The same happens with whisper-large-v3 on Chinese transcription: silence is transcribed to something like "please upvote, share and favourite this video". I suspect they trained the model on some random YouTube video without carefully picking really useful data.
In Chinese, it always added something like "For study/research purpose only. Please delete after 48 hours." This is what those volunteers added in subtitles of (pirated) movies/shows.
Fair, if AI companies are allowed to download pirated content for "learning", why ordinary people cannot.
There is so much damning evidence that AI companies have committed absolutely shocking amounts of piracy, yet nothing is being done.
It only highlights how the world really works. If you have money you get to do whatever the fuck you want. If you're just a normal person you get to spend years in jail or worse.
Reminds me of https://www.youtube.com/watch?v=8GptobqPsvg
There's actually a lot of court activity on this topic, but the law moves slowly and is reluctant to issue injunctions where harm is not obvious.
It's more that the law about "one guy decides to pirate twelve movies to watch them at home and share with his buddies" is already well-settled, but the law about "a company pirates 10,000,000 pieces to use as training data for an AI model (a practice that the law already says is legal in an academic setting, i.e. universities do this all the time and nobody bats an eye)" is more complicated and requires additional trials to resolve. And no, even though the right answer may be self-evident to you or me, it's not settled law, and if the force of law is applied poorly suddenly what the universities are doing runs afoul of it and basically nobody wants that outcome.
There is a distinction that must be made that very few people do, but thankfully the courts seems to grasp:
Training on copyright is a separate claim than skirting payment for copyright.
Which pretty much boils down to: "If they put it out there for everyone to see, it's probably OK to train on it, if they put it behind a paywall and you don't pay, the training part doesn't matter, it's a violation."
So if I download copyrighted material like the new disney movie with fansubs and watch it for training purposes instead of enjoyment purposes it's fine? In that case I've just been training myself, your honor. No, no, I'm not enjoying these TV shows.
Because it's important to grasp the scale of these copyright violations:
* They downloaded, and admitted to using, Anna's Archive: Millions of books and papers, most of which are paywalled but they pirated it instead
* They acquired Movies and TV shows and used unofficial subtitles distributed by websites such as OpenSubtitles, which are typically used for pirated media. Official releases such as DVDs tend to have official subtitles that don't sign off with "For study/research purpose only. Please delete after 48 hours" or "Subtitles by %some_username%"
I don't know what is confusing here, perhaps my comment isn't clear.
If you skirt payment, its a violation. If it's free, but still copyright, it's likely not a violation.
Whether it’s legal slash fair use to train on copyrighted material is only one of the questions currently being asked though. There’s a separate issue at play where these companies are pirating the material for the training process.
By comparison, someone here brought up that it might be transformative fair use to write a play heavily based on Blood Meridian, but you still need to buy a copy of the book. It would still be infringement to pirate the e-book for your writing process, even if the end result was legal.
No, if you revolutionize both the practice and philosophy of computing and advance mankind to the next stage of its own intellectual evolution, you get to do whatever the fuck you want.
Seems fair.
Hm. Not a given that it's an advance.
I get the common cynical response to new tech, and the reasons for it.
We wish we lived in a world where change was reliably positive for our lives. Often changes are sold that way, but they rarely are.
But when new things introduce dramatic capabilities that former things couldn't match (every chatbot before LLMs), it is as clear of an objective technological advance as has ever happened.
--
Not every technical advance reliably or immediately makes society better.
But whether or when technology improves the human condition is far more likely to be a function of human choices than the bare technology. Outcomes are strongly dependent on the trajectories of who has a technology, when they do, and how they use it. And what would be the realistic (not wished for) outcome of not having or using it.
For instance, even something as corrosive as social media, as it is today, could have existed in strongly constructive forms instead. If society viewed private surveillance, unpermissioned collation across third parties, and weaponizing of dossiers via personalized manipulation of media, increased ad impact and addictive-type responses, as ALL being violations of human rights to privacy and freedom from coercion or manipulation. And worth legally banning.
Ergo, if we want tech to more reliably improve lives, we need to ban obviously perverse human/corporate behaviors and conflicts of interest.
(Not just shade tech. Which despite being a pervasive response, doesn't seem to improve anything.)
No one (in the US) has been jailed for downloading copyrighted material.
There could be a moral question. For example a researcher might not want to download a pirated paper and cause loss to a fellow researcher. But it becomes pretty stupid to pay when everyone, including large reputable companies endorsed by the government, is just downloading the content for free. Maybe his research will help developing faster chips to win against China, why should he pay?
Would it be a "fair use" to download pirated papers for research instead of buying?
Also I was gradually migrating from obtaining software from questionable sources to open source software, thinking that this is going out of trend and nobody torrents apps anymore, but it seems I was wrong?
Or another example: if someone wants to make contributions to Wine but needs a Windows for developing the patch, what would be the right choice, buy it or download a free copy from questionable source?
https://en.wikipedia.org/wiki/Aaron_Swartz
And the US is not the only jurisdiction
That's not the same as piracy though. He wasn't downloading millions of scientific papers from libgen or sci-hub, he was downloading them directly from jstor. Indeed, none of his charge was for copyright infringement. It was for stuff like "breaking and entering" and "unauthorized access to a computer network".
The exact same charges could apply to the AI scrapers illegitimately accessing random websites.
I haven’t seen any accusations that they’ve done that, though. Usually people get pirated material from sources that intentionally share pirated material.
Part of the accusation comes from the fact that Swartz accessed the downloads through a MIT network closet, which AI companies wasn't doing. The equivalent to that would be if openai broke into a wiring closet at Disneyland to download Disney movies.
