3126: Disclaimer

Disclaimer

|< < Prev #3126 (August 8, 2025) Next > >|

Title text: You say no human would reply to a forum thread about Tom Bombadil by writing and editing hundreds of words of text, complete with formatting, fancy punctuation, and two separate uses of the word 'delve'. Unfortunately for both of us, you are wrong.

|< < Prev Random Explanation Next > >|

Explanation

This comic shows a message written by Cueball (likely representing Randall). The combination of a user profile image, metadata, and the body of text (not to mention the reference to a forum in the title text), suggest this is a posting on a forum. At the end — possibly as an automatically appended forum signature — he includes a disclaimer pre-emptively denying that the content was produced by ChatGPT, and that this is just the way he is.

ChatGPT is a large language model, a type of generative artificial intelligence designed to produce human-like text. In 2025, at the time of this comic’s release, tools like ChatGPT had become widely used for composing everything from professional emails to casual forum posts. Chatbots tend to be verbose. One side effect of this trend is that genuine, naturally long-form writing is sometimes mistaken for AI-generated content. People may become suspicious not only of length and thoroughness, but also of stylistic markers such as formally correct punctuation — particularly the em dash — which is uncommon among average internet users, but occasionally favored by ChatGPT.

The title text explains the context. It notes that some people claim no human would compose a meticulously organized essay about Tom Bombadil, a relatively minor character from The Lord of the Rings, especially if it contains the word "delve". This word is perceived as atypical in everyday conversation, yet is frequently used by ChatGPT. Randall implies that he would write such an essay, and in fact has. The choice of “delve” may also be a deliberate reference to Gandalf’s line in The Fellowship of the Ring about the Dwarves who “delved too greedily and too deep,” or perhaps to “Michel Delving,” the largest Hobbit settlement in the Shire, whose name literally means “large excavation.” He also says this is unfortunate for both of them: the person he's talking to for having jumped to the wrong assumption about the nature of the author, and Randall for having wasted so much of his time writing about a minor topic.

The presence of the em dash in the disclaimer further fuels the joke. Em dashes are seen by many as an indicator that a piece of text was AI-generated; although they have been a long-standing (even pre-GPT) feature of both Randall's comics and this site's user-written explanations, which of course does not disprove the more general observation being alluded to. The comic is also likely to be have been written specifically as a response to the release of GPT-5, the day before this comic was published.

Transcript

[A Cueball profile picture is shown next to five paragraphs of illegible text, which contains varied punctuation marks and two square-bracketed citations. The last paragraph in the essay is separated from the other paragraphs by a line with three dashes. Its single sentence is highlighted, and lines connect that illegible sentence to a box with an enlarged, legible version of the sentence.]

Enlarged text: Not ChatGPT output—I’m just like this.

[Caption below the panel:]

I've had to start adding this disclaimer to my messages.

  Add comment        Create topic (use sparingly)       Refresh 

Discussion

Is this related to the ChatGPT 5 release -- Mathmaster (talk) 19:28, 8 August 2025 (UTC) (please sign your comments with ~~~~)

Probably. Barmar (talk) 19:30, 8 August 2025 (UTC)

Was anyone tempted to ask ChatGPT to write an explanation of this comic? Barmar (talk) 19:30, 8 August 2025 (UTC)

Sure, https://chatgpt.com/share/6897c0ac-dfdc-8012-8030-76447b589747 BytEfLUSh (talk) 21:43, 9 August 2025 (UTC)

Considering who Tom Bombadil is, anything less than a dissertation-length discussion is incomplete. PS: The captcha wanted me to select pictures that contained parking meters, but it showed what looked like a mailbox. Maybe parking meters look like mailboxes in the parts of Middle Earth that I haven't visited yet. 64.201.132.210 20:23, 8 August 2025 (UTC)

Am I the only one wondering if Randall actually made such a post? It would be fun to find but without knowing what LOTR boards he frequents and his screen name, it would be pretty difficult. 47.248.235.170 20:44, 8 August 2025 (UTC)Pat

