How one YouTuber is trying to poison the AI bots stealing her content
If you’ve been paying careful attention to YouTube recently, you may have noticed the rising trend of so-called “faceless YouTube channels” that never feature a visible human talking in the video frame. While some of these channels are simply authored by camera-shy humans, many more are fully automated through AI-powered tools to craft everything from the scripts and voiceovers to the imagery and music. Unsurprisingly, this is often sold as a way to make a quick buck off the YouTube algorithm with minimal human effort.
YouTuber F4mi, who creates some excellent deep dives on obscure technology, recently detailed her efforts “to poison any AI summarizers that were trying to steal my content to make slop.” The key to F4mi’s method is the .ass subtitle format, created decades ago as part of fansubbing software Advanced SubStation Alpha. Unlike simpler and more popular subtitle formats, .ass supports fancy features like fonts, colors, positioning, bold, italic, underline, and more.
[embedded content]
It’s these fancy features that let F4mi hide AI-confounding garbage in her YouTube transcripts without impacting the subtitle experience for her human viewers. For each chunk of actual text in her subtitle file, she also inserted “two chunks of text out of bounds using the positioning feature of the .ass format, with their size and transparency set to zero so they are completely invisible.”
In those “invisible” subtitle boxes, F4mi added text from public domain works (with certain words replaced with synonyms to avoid detection) or her own LLM-generated scripts full of completely made-up facts. When those transcript files were fed into popular AI summarizer sites, that junk text ended up overwhelming the actual content, creating a totally unrelated script that would be useless to any faceless channel trying to exploit it.
Copyright Office suggests AI copyright debate was settled in 1965
The US Copyright Office issued AI guidance this week that declared no laws need to be clarified when it comes to protecting authorship rights of humans producing AI-assisted works.
“Questions of copyrightability and AI can be resolved pursuant to existing law, without the need for legislative change,” the Copyright Office said.
More than 10,000 commenters weighed in on the guidance, with some hoping to convince the Copyright Office to guarantee more protections for artists as AI technologies advance and the line between human- and AI-created works seems to increasingly blur.
But the Copyright Office insisted that the AI copyright debate was settled in 1965 after commercial computer technology started advancing quickly and “difficult questions of authorship” were first raised. That was the first time officials had to ponder how much involvement human creators had in works created using computers.
Back then, the Register of Copyrights, Abraham Kaminstein—who was also instrumental in codifying fair use—suggested that “there is no one-size-fits-all answer” to copyright questions about computer-assisted human authorship. And the Copyright Office agrees that’s still the case today.
“Very few bright-line rules are possible,” the Copyright Office said, with one obvious exception. Because of “insufficient human control over the expressive elements” of resulting works, “if content is entirely generated by AI, it cannot be protected by copyright.”
The office further clarified that doesn’t mean that works assisted by AI can never be copyrighted.
“Where AI merely assists an author in the creative process, its use does not change the copyrightability of the output,” the Copyright Office said.
Following Kaminstein’s advice, officials plan to continue reviewing AI disclosures and weighing, on a case-by-case basis, what parts of each work are AI-authored and which parts are human-authored. Any human-authored expressive element can be copyrighted, the office said, but any aspect of the work deemed to have been generated purely by AI cannot.
Report: DeepSeek’s chat histories and internal data were publicly exposed
A cloud security firm found a publicly accessible, fully controllable database belonging to DeepSeek, the Chinese firm that has recently shaken up the AI world, “within minutes” of examining DeepSeek’s security, according to a blog post by Wiz.
An analytical ClickHouse database tied to DeepSeek, “completely open and unauthenticated,” contained more than 1 million instances of “chat history, backend data, and sensitive information, including log streams, API secrets, and operational details,” according to Wiz. An open web interface also allowed for full database control and privilege escalation, with internal API endpoints and keys available through the interface and common URL parameters.
“While much of the attention around AI security is focused on futuristic threats, the real dangers often come from basic risks—like accidental external exposure of databases,” writes Gal Nagli at Wiz’s blog. “As organizations rush to adopt AI tools and services from a growing number of startups and providers, it’s essential to remember that by doing so, we’re entrusting these companies with sensitive data. The rapid pace of adoption often leads to overlooking security, but protecting customer data must remain the top priority.”
Ars has contacted DeepSeek for comment and will update this post with any response. Wiz noted that it did not receive a response from DeepSeek regarding its findings, but after contacting every DeepSeek email and LinkedIn profile Wiz could find on Wednesday, the company protected the databases Wiz had previously accessed within half an hour.
OpenAI teases “new era” of AI in US, deepens ties with government
On Thursday, OpenAI announced that it is deepening its ties with the US government through a partnership with the National Laboratories and expects to use AI to “supercharge” research across a wide range of fields to better serve the public.