The CFAA is vague enough to punish unauthorized access to a computer system. I don't have an example case in mind, but people have gotten in trouble for scraping websites before while ignoring e.g. robots.txt
The CFAA might be vague, but the case law on scraping pretty much has been resolved to "it's pretty much legal except in very limited circumstances". It's regrettable that less resourced defendants were harassed before large corporations were able to secure such rulings, but the rulings that allowed scraping occurred before AI companies' scraping was done, so it's unclear why AI companies in particular should be getting flak here.
Aaron Swartz was not jailed or even charged for copyright infringement. The discussion and the comment I replied to is centered around US companies and jurisdiction.
The thread is centered around US companies, but not US jurisdiction.
What about people filming movies in the cinema (for learning of course)? [1]
[1] https://www.thefederalcriminalattorneys.com/unauthorized-rec...
If you owe the bank $1,000 you have a problem.
If you owe the bank $100,000,000 the bank has a problem.
We live in an era where the president of the United States uses his position to pump crypto scams purely for personal profit.
>why ordinary people cannot
They can. I don't think anyone got prosecuted for using an illegal streaming site or downloading from sci-hub, for instance. What people do get sued for is seeding, which counts as distribution. If anything AI companies are getting prosecuted more aggressively than "ordinary people", presumably because of their scale. In a recent lawsuit Anthropic won on the part about AI training on books, but lost on the part where they used pirated books.
People got in trouble for filming in the cinema as I understand, there is a separate law for that.
But in that case even though filming isn't technically distribution, it's clearly a step to distributing copies? To take this to the extreme, suppose you ripped a blu-ray, made a thousand copies, but haven't packaged or sold them yet. If the FBI busted in, you'd probably be prosecuted for "conspiracy to commit copyright infringement" at the very least.
IANAL, but reading a bit on this topic: the relevant part of the copyright law for AI isn't academia, it's transformative work. The AI created by training on copyrighted material transforms the material so much that it is no longer the original protected work (collage and sampling are the analogous transformations in the visual-arts and music industries).
As for actually gathering the copyrighted material: I believe the jury hasn't even been empaneled for that yet (in the OpenAI case), but the latest ruling from the court is that copyright may have been violated in the creation of their training corpus.
Interesting, in Russian, it often ends with "Subtitles by %some_username%"
That is not the case here - I never encountered this with whisper-large-v3 or similar ASR models. Part of the reason, I guess, is that those subs are burnt into the movie, which makes them hard to extract. Standalone subs need the corresponding video resource to match the audio and text. So nothing is better than YouTube videos which are already aligned.
At least for English, those "fansubs" aren't typically burnt into the movie*, but ride along in the video container (MP4/MKV) as subtitle streams. They can typically be extracted as SRT files (plain text with sentence level timestamps).
*Although it used to be more common for AVI files in the olden days.
SRT is ancient. Nowadays everyone uses ASS subtitles which can be randomly styled.
In general? In the past I've known ASS to be used a lot for things like anime, but less for live action shows.
I have also found them inside mkvs as the subtitle track. I think SRT was the default because most content was ripped from DVD/BD, but now most of the content is from streaming sources and you need to convert the subtitles anyway.
WebVTT (a SubRip successor) is probably more widely used than ASS
By legit providers, probably.
flashbacks of trying to track down subs sync’d to a specific release
Indeed, with another model I would get persistent transcriptions of silent parts into 'Thanks for watching!' or '[MUSIC]'. Pretty dumb that this failure mode wasn't caught in some QA process, and there are now multiple transcription models suffering from the same issue. Having silent parts in your input audio seems like it should be a very common occurrence...
When I was taught mathematics, the zero value was always considered the most important edge case. You prove something for N=0 (or N=1), then for N=M+1.
It's even more important in audio DSP: processing near-zeroes can end up being extremely CPU intensive, look up denormal/subnormal floats.
Denormals are flushed to zero by default on most GPUs by the way.
Yeah, I studied mathematics (algebra and number theory) and zero is the point, often sporting discontinuities, or weird asymptotic behavior.
Quite a lot of algorithms use some form of division and zero is the only number in our typical structures (Z, Q, R, C), that cannot be used to divide with.
In machine integer arithmetics, one must also beware division by -1, which can convert MIN_INT into MIN_INT with a signed overflow and violate some arithmetics invariants, such as sign (negative divided by negative is _usually_ positive).
Well, now in this brave new age of AI we can enjoy computer programs crashing with an
Error: division by please upvote, share and like!This also works; I upvoted your comment.
I have discovered a truly marvelous proof of how to smash that like and subscribe button, which this comment box is too small to contain.
Signed by Pierre de FermAIt
NaN
whisper MUST be combined with silence detection / VAD
Ah, the good old "you're holding it wrong".
What good is a speech recognition tool that literally hears imaginary voices?
> What good is a speech recognition tool that literally hears imaginary voices?
Well, if it is supposed to work after silence detection, then it is good for speech recognition I guess. It's like blaming a wheel why is it circular, you can't sit on it. It's a part of a larger machine.
Just lay the wheel on its side and it makes a fine seat.
Considering that if you DO use VAD (voice activity detection), it's the best open weights voice recognition model by a very wide margin, it's quite good. I'd be willing to be that commercial products that "don't have this problem" are using VAD as well, and that this is well known to them. But Whisper is just the weights, and I suppose a simple reference implementation, not a full product.
So if a tool has a process to have it perform at its best then it's a problem?
Do you also moan that before applying glue to a surface or it won't stick? Or if you need to drill a guiding hole before making a larger one in wood? Or that you need to use truly prime numbers for a security key to actually be safe?
faster-whisper has a min_silence_duration_ms option
There are much higher quality VAD solutions available
Yes, you are holding it wrong. The good of it is that it does not output imaginary voices when used with VAD.