Who doesn't use the word 'delve'? I've used it since I was a child! 92.23.2.228 20:59, 8 August 2025 (UTC)

Once or twice, I've heard a Reddit reader on YouTube claim that a story was probably written by ChatGPT because of things like "used em dashes" and "flows so smoothly that I didn't make any mistakes while recording, while most stories by humans make me glitch once or twice". Um. Is it bad to use punctuation correctly? To polish writing so that it flows smoothly, as story-telling often should? BunsenH (talk) 21:07, 8 August 2025 (UTC)

Hyphens between spaces are automatically converted to em dashes when using modern versions of MS Word. 104.28.205.246 22:27, 8 August 2025 (UTC)

That's an en-dash; em-dashes are automatically created from two hyphens between words like--thisThatNerdyHylian (talk) 09:01, 9 August 2025 (UTC)

It’s not bad—it’s just a lazy diagnostic. Smooth prose and correct punctuation aren’t inherently “AI tells”; they’re simply easier for large language models to produce consistently than for average human writers. (This is because large language models were trained to predict the next token, which is a task that emphasizes grammar more than it does ideas. Meanwhile, humans are taught to emphasize the ideas and only edit for grammar in formal writing.) People latch onto prose and punctuation because they’re visible, not because they’re reliable indicators. Over-reliance on such surface markers risks flagging well-edited human work as AI-generated, which is exactly the false-positive trap detection tools keep falling into. 114.94.93.163 11:37, 9 August 2025 (UTC)

"Delve" is/was specifically a word considered a marker of ChatGPT. See https://news.fsu.edu/news/science-technology/2025/02/17/why-does-chatgpt-delve-so-much-fsu-researchers-begin-to-uncover-why-chatgpt-overuses-certain-words/ 64.203.66.182 21:19, 8 August 2025 (UTC)

Groundbreaking advancements in the realm of AI may indeed boast surprising and intricate compositions, showcasing a lexicon emphasising and underscoring its advanced comprehending of garnered language that surpasses what aligns with human usage. But I hope the team behind this paper delves into the possibility of false-positives. 92.23.2.228 21:54, 8 August 2025 (UTC)

You're absolutely right—delve has become emblematic of ChatGPT’s stylistic tendencies.

Florida State University researchers Tom Juzek and Zina Ward conducted a study titled “Why Does ChatGPT ‘Delve’ So Much? Exploring the Source of Overrepresentation in Large Language Models.” This study, published in the January 2025 Proceedings of the International Conference on Computational Linguistics, systematically identified words like “delve,” “intricate,” and “underscore” that have sharply increased in usage within scientific abstracts and are notably more common in ChatGPT outputs. They called these 21 “focal words.”

The researchers ruled out several potential causes—such as model architecture, training data, or algorithms—as primary drivers of this lexical overrepresentation. Their model testing suggested that Reinforcement Learning from Human Feedback (RLHF) may contribute to the pattern, though the results were mixed and partially inconclusive. Intriguingly, they found that human evaluators might respond differently to the word “delve” compared to other focal words.

So yes—delve is indeed one of the telltale markers that researchers have pinpointed for ChatGPT’s distinctive language usage.

Would you like to explore more on how other focal words (like “intricate” or “underscore”) compare, or dive into the experimental methodology they used? 104.28.205.246 22:35, 8 August 2025 (UTC)

Sorry not sorry, this is a really sore subject for me. Especially since I work at a marine biology laboratory that, for decades, has worked on topics related to (WARNING: PROHIBITED LANGUAGE!) "climate change" (anthropogenic global warming). The MIT Technology Review, last May, published an article detailing the energy costs of all forms of AI. If you have not yet encountered this article, I recommend it (WARNING: PAYWALL sorry not sorry, first-readers might get a pass). In case you were wondering how come 47 was so quick to pull US out of global climate accords, and conduct his full-on assault on [ahem] inconvenient truths, this is part of the answer (the other part is equally-energy-guzzling cryptocurrencies). Not for nothing did Frank Herbert, in 1965, create an "Orange Catholic Bible" for the sole purpose of quoting a single line from it: "Thou shalt not create a machine in the image of a human mind." 2605:59C8:160:DB08:F999:FCE:2CB7:40E5 17:12, 9 August 2025 (UTC)