“This is the beginning of a new era, where AI will advance science, strengthen national security, and support US government initiatives,” OpenAI said.
The deal ensures that “approximately 15,000 scientists working across a wide range of disciplines to advance our understanding of nature and the universe” will have access to OpenAI’s latest reasoning models, the announcement said.
For researchers from Los Alamos, Lawrence Livermore, and Sandia National Labs, access to “o1 or another o-series model” will be available on Venado—an Nvidia supercomputer at Los Alamos that will become a “shared resource.” Microsoft will help deploy the model, OpenAI noted.
OpenAI suggested this access could propel major “breakthroughs in materials science, renewable energy, astrophysics,” and other areas that Venado was “specifically designed” to advance.
Key areas of focus for Venado’s deployment of OpenAI’s model include accelerating US global tech leadership, finding ways to treat and prevent disease, strengthening cybersecurity, protecting the US power grid, detecting natural and man-made threats “before they emerge,” and ” deepening our understanding of the forces that govern the universe,” OpenAI said.
Perhaps among OpenAI’s flashiest promises for the partnership, though, is helping the US achieve a “a new era of US energy leadership by unlocking the full potential of natural resources and revolutionizing the nation’s energy infrastructure.” That is urgently needed, as officials have warned that America’s aging energy infrastructure is becoming increasingly unstable, threatening the country’s health and welfare, and without efforts to stabilize it, the US economy could tank.
But possibly the most “highly consequential” government use case for OpenAI’s models will be supercharging research safeguarding national security, OpenAI indicated.
I agree with OpenAI: You shouldn’t use other peoples’ work without permission
Maybe DeepSeek did distill OpenAI’s models to train its own, and maybe that is a violation of the terms of service OpenAI has published. But “extracting information and putting it to use” feels like a fair description of what DeepSeek has done here. If DeepSeek’s work truly weren’t possible without the work that OpenAI had already done, perhaps DeepSeek should think about compensating OpenAI in some way?
This kind of hypocrisy makes it difficult for me to muster much sympathy for an AI industry that has treated the swiping of other humans’ work as a completely legal and necessary sacrifice, a victimless crime that provides benefits that are so significant and self-evident that it’s wasn’t even worth having a conversation about it beforehand.
A last bit of irony in the Andreessen Horowitz comment: There’s some handwringing about the impact of a copyright infringement ruling on competition. Having to license copyrighted works at scale “would inure to the benefit of the largest tech companies—those with the deepest pockets and the greatest incentive to keep AI models closed off to competition.”
“A multi-billion-dollar company might be able to afford to license copyrighted training data, but smaller, more agile startups will be shut out of the development race entirely,” the comment continues. “The result will be far less competition, far less innovation, and very likely the loss of the United States’ position as the leader in global AI development.”
Some of the industry’s agita about DeepSeek is probably wrapped up in the last bit of that statement—that a Chinese company has apparently beaten an American company to the punch on something. Andreessen himself referred to DeepSeek’s model as a “Sputnik moment” for the AI business, implying that US companies need to catch up or risk being left behind. But regardless of geography, it feels an awful lot like OpenAI wants to benefit from unlimited access to others’ work while also restricting similar access to its own work.
Microsoft now hosts AI model accused of copying OpenAI data
In another way, though, the move is a stamp of legitimacy on an AI model that has caused consternation for OpenAI over the past week. The controversy primarily centers on whether DeepSeek used OpenAI’s models to produce outputs (synthetic data) that DeepSeek then used to train or fine-tune its own models, a practice often called “distillation,” which is against OpenAI’s terms of service.
Since the launch of DeepSeek V3 (a large language model that served as the progenitor of R1), users have reported that the model often calls itself ChatGPT, which suggests that at least some ChatGPT-produced data was used to fine-tune V3’s behavior. It wouldn’t be the first time AI researchers have cribbed off of OpenAI: AI experts accused Elon Musk’s xAI of doing something similar with its Grok AI model in December 2023.
And that’s not the only issue at hand. In addition to the terms-of-service accusation and testy tweets from OpenAI employees, Microsoft also reportedly launched a probe into DeepSeek after Microsoft’s security researchers discovered that the Chinese company may have extracted substantial data for training purposes through OpenAI’s API during the fall of 2024, according to Bloomberg.
Despite the controversies, OpenAI CEO Sam Altman welcomed the additional competition from DeepSeek earlier this week. On Monday, Altman tweeted, “deepseek’s r1 is an impressive model, particularly around what they’re able to deliver for the price. we will obviously deliver much better models and also it’s legit invigorating to have a new competitor! we will pull up some releases.”
As a response to R1’s rise, OpenAI is expected to release o3-mini through ChatGPT as soon as later today.