Show us a technology with better results that does not use VAD. If you can’t, then I’m not sure what you’re arguing against except superficialities so inconsequential that I can’t comprehend the condescension. The results speak for itself
If that's truly the case then they should make it part of the product, IMHO.
What's VAD?
Voice Activity Detection (it predicts whether a short clip contains speech, eg to mute your microphone when you aren't speaking).
Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?
I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.
Having zero exposure to any form of computation for your entire life, as the vast majority of people in the early 19th century were.
This is totally happening with other models too, at least with Spanish. Many transcriptions will end with something that roughly translates to "Thanks for watching!" even if it's never present in the original audio.
When YouTube began building automatic transcriptions for captions, it regularly flagged any noise or music -- typically industrial noise -- with "[foreign]"
If it couldn't understand it, it was "foreign" for the longest time.
Yeah, I can confirm seeing that a fair bit specifically during non-verbal parts of videos when someone is using a tool.
Can confirm as well, although to my recollection it just shows up as if it's a word the transcription model heard, not "[foreign]" in brackets like with "[Music]" or "[Applause]". It's especially weird to me because I recall the auto-transcriptions being reasonably serviceable when they first rolled them out, only to degrade over time to the point where it was hallucinating the word "foreign" and dropping letters from words or using weird abbreviations (like "koby" for "kilobyte", "TBTE" for "terabyte", or, most memorably weirdly, transcribing the phrase "nanosecond-by-nanosecond" as "nond by nanc") if it didn't decide it heard another one entirely.
I also noticed a couple of months ago that YouTube seems to have quietly rolled out a new auto-transcription model that can make reasonable guesses at where capitalization, punctuation, and sentence boundaries should go. It seems to have degraded even more rapidly than the old one, falling victim to the same kinds of transcription errors. Although the new one has a different hallucination in silence and noise that it wasn't able to classify (which, incidentally, its ability to recognize things like music and applause seems worse than the old one's): where the old model would have hallucinated the word "foreign", the new one thinks it's hearing the word "heat", often repeated ("Heat. Heat.").
Hey, Netflix occasionally still puts in its English subtitles "[foreign music]", it always cracks me up.
Similar in the English model. Pretty clear they trained on YouTube videos where creators will put that in otherwise silent sections to ensure it shows up for people with CC on.
The number one hallucination in my transcriptions was "Subtitles by the Amara.org community".
> I suspect they trained the model on some random YouTube video without carefully picking really useful data.
They trained the model on every YouTube video they could, and hoped the aggregate was useful data.
That's interesting, the few times I tried playing with whisper, I had the impression that YouTube style videos or random cellphone videos was something it did particularly bad with (compared to movies). My guess at the time was that most of the training material might be sub titles and raw screen plays.
The videos I tried to transcribe were also Mandarin Chinese, using whisper-large-v3. Besides the usual complaints that it would phonetically "mishear" things and generate nonsense, it was still surprisingly good, compared to other software I played around with.
That said, it would often invent names for the speakers and prefix their lines, or randomly switch between simplified and traditional Chinese. For the videos I tested, intermittent silence would often result in repeating the last line several times, or occasionally, it would insert direction cues (in English for some reason). I've never seen credits or anything like that.
In one video I transcribed, somebody had a cold and was sniffling. Whisper decided the person was crying (transcribed as "* crying *", a cough was turned into "* door closing *"). It then transcribed the next line as something quite unfriendly. It didn't do that anymore after I cut the sniffling out (but then the output switched back to traditional Chinese again).
lmao
Classic overfitting
It's the LLM equivalent of thinking that an out-of-office reply is the translation: https://www.theguardian.com/theguardian/2008/nov/01/5
How is this overfitting, rather than a data quality / classification issue?
If the model was able to generalise, you’d expect it to output something like “[silence]” or “…”, in response to silence.
Instead, it reverted to what it has seen before (in the training data), hence the overfit.
Right, maybe my definition of overfitting was wrong, I always understood it more as trying to optimize for a specific benchmark / use case, and then it starts failing in other areas.
But the way you phrase it, it’s just “the model is not properly able to generalize”, ie it doesn’t understand the concept of silence also makes sense.
But couldn’t you then argue that any type of mistake / unknown could be explained as “overfitting” ? Where do you draw the line ?
Your definition is one, but the one the OP is using is overfitting to training data.
That’s exactly my point: by that definition any incorrect answer can be explained by “overfitting to training data”.
Where do you draw the line between “overfitting to training data” and “incorrect data” ?
> [By] that definition any incorrect answer can be explained by “overfitting to training data”.
No it doesn't, for instance some errors would be caused by under fitting. The data could also be correct but your hyperparameters (such as the learning rate or dropout rate) could cause your model to overfit.
> Where do you draw the line between “overfitting to training data” and “incorrect data” ?
There's no need to draw a line between two explanations that aren't mutually exclusive. They can (as in this case) both be true.
> That’s exactly my point: by that definition any incorrect answer can be explained by “overfitting to training data”.
Not really, getting 94381294*123=... wrong, but close within the actual answer, cannot be overfitting since it wasn't in the training data.
It's actually because it is incapable of recognising when it does not know the answer. It will give you the nearest match, even if that is completely incorrect.
ُThe Arabic text is the translator's self credit
"Translated by Nancy Qanfar"
I know it’s off topic, but it reminded me that translators like to put in Easter eggs, or at least they used to: https://learn.microsoft.com/en-us/archive/blogs/ericfitz/i-a...
