AI technology thread

[Richard W]

I hope I am talking of the right technology. Is the conversion of text to speech now likely to be done by AI? I heard, at 0:28 in the Youtube video entitled "Her Family Sold The Orc Girl As Useless But I Built Her A Small Home She Could Finally Call Own", two very different pronunciations of 'wagon' in "A man started shouting as if shouting could hold a wagon on a mountain. The wagon sat crooked in the cut of the road. ". The first had the first vowel as TRAP and the second occurrence has the vowel as FACE. What's going on? It's the sort of error I would associate with a human reading out the narrative. Surely AI isn't deliberately mimicking human error?

I'm not sure the written text I am quoting was upstream of the speech in the creation of this content. The text shown on the screen isn't punctuated. Occasionally the text as read out and as displayed on the screen have a different choice of words, as though a man were reading what ought to have been in the text, rather than the text he is presented with.

The vowel lengths in 'read' and 'live' suggest to me that the text is not being parsed for reading, which argues against the text to speech converter being a human that knows English well. Some of these videos at least are composed in Portuguese and then converted to English, but the author's accent I've heard is different to what is presented in the video.

[bradrn]

All I can think of is that it takes a very special combination of interests for one to be pondering vocalic inconsistencies while watching a probably AI-generated video about orc romances…

[alice]

I noticed the same during my investigations, and I avoided films about orcs completely Here's one particularly egregious offender; one of the names seems to be /lɔra/ but also appears as /lura/ and at least two other variants, sometimes within less than a minute of each other (IIRC; it was a few days ago that I listened to it). In others, when there are English subtitles, I've noticed that every so often one of the words is spoken as gibberish but is printed clearly on the screen.

I've also noticed that the AI films on YouTube seem to be cut from very few pieces of cloth; minor variants on "they ignored the cleaner until she turned out to be a Mafia boss's daughter" and "they ignored the loser until he turned out to be a dragon god" turn up a lot.

[Richard W]

I noticed the same during my investigations, and I avoided films about orcs completely

I suspect the same production process. What looks like plagiarism might just be the same creator reusing his own ideas for different genres.

Here's one particularly egregious offender; one of the names seems to be /lɔra/ but also appears as /lura/ and at least two other variants, sometimes within less than a minute of each other (IIRC; it was a few days ago that I listened to it).

My hypothesis from the orc romances is that an outline has been dictated, which results in the proper names being handled multiple ways. I suspect there is a lot of conversion going on between audio and text forms. Curiously, there seems to be less variation in the spoken forms of the names, but I haven't investigated and may be mistaken. There seem to be some consistent errors - what I think is 'Thomas' often comes out with what may be an Anglicisation of a Spanish pronunciation.

In others, when there are English subtitles, I've noticed that every so often one of the words is spoken as gibberish but is printed clearly on the screen.

True gibberish, or just mangled? There are also many cases where a stream of consciousness seems to leak through from the AI. What I've been watching hasn't reached the level of an AI film, though there are some attempts at the starts of a few videos. I'm not sure that they're consistent with the story, though; the orc content creators have a long way to go.

I've also noticed that the AI films on YouTube seem to be cut from very few pieces of cloth; minor variants on "they ignored the cleaner until she turned out to be a Mafia boss's daughter" and "they ignored the loser until he turned out to be a dragon god" turn up a lot.

Variants of these turn up in the 'manga dubs' (terminology?), so I think it's just a dearth of originality rather than an AI artefact. The manga dubs seem to employ human voice actors, seemingly performing on a tight schedule. I'm pretty sure that sometimes the wrong actor speaks the lines, but the actors often improve what I take to be translations from Japanese. It's a shame these improvements don't feed back to the speech captions. Unlike the English language tradition, the speech bubbles aren't clearly connected to the characters.

[zompist]

I hope I am talking of the right technology. Is the conversion of text to speech now likely to be done by AI? I heard, at 0:28 in the Youtube video entitled "Her Family Sold The Orc Girl As Useless But I Built Her A Small Home She Could Finally Call Own", two very different pronunciations of 'wagon' in "A man started shouting as if shouting could hold a wagon on a mountain. The wagon sat crooked in the cut of the road. ". The first had the first vowel as TRAP and the second occurrence has the vowel as FACE. What's going on? It's the sort of error I would associate with a human reading out the narrative. Surely AI isn't deliberately mimicking human error?

I found the video, and I'd say: definitely an AI voice. It's competent, but every line is delivered the exact same way. A human actor would inject more life into the lines.

Does anyone have /ej/ in wagon? That also seems like a tell.