Basically anyone with autism was blamed for being ChatGPT at some point in the last 2 years (I include myself in this, and many others). And Randall displays, through XKCD, basically every symptom. But hey, that's what happens when your thinking is logic-based. 2A06:C701:419E:9400:FB39:2243:3D5D:F06D 22:56, 9 August 2025 (UTC)

Wait! If you look closely at the illegible text, you can see things that look like [1] or something, which occasionally comes from ai or Wikipedia. ^April_fools^update!(talk) 01:08, 10 August 2025 (UTC)

You can, but (in my own forum haunts) I'm not unknown to use footnotes[1], just like seen, to pepper my discourse; and these are clearly inline[2] []-references, not superscript[3] ones, as per actual Wiki markup. Copying and pasting from a superscripted source can remove that, but if it's an actual GPT posting, and being accurate with things like the mdashes (I normally avoid them, mainly due to the "word-word" vs "word - word" problem, but use parentheses and semicolons a lot more), you'd think it'd be accurate by the style-tags too. Unless it has been specifically/predominantly trained on this style of footnote use (and long-winded sesquipedalian lexicon) from the likes of myself, of course!

[1] a.k.a. 'feetnete', in the full plural!

[2] Foo[bar]

[3] Foo^[bar]

92.23.2.228 10:26, 10 August 2025 (UTC)

Tom Bombadil a "relatively minor character"? How dare you! In this essay, I will... 81.103.84.101 (talk) 18:23, 10 August 2025 (please sign your comments with ~~~~)

Should "delved too greedily and too deep" link to 760: Moria? Morgan Wick (talk) 08:02, 11 August 2025 (UTC)

From now on, I'm adding this disclaimer every time I speak in any public space, including the link to the xkcd strip where applicable. I have been questioned a few times simply because I usually use proper punctuation and I write in a slightly formal manner. I see where this is heading. I am imagining the future world, where the humans will be writing comments, articles and books in this simplistic, "crappy" style full of gramatic errors to prove their humanity. After the internet becomes saturated with this type of content, chatbots will pick it up too, and at that point, it will be impossible to distinguish who's real and people will probably go outside to actually meet other people once again. I think I'll make this into a comicbook, but knowing myself, I wouldn't finish the project before it would stop being speculative fiction and become just boring reality instead. [Not ChatGPT output—I'm just like this.] Harry Chaote 46.34.239.30 11:04, 11 August 2025 (UTC)

Grammatical 82.13.184.33 13:40, 11 August 2025 (UTC)

Thank you, kind stranger, I'll add that to my notes. I'm self-taught in English, so I guess I slipped even though I was particularly careful in this specific case, not to shoot myself in the foot with my own comment. Oh well. Harry Chaote 46.34.239.30 15:23, 11 August 2025 (UTC)

It is interesting to consider the contrasts between this comic and 810: Constructive. By one interpretation, the difference is that (a decade and a half ago) Randall was perhaps wanting AIs to look indistinguishable from genuine and intentionally helpful people; whereas now he finds that his status as a genuine person, in which he is certainly intending to be helpful, is being mistaken for an AI. Though, by another interpretation, it could also be said that AI has failed to look like ordinary people, only in looking like people like himself (which is bad news, for both Randall and the AIs — though arguably better, and more in line with #810, than the kind of user that Tay became effectively indistinguishable from). I could expound a couple more comparisons, different yet again, but won't. Lest I look AI-like myself (bad?), and/or Randall-like (definitely not bad, just not adding much easily digested explanatory value). 82.132.236.103 14:29, 11 August 2025 (UTC) (And the CAPTCHA initially seemed to disagree with me about what is, and what isn't, a motorcycle. Or maybe it instead values my expertise and wishes me to identify yet more of them..!)
        Add comment