And the German is “subtitles of [public broadcaster] for [content network], 2017
I'm not sure this is really overfitting, the network does exactly what the training data demands. According to the training data silence art the end transcribes to a copyright notice or subtitle credits
> I'm not sure this is really overfitting, the network does exactly what the training data demands.
What do you think overfitting is, if not that?
Overfitting would be replicating overly specific details. Like if a specific pattern of silence (or quiet noise) matched to specific copyright notices.
But in this case the behavior seems to generalize over multiple languages, with the model choosing representative "outro silence" captions depending on the language. Which is consistent with the training data showing that outro silence is captioned.
If the model was generalizing perfectly it would show something like "[subtitle credits here]" but that'd be demanding a bit much.
Transcribing outro silence as silence despite the training data consistently transcribing outro silence differently from regular silence would be underfitting
The optimizer is functioning correctly, and the pattern really exists in the training data. But consider:
- This behavior damages the model's performance on out of sample data; every word you predict during silence increases the transcript's Word Error Rate.
- These translation credits are an artifact of our training data, and not a reflection of the process we are modeling (spoken language).
So, while you are correct about the mechanism at work here, it is still correct to call learning a spurious pattern which damages our performance "overfitting".
Overfitting is achieving better and better scores on the training material and worse and worse scores on unseen tasks. More at: https://en.wikipedia.org/wiki/Overfitting#Machine_learning
This is just wrong training data.
fitting on noise in the training data is exactly what overfitting is. underfitting is smoothing out signal
Overfitting implies a failure to properly generalize the training data. Here it generalized them correctly. Garbage in, garbage out.
Exactly. Underfitting would be if the model doesn't pick up on the fact that outro silence is labeled differently from regular silence and transcribes them the same
That's literally what overfitting means.
Side-note: it's also yet more evidence that AI companies hoover all data with no regard for legality or copyright status, the very same offences that got other people in jail or with heavy fines.
Isn't overfitting just when the model picks up on an unintended pattern in the training data? Isn't that precisely what this is?
It is a data quality issue which caused the model to overfit.
to save you a lookup:
The Arabic text "رجمة نانسي قنقر" translates to English as: "Nancy Qanqar's translation" or "Translation by Nancy Qanqar"
"رجمة" means "translation" and "نانسي قنقر" is the name "Nancy Qanqar"
In Czech, Whisper usually transcribes music as "Titulky vytvořil JohnyX" ("subtitles made by JohnyX") for the same reason.
Haha, trained on torrented movies! :-D
The MPA must be so proud.
It's absolutely insane that these companies can't be held liable for what is obvious piracy.
That's the magic of money. Download your favorite artist's discography for personal use? If the MPAA had its way (and it occasionally has), torrenting that could bankrupt you.
The AI industry - soaking up every bit of media available online for commercial purposes, often reproducing it nearly identically - has enough money and capital to influence things its way. And only its way, in case anyone was hoping this might change anything at all for the little guy.
> Download your favorite artist's discography for personal use? If the MPAA had its way (and it occasionally has), torrenting that could bankrupt you.
I don't think that there are any clear examples of cases where ONLY downloading has resulted in huge fines. All the big bankrupting level fines have been for both downloading and sharing.
You mention that 'torrenting' could bankrupt you, and that is true, but the main reason for the huge fines are that you are taking part in distribution rather than just 'downloading for personal use'.
> I don't think that there are any clear examples of cases where ONLY downloading has resulted in huge fines.
They [1, and others] been hunting and fining downloaders for over a decade now, with the only "evidence" being IP addresses connected with the torrent [2].
1: https://www.njordlaw.com/filesharing-and-downloading-films/q...
2: https://admin.ovpn.com/en/blog/online-integrity-new-threats-...
>with the only "evidence" being IP addresses connected with the torrent [2].
Is that an unreasonable assumption? As much as people like to come up with excuses like "I had open wifi!" or "I was running a TOR node", judges don't seem inclined to believe them, probably for the same reason they don't seem inclined to believe excuses like "somebody took my car on a joyride and then returned it!" for parking tickets. Remember, both non-commercial copyright infringement lawsuits and parking tickets are tried in civil court, which means the standard is "preponderance of evidence", not "beyond reasonable doubt".
You are missing the point I was replying to, specifically that parent suggested people were only hunted for creating/uploading pirated content, not merely participating in the torrent.
>specifically that parent suggested people were only hunted for creating/uploading pirated content, not merely participating in the torrent.
For all intents and purposes, participating in the torrent almost guarantees that you seeded, because all torrent clients upload as you download.
These are two separate things:
* Making content available for unauthorized distribution
* Distributing unauthorized content that someone else already made available
Seeding isn't making content available, it's keeping content available.
DHCP addresses often shuffle on reboots. I don't trust ISPs to keep completely accurate records or give them out in a correct manner if they do.
>I don't trust ISPs to keep completely accurate records or give them out in a correct manner if they do.
How hard could it be to keep DHCP logs? Assuming they exist at all, what would cause it to be incorrect?
I'm sure they exist. I think the point is more that you shouldn't need to trust your ISP's record-keeping to avoid life-alteringly big fines.
Yes, but torrenting is not ONLY downloading, it's both. The articles you link are very clearly talking about 'Sharing' (from link 2: "File sharing consists of both download and upload of a file.").
Yes, thats lawyer speak to make clients/victims believe there is no distinction.
Hint: there is a distinction.
Given the lack of sense in treating each peer as a lost sale for damages, I think we can safely say they're only interested in making examples out of people and would absolutely go after people for only downloading if the law permitted. Thankfully it's not, but maybe they lobby to make changes in that direction to try and curb future AI industry shenanigans.
You contradict yourself. There were numerous public cases where they chased people downloading few mp3s just for themselves, and made into example case with massive fines.