Variability is built into AIs... the hard part is making them consistent instead. It's not algorithmic, such that every instance of "wagon" will be pronounced the same. I'm relying on an understanding based more on LLMs, but the general idea is that AIs work as well as they do because they have a broad idea of context. They may or may not have a node which encodes each word's pronunciation: they could easily have multiple nodes for a word, or handle things based on more general correlations. The context in sentence A may not be the same as that for sentence B.

It sounds like a bug, but it's really not. Repeating the same thing every time for the same prompt strikes humans as very "computery". Plus some built-in variation helps the AI get unstuck from loops, or provide alternate solutions/wordings/art for the same prompt.

[Richard W]

I'm relying on an understanding based more on LLMs, but the general idea is that AIs work as well as they do because they have a broad idea of context.

It sounds like a bug, but it's really not. Repeating the same thing every time for the same prompt strikes humans as very "computery". Plus some built-in variation helps the AI get unstuck from loops, or provide alternate solutions/wordings/art for the same prompt.

Of course, context is what we need to know the pronunciation of read and live, but it strikes me that it is mimicking human error, sometimes pronouncing wagon from memory, and on this one occasion applying reading rules and getting the FACE vowel. And I have grave doubts about it working out when to rime leading with Reading rather than with reading.

[Richard W]

Of course, context is what we need to know the pronunciation of read and live, but it strikes me that it is mimicking human error, sometimes pronouncing wagon from memory, and on this one occasion applying reading rules and getting the FACE vowel. And I have grave doubts about it working out when to rime leading with Reading rather than with reading.

I'm beginning to wonder of these mispronunciations are deliberate. In another video, I just heard ordinariness pronounced with that 'a' surfacing as something like [ɒ], just a sentence or two after it had surfaced with something like [ɛ]. This would mimic confusion between handwritten 'a' and 'o'. A cynical view is that this is an attempt to defeat Youtube's limits on AI, like dynamic snow and panning on videos with basically a single picture. Youtube is officially requiring effort or originality for content creators to be paid for advertising.

[linguistcat]

I hope I am talking of the right technology. Is the conversion of text to speech now likely to be done by AI? I heard, at 0:28 in the Youtube video entitled "Her Family Sold The Orc Girl As Useless But I Built Her A Small Home She Could Finally Call Own", two very different pronunciations of 'wagon' in "A man started shouting as if shouting could hold a wagon on a mountain. The wagon sat crooked in the cut of the road. ". The first had the first vowel as TRAP and the second occurrence has the vowel as FACE. What's going on? It's the sort of error I would associate with a human reading out the narrative. Surely AI isn't deliberately mimicking human error?

I found the video, and I'd say: definitely an AI voice. It's competent, but every line is delivered the exact same way. A human actor would inject more life into the lines.

Also those inconsistencies in pronunciation are exactly the kind of thing I look for to avoid AI voiced videos. If one word is pronounced weirdly but it's pretty consistent in the weirdness, I give it a pass. People pronounce things differently from other people but are pretty consistent in their own speech. If the same word is pronounced various different ways by the same reader, then it's probably AI. And if a human pronounced something weirdly, I'd assume that they'd retake it and fix things in editing unless there's a good reason not to (livestream or similar). AI voicing won't, so the weird mistakes get left.

[Raphael]

Minor point: Does anyone else look at the photo at the top of this BBC piece on "10 of the most striking ocean images of 2026" and think, "If I wouldn't have an assurance from a major news outlet saying this photo is real, I'd guess it's AI"? https://www.bbc.com/future/article/2026 ... es-of-2026

[Richard W]

Also those inconsistencies in pronunciation are exactly the kind of thing I look for to avoid AI voiced videos. If one word is pronounced weirdly but it's pretty consistent in the weirdness, I give it a pass. People pronounce things differently from other people but are pretty consistent in their own speech. If the same word is pronounced various different ways by the same reader, then it's probably AI. And if a human pronounced something weirdly, I'd assume that they'd retake it and fix things in editing unless there's a good reason not to (livestream or similar). AI voicing won't, so the weird mistakes get left.

Does the desire to avoid cuts seem a good reason to not make corrections? I've been surprised by how many times fully human voice over is not corrected. It may be a difference between videos made to educate or persuade, and those made to earn money. With the latter, fixing faults might not be economic, which is a sad observation on one of the supposed merits of free competition. (An observation often implicit in the phrase 'race to the bottom'.)

EDIT: One thing I've noticed is that AI-generated speech does not split the text into chunks. This may be an additional problem for applying corrections if there were to be active editing.

