Tag: LLMs

Posted on
by

…and sometimes, unreality is intentional and benign.

Today, Ars Technica posted an article with a header image featuring a keyboard modified to add an “Ars Subscribe” button. But the keyboard looked weird: the visible keys looked like this:

OPGU↵
LS.,
OC/[Ars Subscribe]

The extra letters, especially the two O keys, immediately drew accusations of AI slop generation.

And yet the image was credited to Aurich Lawson / Getty.

I figured Lawson had grabbed a stock image without noticing it was slop and then added the Subscribe button. So I did some reverse image searches. And it turned out that the apparent source photo was uploaded to iStockPhoto in 2016 (Getty owns iStock now), long before the advent of modern “AI” image generation.

That left the question of why the 2016 image had weird keys, though.

Then @lw replied on Mastodon that it matches a Turkish keyboard layout with the diacritics missing. I looked back at my search results and noticed another image that looked the same, only with the extra keys blanked out. The same account uploaded this one in 2018, and this time iStock had tagged it as having been uploaded from Turkey.

So, mystery (mostly) solved! It’s a real keyboard layout. Almost certainly a real photo, deliberately edited by hand to remove the extra markings. Why? Probably the original uploader figured it would be less distracting to a global audience. Who could have predicted that a decade later it would end up looking more suspicious because of it?

The discussion picked up in the article’s comment thread, where narteb found an even older version from 2014, still with the missing diacritics, but with a different action button and a darker (presumably original) skin tone on the finger.

A Proxy for Quality

It’s no surprise that readers were so quick to assume the photo was flat-out generated slop based on the apparent mismatch (not counting the obviously intentional subscribe button). “AI”-generated images have flooded the zone, and while there used to be some obvious tells — badly drawn hands, for instance — they’re getting to be a lot subtler.

Now we all second-guess everything we see, because to do otherwise would leave us open to being manipulated by falsehoods. And it’s not just knowing whether the image shows something real, but to know whether the image represents real human intention. The *ahem* key aspect of this image is the big red subscribe button, which is obviously not real, but it’s intentionally not real.

No one was under any illusion that the photo was an actual, unmodified keyboard. But when people assumed it was generated slop, they also assumed that was an indicator of laziness. When you see an article accompanied by an obviously junk image, you wonder, what else did they skimp on? Did they generate the article too? Did they research it properly? Did they fact-check it? This particular article was just a “here’s our new forum” announcement, but what about the other stuff that they publish?

Trust

Earlier this year, Ars retracted an article that contained fabricated, LLM-generated quotes. They fired the reporter and spent time developing an AI usage policy. This kind of thing burns trust, and trust is a news publication’s most important asset.

And it doesn’t even have to be a real mistake, or a real bias, to do damage. Conservatives spent decades painting NPR and PBS as some sort of bizarre “Marxist” strawman parody of what they actually broadcast, forcing them to dig out of a hole with large parts of their potential audience, until they’d lost enough trust that it became politically expedient to defund the networks.

Gut reactions and snap decisions have a purpose. But they’re still only heuristics. When we have time, it’s good to second-guess our impressions. Sometimes they’re wrong. Sometimes they’re right. But we have to look at those impressions clearly, not just spend the effort justifying the conclusion we’ve already reached. That’s something the human brain is also really good at, and when we’re cherry-picking evidence for the wrong conclusion, that makes things even worse.

I saw at least one commenter on the thread following up and insisting it couldn’t possibly be a Turkish keyboard, because those would have the diacritics on them, and this one doesn’t. It reminded me of the character in Mystery Men rejecting the idea that the Clark Kent expy couldn’t possibly be the Superman expy on the basis that he wears glasses, and not-Superman clearly has better eyesight than that.

Blaugust 2026: My Posts

Posted on
by

I recently added “Your LLM was trained on my blog” to the intro on my home page. I’m sort of being facetious, but it’s also true: my old site (hyperborea.org) has been in the Common Crawl data set for years, this blog has been online since 2002, and the major LLMs have used Common Crawl’s data in their training.

Way back when I switched to WordPress in 2003, one of the things I found annoying about it was that it would auto-convert punctuation from ASCII to “fancy” characters. Curly quotes, curly apostrophes, em-dashes instead of double-hyphens, ellipses…These were things that bugged the heck out of me when I saw them in broken emails and Word documents. Over the next year, seeing the difference it made in legibility, I came to appreciate real typography. Since then I’ve been happy to let WordPress, and later ClassicPress and Eleventy convert my plain typing.

Side note: I still have trouble using some apps with code snippets (ex: Apple Notes and Stickies), ’cause I haven’t figured out how to stop them from converting ASCII punctuation that means something in the code to Unicode punctuation that means something more specific in printed human language. Also, notice in that sentence that ClassicPress was smart enough to convert a straight up and down apostrophe used on a contraction to a curly apostrophe, and not to a curly single quote like it does with ‘here.’

Anyway, WordPress powers a huge amount of the web. I’ve seen stats bandied about such as 42% of all websites, 60% of content management systems, >50% of the top 10,000 websites (by some measure or another). The exact numbers aren’t important, it’s just important to know that a whole lot of the web uses WordPress to convert what people write into HTML pages for the web.

Now let’s get back to LLMs:

  • People think they can identify LLM-generated text because it tends to include em-dashes.
  • LLMs were trained on writing by actual people on the web.
  • A whole lot of that writing by actual people is presented by WordPress.
  • WordPress has auto-converted typography including em-dashes since 2003.
  • That means a large chunk of LLMs’ training data includes text by real people that uses real em-dashes.

“AI” doesn’t use em-dashes because real people don’t, it uses em-dashes because real people do [note: lots of swearing] (via).

The only place I can imagine it would be even a semi-reliable indicator would be in social media posts, on platforms where that kind of auto-conversion doesn’t happen. Even then, I can easily imagine someone writing a draft in Word, or copying a sentence or two from their WordPress post, or just hitting option+shift+hyphen on a Mac (or whatever the key combo is on Windows).

Posted on
by

A story’s been making the rounds about a software project that enforced a no-LLM-use policy by using prompt injection to delete itself. An “AI” agent-using coder filed a bug report (understandable), but filled it with a bunch of long-winded, clearly LLM-generated comments.

I looked at those comments. I can’t say I read them, because my eyes started glazing over a couple of paragraphs in. The contrast with the posts by the maintainer and other commenters is…stark.

Though I did notice the bit about how nobody reads the docs, which seems rather telling.

One of the problems with letting an “AI” write for you: If you aren’t reading it, and you assume the person at the other end is just going to summarize it anyway, there’s no motivation to make it readable. And no motivation to think about it and narrow down what’s important. And if you’re rewriting the prompt to focus on what matters most, consider that the prompt would get the idea across more effectively.

»All pages site-wide with this tag