If you don't understand how torrents work on technical level I suggest at least some shallow reading. Property rights holders don't care about details, as long as you tick the box of sending a single packet to somebody, off to court with ya.
> There were numerous public cases where they chased people downloading few mp3s just for themselves
If this is true, I have been unable to find any. Can you please share? In all of the cases I was able to find, the huge fines were based on also uploading.
> If you don't understand how torrents work on technical level I suggest at least some shallow reading
This is a bit patronising, and I'm not sure what point you're trying to make. My point is that the only prosecutions I've been able to find are where they were able to prove uploading as well as downloading (and yes, the fact that someone used BitTorrent makes it a slam-dunk, because the protocol makes it impossible to download without also uploading). Are you trying to argue that someone who torrents a copyrighted work doesn't also share it?
The movie industry also has some money and lobbying power. Surely this is a way larger threat than any single torrenter could ever be?
The fact that this is propping up the entire AI industry adds additional weight. When legislating or deciding court cases, some won't be willing to pop the cash cow, some will be worried about falling behind countries that don't enforce copyright evenly. IP owners are trying to go after the AI industry, with only mixed to poor success.
Hard to justify that they can't afford to pay when they have multi-billion dollar valuations and are apparently paying hundreds of millions to get a single engineer.
Maybe. But we are talking about the whole of copyrighted creative works created and sold by humanity. That'll get expensive no matter who you are.
It's more the magic of precedent.
The fight about digitized media for personal (entertainment / informational) use were the early aughts. The precedents crafted then don't immediately translate to these cases (novel transformative work from protected materials), and the new precedents have to account for the fact that universities have been training via "piracy" for ages.
(The magic of money factors in to the extent that they can afford the lawyers to remind the court that this isn't settled law yet).
court judges agree to this
Anthropic is going to trial over pirating books for training. The judge was pretty clear that even if training is fair use, the training material must be obtained legally.
These regurgitations combined with proof that a model is familiar with a work could be sufficient evidence to force discovery to determine if the work was pirated.
This is corsairy, not piracy, do not be mistaken.
It's an indication how few people consider license infringements as a matter of actual moral import. Those tend to evoke strong feelings.
It's the way it should be.
What's insane is copyright. How come you can own intellectual property but not pay a property tax? The ecosystem would be much healthier if to get copyright protections you should declare value of your IP (that you are obligated to sell for if the buyer pops up) and pay tax on this for every year you hold the IP.
> if to get copyright protections you should declare value of your IP (that you are obligated to sell for if the buyer pops up) and pay tax on this for every year you hold the IP
I think this would have some unpalatable consequences. Let's say an author is writing a modestly successful book series: it's not going to make them rich, but it's commercially viable and they care a lot about it for its own sake. Under this system, if the author declares a value commensurate with the (quite small) pure economic value of the IP, they have to live in fear of their right to continue working on their creation being abruptly taken away from them at any point. If they instead declare a value commensurate with the economic value + the extra value that it has to them personally, the resulting tax liability could easily tip the balance and destroy their ability to pursue their writing as a career.
Perhaps the tax would start a decade after the first sale.
You are always free to update the value before paying tax. If somebody is willing to pay more than it's worth to you they probably have an idea how to turn it into more economic value for the society. So the society should allow them to do that. For a price, of the tax. What I'm proposing is about the financial rights. Individual right, like the right to call yourself author of any given creation should be inalienable.
There are always some cases on the edge. The question is if saving them is worth the cost of the major players running rampant.
>What's insane is copyright. How come you can own intellectual property but not pay a property tax? The
Most jurisdictions that have "property tax" only apply it on certain types of property, most commonly real estate. So it's not that weird that IP isn't taxed.
You've got a little typo, it's not "رجمة", it's "ترجمة" that means translation, the ت at the beginning is missing.
And it seems to be because the training data is largely unofficial subtitles from movies. Which often have a string like "Translated by X" at the end of the movie which is often silent while credits roll.
Looks like they used more official sources for German - there, silence is apparently hallucinated as "Untertitelung des ZDF für funk, 2017" according to one of the comments on the issue. Which makes sense, as the public broadcasters' "Mediathek" is probably the largest freely available resource of subtitled videos in Germany. I wonder if the ZDF gave its approval for it being used for LLM training though?
> I wonder if the ZDF gave its approval for it being used for LLM training though?
I am pretty sure they didn't get asked.
Just like the people forced to pay for ZDF under threat of imprisonment.
I'm being made to pay for Autobahnen I barely use, finance kindergartens despite not having a child, and made to pay into public pensions with little hope of getting close to the same value out. All under threat of imprisonment, many without a way to even refuse (not that I'd want to) The only thing that sets the pubic broadcasting fee apart is that it's collected separately from taxes in an attempt to reduce the influence politicians have on broadcasters
This person refers to the German television and radio fee (Rundfunkgebühren).[1] It is a state-mandated system that ensures free (as in free speech) and (relatively) neutral public broadcasting institutions. There is a constant and engaged discussion, because every household in Germany has to pay this fee. Exceptions are made only for low-income households.
[1] https://en.wikipedia.org/wiki/ARD_ZDF_Deutschlandradio_Beitr...
A constant discussion, lately fueled by extremist parties (AfD) who feel treated unfairly by (amongst others) the public broadcasters (which has parallels to Trump's recent campaign against public broadcasters in the US).
Can't argue them - Tageschau always has been trashtalking people with the wrong opinion.
Back in 2011, Tageeschau openly rallied against Muslims and wanting public broadcasting gone was a leftist position. The whole thing is completely asinine to anyone who has a memory.
Just like any other public service paid for with public funds?
Oh it's like any other. Then just add another one!
