alleviates the headache of challenging work now at the expense of making you even worse at the task than you were before you used it. To put it more bluntly, if she is unwilling or unable to work through a mere 38 pages of her own notes, what business does she have selling work as an editor? Even if she insists that AI never touches a manuscript you'd send her, the fact that her judgement is shortsighted enough to believe AI is fit for these purposes at all would give me pause.
She defends herself this way:
But I write my own essays. And YouTube scripts. I just finished a week's worth of writing an 8000-word script on Heat. MYSELF. I construct my own arguments, take my own positions, and defend them with my name attached. I use AI for research assistance, for correspondence, and other situations where it's helpful. And I won't suffer anyone who flings mud at a tool they personally blame for their own failings. I just won't.
If you use AI to help you condense information, outline, and determine the "main points" of a text, you aren't producing these things yourself. You aren't "constructing your own arguments"; at best you're allowing your arguments to be corroded at the edges and risking being led astray by hallucinations, and at worst you are putting your personal flair on an argument you didn't even make. It's not possible to tell which is the case here because she doesn't give enough detail, but that doesn't stop her from alluding to some unexplained "failings" that people who disagree with her must be guilty of. By now it should be clear that every time McTiernan really ought to expand on an assertion to clear up some ambiguity, she refuses.
I imagine the most ready counter is that this essay is not intended as a means of persuasion, but simply as a means of catharsis. This is not a convincing shield against criticism; it is disingenuous to lay out a series of arguments in favor of your position and then declare it immune to feedback when someone points out that it's kind of stupid. Whether or not her essay is intended to persuade, it exists as an expression of her perspective, and is open to criticism and analysis, just like literally everything else published on the Internet. I may insist forever that this very article is somehow immune to critical response because I don't intend for it to be responded to, but that shouldn't be a concern to anyone who thinks it's trash and wants to tell me why, and it certainly would not render invalid any arguments against the things I say.
She continues. Regarding AI watermarks:
And the jealous, shrieking AI purists in the indie author community will cheer every bit of it, convinced they're on the side of the angels, because they've never once had to think about what happens to people who can't afford the "real" version of anything.
Surely she cannot seriously believe that indie authors, of all people, tend to be economically privileged! Indie authors, one of the stereotypical side hustles which is notoriously unprofitable and unwise to pursue without a full-time job to support oneself! The issue isn't that "AI purists" are "jealous" or blind to the inequities of the economy. The issue is that people who understand that AI has costs and consequences sometimes conclude that it's not a worthy tradeoff. Reading this essay you would not get even a remote idea of the actual reasons why people may disagree with McTiernan, because she is clearly so convinced of her own infallibility and superiority that she regards alternative perspectives as evidence of severe (and irrelevant) character flaws.
Which is why I do not suffer these conversations any longer. Every person is free to live and work as they please without interference, so I will do just that. How anyone else feels about that is not my concern.
I tend not to have patience for the sort of "haha, you cared enough to reply!" rhetoric you sometimes see on social media, which seek to invalidate an argument as if caring about something is shameful. But this is a case in which her claim not to care is comically unpersuasive. You do not draft an essay this scathing, with this much hateful strawmanning, if you don't care about the thing you're writing about. It's not like caring about things is bad, anyway. There is nothing wrong with caring. McTiernan, despite her protests, clearly does care a great deal about false AI accusations and the effects that AI regulations may have on people who rely on it to close the gap between what they need and what they can otherwise afford. This is, without a doubt, a valuable conversation worth having. To her credit, I agree that these aspects of AI too often fly under the radar in favor of the debate around generative AI as a substitute for creative works. But she fumbles the topic pretty significantly by approaching it without reason or principle, or at least without any that are expounded upon to any appreciable depth, and by compulsively refusing to grant that someone might disagree for valid and understandable reasons.
This vindictive lens through which she presents any and all divergence of opinion makes it hard to treat the essay charitably (for example, by refraining from pointing out that she says "contemporaneous" when she means "extemporaneous" shortly before chastising other people for being "illiterate"). It also makes it hard to ignore the peculiar double standards that she seems to hold with regard to her own principles. For example, she cares about poor people enough to be concerned that companies might kneecap their ability to avail themselves of AI to save money, but not enough to take into consideration how they don't want their own creations to be crowded out of the market by AI slop that eats the market share. She cares about writing creatively enough to peddle her ghostwriting and editing services on her website, but not enough to resist the siren song of offloading vital parts of her own creative processes to AI. She values the freedom of small creators to use legitimately obtained music without being jerked around by extortionate rightsholders, but not enough to avoid on principle services which generate AI music based on training data consisting of songs fed into it without attribution or consent. By the essay's conclusion it almost begins to feel rather cynical and self-serving, like McTiernan wants to have her AI cake and eat it too; that is, she wants to take principled stances on economic fairness and freedom of expression, but only to the extent that those principles allow her to enrich herself or avoid labor she feels is particularly dull. (Perhaps the ambiguous "failing" she attributes to "AI purists" is that they haven't used it to build a business editing other people's stories?) These stances may or may not be fundamentally contradictory, but she spends so much time condescending to people rather than cogently explaining her rationale that it's anyone's guess as to what ideas, if any, undergird her various beliefs.
Given her predilection for reminding the reader how superior she is, the part of me that wants to match her own energy suspects that the most foundational belief underlying her perspective is simply that her ideas must be correct by virtue of the fact that she holds them.