If one is aiming for a continuous take, then natural error rates prevent perfection - one eventually has to abandon the attempts. I did wonder if there were a similar problem with image generation - the AI artist would just keep finding fresh errors to introduce. I do wonder if AI costs are beginning to bite, now that Youtube has said that a single image for an entire video is not acceptable. However, I fear that for narration over a series of images, the content creator just can't be bothered to act as a reviewing editor - it takes too long.

[linguistcat]

Also those inconsistencies in pronunciation are exactly the kind of thing I look for to avoid AI voiced videos. If one word is pronounced weirdly but it's pretty consistent in the weirdness, I give it a pass. People pronounce things differently from other people but are pretty consistent in their own speech. If the same word is pronounced various different ways by the same reader, then it's probably AI. And if a human pronounced something weirdly, I'd assume that they'd retake it and fix things in editing unless there's a good reason not to (livestream or similar). AI voicing won't, so the weird mistakes get left.

...
If one is aiming for a continuous take, then natural error rates prevent perfection - one eventually has to abandon the attempts. I did wonder if there were a similar problem with image generation - the AI artist would just keep finding fresh errors to introduce. I do wonder if AI costs are beginning to bite, now that Youtube has said that a single image for an entire video is not acceptable. However, I fear that for narration over a series of images, the content creator just can't be bothered to act as a reviewing editor - it takes too long.

Yeah, though that's also why I look at other things, like the number of different pronunciations, general tone and rhythm, etc If someone misspeaks once, or puts the emphasis on the wrong syLAHble, that's human error. If I hear three wildly different pronunciations for a common word in a paragraph, that's either AI voicing or VERY lazy human made stuff, both things I want to avoid if possible.
Although, I do know a WH40k channel that uses AI images for background art but the voicing (AFAICT) is by a human narrator. I'd prefer the guy commissioned someone for the art, but I also understand that the main draw is the voice acting.

[alice]

Minor point: Does anyone else look at the photo at the top of this BBC piece on "10 of the most striking ocean images of 2026" and think, "If I wouldn't have an assurance from a major news outlet saying this photo is real, I'd guess it's AI"? https://www.bbc.com/future/article/2026 ... es-of-2026

Yes. Are we getting to the stage that images are now considered by default to be AI?

[jcb]

We gave Claude Fable 5.1 an open task: solve an unsolved 370-year-old cipher. It solved it within a day.

[jal]

Watching AI-narrated technology videos is a pain. I recently watched one that consistently called "DOS" dee oh es. It was cringeworthy.


JAL

[Richard W]

Anthropic researcher believes more than 10% chance AI 'could kill all humans'. Not the first time someone has said something has said something like this; the question is: "how"?

One scenario that nearly came to pass (admitted to in the past week, nearly happening this spring) was the US navy attempting to board a Chinese vessel because they believed, because of AI, that it was transporting nuclear weapon components to Iran. If such an action led to all-out war between China and the USA, would it count as AI killing us?

[bradrn]

If such an action led to all-out war between China and the USA, would it count as AI killing us?

No, just good ol’ human stupidity. Fortunately, this time round someone was sufficiently clever to avoid it. (cf. Stanislav Petrov)

[Travis B.]

"Kill all humans" scenarios, when those positing them are pressed for just how AI would kill all humans, usually involve things like bioengineered pathogens or like.

[WeepingElf]

I think that there is an important aspect that is hardly addressed in the current public debate about AIs, and that is how AI scrapers overrun website after website. I have a feeling that this adverse side effect of AI technology is intended by the techbros because they want to peddle their concept of a CompuServe-style Metaverse as an alternative to the failing Internet. (And the autocrats also want something that is easier to censor than the Internet.)

[Ares Land]

I think that there is an important aspect that is hardly addressed in the current public debate about AIs, and that is how AI scrapers overrun website after website. I have a feeling that this adverse side effect of AI technology is intended by the techbros because they want to peddle their concept of a CompuServe-style Metaverse as an alternative to the failing Internet. (And the autocrats also want something that is easier to censor than the Internet.)

I think the scrapers just don't care. I don't think small websites are really on their radar much and the effects are adequately explained by indifference and/or incompetence.

[jal]

I think the scrapers just don't care. I don't think small websites are really on their radar much and the effects are adequately explained by indifference and/or incompetence.

Obviously. If you mean by "on their radar" "they don't scrape them" that's wrong, if you mean "they don't give a rat's ass about them" you are very right. (A year ago, a small site I regularly visit, posted a message after some downtime: "LLM crawlers have been wreaking DDOS-like havoc around the world, causing site outages and huge headaches for sysadmins of small/independent websites around the world. The spate of recent gateway errors and other problems is because of these botnets.").


JAL