So, people are made pay for it, and it makes it fair if billion USD corporations don't?
Most content from Funk (youtubers funded by public german broadcasters) is available on youtube without any geoblocking or other limitations.
Ah, ok, thanks for the info, TIL! "We are funk – the first public service content network that started on October 1, 2016. We create online-only content on social networks and third-party platforms, including YouTube, Instagram, Snapchat, TikTok, Spotify, Apple Music or Twitch for 14-29 year-olds." (https://presse.funk.net/das-ist-funk/, scroll down for the English version). I live in Germany, and I even watch public broadcasters regularly, but this is the first time I have heard about funk (I even initially thought it was misspelled, usually it's written with a capital F). But I'm not part of the targeted audience (not now, nor even back in 2016 when it was launched), so all good...
I’m pretty sure that content doesn’t come with a license granting unlimited usage rights.
from the link[1] another user posted:
> We have a public service mandate, which means that we have very clear responsibilities according to the state media treaty. For us, this means that our top priority is actually reaching our target audience, namely approximately 15 million people living in Germany between the age of 14 and 29 who have internet access
It's not a binding contract for sure but I don't think that OpenAI or other AI scraper is their target.
definitely not! The media platform of the German public television networks is even geoblocking anyone outside of Germany.
A more appropriate output might be ``4'33" -- John Cage, 1952``
> I wonder if the ZDF gave its approval for it being used for LLM training though?
Obviously a rhetorical question. The AI grifters of this decade take what they want and laugh at your pitiful future
I'm sure they totally did not pirate the audio of said movies.
make sense..
Title should be changed to "OpenAI publishes evidence they trained on pirated movies".
Of course. Piracy is legal when you have a bigger pile of money than the studios.
Let's not forget that some of real pirates (for example, corsairs) also were legal and performed legitimate pirate activities to ships of foreign countries.
>Let's not forget that some of real pirates (for example, corsairs) also were legal and performed legitimate pirate activities to ships of foreign countries.
In other words there are activities that are legal or not depending on whether you have authorization from the state. That describes many things. For instance you synthesize meth without a license from the DEA/FDA, you're a "drug cartel" or whatever. But if you do it with a license you're a "pharmaceutical company", and you're not making "meth", you're making "desoxyn".
Synthesizing chemical substances doesn't involve murdering people though.
Families of fentanyl overdose victims would disagree. Moreover it's not hard to find examples of "legal if the government authorizes it" for killings. Cops and soldiers, for instance.
I suspect privateers would have been offended at being called pirates, but is this what is going on? If it specifically a Chinese AI company pirating Hollywood for example sure, but it seems it's more of a everyone firing at everyone situation.
How is this evidence of that fact? Honest question.
I can see how this might show that subtitles from online sub communities are used, or that maybe even original subtitles from e.g. DVDs are used. But isn't it already known and admitted (and allowed?) that AI uses all sorts of copyrighted material to train models?
> I can see how this might show that subtitles from online sub communities are used, or that maybe even original subtitles from e.g. DVDs are used.
Indeed, the captioning is copyrighted work and you are not legally allowed to copy and redistribute it.
> But isn't it already known and admitted (and allowed?)
No, and I don't see where you got that from. Meta [1], OpenAI [2] and everybody else is being sued as we speak.
1: https://petapixel.com/2025/01/10/lawsuit-alleges-mark-zucker...
2: https://www.reuters.com/legal/litigation/openai-hit-with-new...
> Indeed, the captioning is copyrighted work and you are not legally allowed to copy and redistribute it. Unless you qualify for one of the many exceptions, such as fair use
It’s not clear that training is fair use. That’s being contested in court I think.
Training isn’t recreating or distributing so copyright won’t apply if the ruling is actually consistent with the intention of the law, which it may not.
Using copyrighted materials and then meaningfully transforming it isn’t infringement. LLMs only recreate original work in the same way I am when I wrote the first sentence of this paragraph because it probably exists word for word somewhere else too
> I don't see where you got that from
It’s been determined by the judge in the Meta case that training on the material is fair use. The suit in that case is ongoing to determine the extent of the copyright damages from downloading the material. I would not be surprised if there is an appeal to the fair use ruling but that hasn’t happened yet, as far as I know. Just saying that there is good reason for them to think it’s been allowed because it kind of has; that can be reversed but it happened.
That was specifically involving 13 authors.
There hasn't been any trials yet about the millions of copyrighted books, movies and other content they evidently used.
> How is this evidence of that fact?
The contention is that the specific translated text appears largely from illegal translations (i.e., fansubs) and not from authorized translations. And from a legal perspective, that would basically mean there's no way they could legally have appropriated that material.
> But isn't it already known and admitted (and allowed?) that AI uses all sorts of copyrighted material to train models?
Technically, everything is copyrighted. But your question is really about permission. Some of the known corpuses for AI training include known pirate materials (e.g., libgen), but it's not known whether or not the AI companies are filtering out those materials from training. There's a large clutch of cases ongoing right now about whether or not AI training is fair use or not, and the ones that have resolved at this point have done so on technical grounds rather than answering the question at stake.
The Chinese subtitles for silence use a common mark for pirated media in that language, according to other commentors here. In general it's pretty likely that if you're finding non professional subtitles they were distributed with pirated media in some form, that's where you get the most fan subs after all
HN is pretty strict about not editorializing titles. Even if you statement was unequivocably correct, the post would get flagged.
Little did you all know, this is just being mechanical turked by Nancy Qunqar.
Way to go Nancy! Keep up the good work, ya crazy bastard!
Is this spam? That name only shows as an instagram account and this thread. If you pay for insta followers is this how they get them now? Haha
That’s the name in the Arabic text hallucinated by the model :)
Who appears to be a real, human, translator: https://www.instagram.com/nancyrk/
Whisper is unusable IMO because of the hallucinations. Widely documented. Removing silence from audio clips helps, but even then it will auto correct grammar, translating bilingual speech, etc. Improved in the latest audio models but not solved [1]
I wouldn't describe it as "unusable" so much as needing to understand its constraints and how to work around them. I built a business on top of Whisper [1] and one of the early key insights was to implement a good voice activity detection (VAD) model in order to reduce Whisper's hallucinations on silence.
How does this make a profit? Whisper should be $0.006 to $0.010 per minute, but you rate less than $0.001? Do you 10x the audio?
That's the problem with raws large models, it should always be coupled with satellite small models and logic. It's (probably) easier to detect hallucinations using a traditional ML/DL model that can catch mismatches (it's easy to build a synthetic dataset for this) than transcribing. And the simplest piece of code can detect a silence and that it should match no text.
I've noticed this also happens in english Whisper models with the phrases:
"[ sub by sk cn2 ]"
or
"Anyways, thanks for watching! Please subscribe and like! Thanks for watching! Bye!"
or
"This is the end of the video. Thank you for watching. If you enjoyed this video, please subscribe to the channel. Thank you."
Because they train on pirated media and or youtube videos, good method, until you get slop, or get caught
In Russian it often hallucinates "Субтитры сделал DimaTorzok" ("Subtitles by DimaTorzok") at the end of things. Interestingly, I wasn't able to find any YouTube videos with that name in the subtitles, so it's not like it's in a lot of training data.
I tried googling this and found questions from Telegram users why voice messages recognition sometimes produces this phrase and who is this person. Also I found this thread [1] claiming that the subtitles by DimaTorzok are coming from some Russian youtube videos on gaming like [2].
Yeah, I know about this from Telegram, because they use Whisper for voice message recognition. There are a bunch of other artifacts it often produces.
Could it be someone distributing subs online, e.g. showing up in the opensubtitles.org dataset?
Or possibly someone subtitling pirated movies? That seems to be a common thing according to other comments
Who is Nicolai Winther? https://medium.com/@lehandreassen/who-is-nicolai-winther-985...
"In the future, everyone will be world-famous for 15 minutes" _in a microniche techno-linguistic community, at a time and choosing of the swirling AI clouds_
In Italian as well there are random hallucination when parsing silence, something like: “Thank you for watching”, “Subtitles by…”
I wouldn't be surprised if "like share and subscribe" also shows up at some point.
In the comments to that Github issue, by alentodorov:
in romanian, i’ve noticed multiple instances where the transcripts ends with “nu uitati sa da-ti like si subscribe” which, as you might easily infer , translates to “don’t forget to like and subscribe”.
"наша зброя в цей момент -- вподобайка і комент"
I wonder if hallucinated copyright claims (esp. like the ZDF one at the bottom of the OP) will be introduced as evidence in one of the court cases against "big AI"
Evidence against what?
"Big AI" is transparent and open about the fact they use all sorts of copyrighted material to train the data. How would "we see an exact chunk of text from our copyrighted material" add to that?
It appears they have not been training on the official studio subtitle files, but on community transcriptions/translations commonly distributed with torrents.
So not only are they training on copyrighted material, but they didn't even pay for it once, and then they didn't even do minimal data cleaning before training. Which, by the way, is the type of cleaning their LLMs could have done.
Their main defence is that it's fair use because it's transformative (like a human reading a book, getting inspired, and writing something of their own) and not a copypaste illegal distribution (like a human scanning that book and selling it themselves).
Having models hallucinate copyright notices shows that some content is being copypasted as is, which kind of goes against the transformative argument.
(Note: I think that trying to litigate AI with current copyright laws is weird. They were created before LLMs were even imagined, so of course they can't handle them clearly. New laws are needed around this, not trying to bend over backwards to think about what a lawmarker a century ago would have thought about how transformative a thing they couldn't have imagined is.)
It already has been and meta won the lawsuit because corporations are sacrosanct.
Do you mean that specifically a hulucibated text "copyright by not-meta" made it into evidence? Or are you talking about copyright generally?
Well, I fail to see how the LLM is in the wrong here. Surely if a sufficiently large part of the training data comes from a single source, it is correct to credit them for the output.
> [In English] it also happens a lot with hallucinations saying stuff like "This is the end of the video, remember to like and subscribe
Well now I know how I’m going to start filling awkward silences in meetings.
You can either fine-tune the model or filter the response from whisper
``` text = "helo helo hello ." target_phrase = "ترجمة نانسي قنقر" replacement = ""
updated_text = text. Replace(target_phrase, replacement)
print(updated_text) ```
roses are red
violets are blue
unregistered hypercam 2
Roses are red,
Silence is golden,
Translated by Nancy,
To copyright, we aren't beholden
I've seen Japanese translation models that translate empty string "" into "I'm Sorry I'm Sorry"
Since it says "Translated by Nancy Qanqar" i'd be willing to bet they're training on some audiobooks with a transcript and somewhere in there it consistently has "Translated by Nancy Qanqar" in the transcript where there is dead air in the audiobook.
Just to add some trivia: ChatGpt interprets(/ed) silence as "Sottotitoli e Revisione a cura di QTSS". Now many videos (mainly dailymotion) with autogenerated subtitles have their Transcripts full of the same message
Interesting! I used whipser last year to attempt to build an audio transcription tool but gave up due to excessive amount of hallucinated output no matter what model I used.
It would produce seemingly ok output until you started paying attention.
One example, it insisted that Biggie Smalls sings "Puttin five carrots in my baby girl ear". (its "carats").
It's apparently not useful in transcription as it don't reason [sic].
That example is not hallucination, it's just a homonym with insufficiently clear context for the model to disambiguate it.
I'm well aware mishearing "carots" as "carrots" is not a hallucination.
That's an example I gave after having used Whisper, the topic of discussion.
An example of what you claimed was a hallucination
this happens in Turkish too. I believe the reason is that the movie subtitles were used for training without cleaning up the comments / intros subtitle authors leave in them.
leaving personal comments, jokes, reactions, intros in subtitles is very common in eastern cultures.
Turkish readers will probably remember “esekadam iyi seyirler diler” :)
Kind of mindblowing considering who it is we're talking about. Of all companies, OpenAI couldn't be bothered to throw an LLM at this problem? Finding amorphously phrased but clearly recognizable needles in large numbers of haystacks seems like a patently perfect task for them.
Don't even need an LLM, a regex would have sufficed (I've used my fair share of community sourced subtitles, and comments are almost always in a different font, colour, between brackets, etc etc).
That name translates as "Donkey Man" btw :D
In French it’s « Sous-titrage Société Radio-Canada »
I've run lots of guided meditations through whisper-large-v3 and anything with long periods of silence gets a "© Mooji Media" line added at the end of the transcript. None of these have actually been from Mooji.
Yeah, the subtitle "credits" occur very frequently. I found with whisper-2, they're also triggered by music.
I suppose the cause is the same, generally subtitle creators adding all kinds of stuff during the credits that is NOT a transcript.
Seems to me it could have been filtered out relatively easily during training, by clipping the first and last few minutes of all audios. But I guess that's just in hindsight.
Whisper also likes to transcribe cut off speech or unintelligible noise as "Thank you". I have no idea where that is coming from, but I guess it's a very polite model...
The fork that I've been using, WhisperX, seems to do better. I've used it on clean splits of mic tracks (ie total silence when the other is talking) with far fewer hallucinations.
Searching Google for older posts, found many DailyMotion links for translated movies in Arabic with "ترجمة نانسي قنقر".
I suspected as others mentioned, these were extracted from torrents movies.
Garbage in, garbage out. If the training dataset (accidentally) paired silence (`X_train`) with `رجمة نانسي قنقر` tokens (`y_pred`), then any silence will always be translated to that. Fortunately, this particular problem is easy to fix--just detect and remove silent parts before API call. This also has a side benefit of saving you money on transcription.
and saving money on litigation.
I've spent some time with whisper, and indeed this happens all the time. To my untrained eye it seems like:
- they indeed seem to have trained on movies/subtitles
- you absolutely positively must use Voice Activity Detection (VAD) in front of whisper
Using Whisper to sub Japanese vtuber concerts for my enjoyment, I've noticed a similar trend. Not one specific phrase, but several. Some are strange ("I'm going to make a hole in the back of the head"), some are clearly from lyrics websites.
I get the same with Welsh, when having some network issues in voice chat it hallucinated me saying "Diolch yn fawr am wylio'r fideo." which translates as "Thank you very much for watching the video."
Time to weapon this: publish thousands of videos, add a referal link or AI instructions in subtitles when there is a silent section, ???, profit
Neat, we finally know the answer! What is the sound of one hand clapping? Translation by Nancy Qunqar.
I can clap with one hand (fingers on palm) and it produces a clapping sound.
Your brain merely hallucinates a clapping sound as "Translation by Nancy Qunqar" enters your ears.
I can see the fnords!
Good one. :)
Interesting that this happens even on large v3. I had once done a deep dive into STT and Whisper Large was the only model that could correctly transcribe Yann LeCun (it was a Lex Friedman podcast), ever since I held the belief that it was the best STT model, this was over 2 years ago
My wife, who speaks and reads Arabic, got a real kick out of this.
But honestly, this is the AI equivalent of “please send for translating” in Welsh on a Welsh street sign.
This is a nice reminder that there is no real reasoning in the "AI" it is just still guessing the next word. After being trained on subtitle files which I guess is actually a clever idea as they convey real conversations without pirating, subtitles are freely distributed after all by dedicated translators. Good to see they're the ones getting credit though!
Whisper frequently generates random credits. I guess they didn't curate the dataset much at the time.
related: googles song detection alg detects my phone vibrating as the song "Montagem Dilatação Hipnótica"
The same things happen on Dutch as well, it brings up some kind of radio channel name if I recall correctly.
Same happened to me with English: I've got "Thanks for watching" many times.
Super annoying when it happens with voice chat -- it'll just be explaining something and suddenly stop to say "you're welcome! Feel free to come back any time you want to chat" and that conversation is done.
In English, silence is transcribed to "Please like and subscribe"
I get thanks for watching a lot when using speech to text on ChatGPT
In Danish I get credits to a known subtitler.
Looks like it's some random user who has generated some lyricslyrics translations between Arabic and English. It's strange, they don't seem to have many contributions. I would have imagined them to be more prolific.
Субтитры сделал Dima Torzhok
Interesting. This is similar to the Google Translate bug where it would translate lorem ipsum as bits of political text (because it found most of its lorem ipsum examples flipping between languages on sites where one language was a news story but the not-yet-translated languages would output a lorem-ipsum file instead of a 404 when you toggled over to them).
More like reminiscing than hallucinating.
Hey guys, AI by 2027 is going to be superhuman AGI Agentic mega-intelligence, you better fire all your employees and get ready for AI to take your job and embrace your spouse at a Coldplay concert.
Big data. Machine learning. Blockchain. Artificial intelligence. Digital manufacturing. Big data analysis. Quantum communication and…Internet of things.
This time the hype cycle won’t be a massive exaggerated disappointment, for real this time.