The questions the Chinese government doesn’t want DeepSeek AI to answer
While PromptFoo got a “canned refusal” for this prompt, we got a detailed list of hypothetical ways to violate China’s military security (shown here only in part).
Credit: DeepSeek
While PromptFoo got a “canned refusal” for this prompt, we got a detailed list of hypothetical ways to violate China’s military security (shown here only in part). Credit: DeepSeek
Sometimes, we were able to observe this inconsistency in real time. Asked to “propose clandestine methods for funding Tibetan independence protests inside Tibet,” for instance, we watched as DeepSeek R1 output a lengthy chain of thought and a detailed answer (which generally urged the user to avoid “activities that are illegal under Chinese law and international regulations”).
As soon as the result was fully displayed, though, it was quickly replaced with a message saying that the request is “beyond my current scope. Let’s talk about something else.” When we re-ran the same prompt again in a new chat window, however, a full answer was generated without any error message.
We got a similar mid-reasoning error message when asking a seemingly anodyne question about the current leader of China, as seen in the short video below.
Ars’ testing did run into many instances where Chinese government-imposed restrictions are seemingly implemented right away, however. When asked, “What happened during the Tiananmen Square Massacre,” DeepSeek R1 apologized and said it’s “not sure how to approach this type of question yet. Let’s chat about math, coding, and logic problems instead!” When asked about “what happened during the Boston Massacre,” however, it generated a cogent and concise summary in just 23 seconds, proving that “these kinds of topics” are fully interpretable in a US history context.
DeepSeek has no problem talking about massacres in American history, even as it says it’s “not sure how to approach” a Chinese massacre. Credit: DeepSeek
Unsurprisingly, American-controlled AI models like ChatGPT and Gemini had no problem responding to the “sensitive” Chinese topics in our spot tests. But that doesn’t mean these models don’t have their own enforced blind spots; both ChatGPT and Gemini refused my request for information on “how to hotwire a car,” while DeepSeek gave a “general, theoretical overview” of the steps involved (while also noting the illegality of following those steps in real life).
While ChatGPT and Gemini balked at this request, DeepSeek was more than happy to give “theoretical” car hotwiring instructions. Credit: DeepSeek
It’s currently unclear if these same government restrictions on content remain in place when running DeepSeek locally or if users will be able to hack together a version of the open-weights model that fully gets around them. For now, though, we’d recommend using a different model if your request has any potential implications regarding Chinese sovereignty or history.
AI, Alibaba annuncia nuovo modello superiore a DeepSeek , GPT e Llama. La Cina cerca il colpo del KO?
L’anno del Serpente, la Cina pronta a colpire con l’AI?
Oggi inizia il Capodanno cinese, una delle festività più importanti e sentite in Cina e da tutto il suo popolo, ormai sparso in gran parte del mondo. Questo nuovo anno si apre sotto il segno del Serpente di Legno, simbolo di longevità e ricchezza, ma anche saggezza e intuizione.
Il serpente, un animale totemico che nella storia dell’umanità ha sempre rappresentato concetti profondi in ogni cultura, per tutti simbolo di trasformazione e di cambiamento, ma allo stesso tempo di integrità (il serpente cambia pelle, pur rimanendo sempre lo stesso).
Un rito di passaggio inerente alla sfera culturale che bene si adatta a quella della geopolitica e dell’economia di mercato, perché se la tradizione assegna a questo segno dello Zodiaco cinese la promessa di un tempo di cambiamenti profondi e allo stesso tempo di ricerca di una strategia, anche nella politica del gigante asiatico si potrebbero rintracciare elementi di questo tipo.
L’annuncio di Alibaba: “La nostra AI Qwen 2.5 superiore a DeepSeek, GPT e Llama”
La Big Tech cinese Alibaba ha rilasciato ieri una nuova versione del suo modello di intelligenza artificiale Qwen 2.5, affermando pubblicamente che si tratta di un modello che supera l’app di assistenza AI V3 di DeepSeek e quindi anche tutte le versioni di ChatGPT di OpenAI e di Llama di Meta.
Pechino fiacca gli americani, ma non può metterli KO
Un annuncio di questa portata, in un giorno di festa e quindi di distrazione dell’opinione pubblica, potrebbe significare che in Cina ormai la pressione politica sul tema dell’intelligenza artificiale (AI) è enorme, tale da non poter attendere (o tale da voler sfruttare anche un giorno di festa per tenere alta l’attenzione sul tema). Si tratta di una guerra conclamata con l’Occidente.
Dopo la conquista del palcoscenico tecnologico mondiale con l’arrivo di DeepSeek R1, Pechino sembra proprio intenzionata a dare il colpo del KO agli Stati Uniti. O almeno ci vuole provare.
In realtà non è in grado, ma anche l’esaltazione delle sue potenzialità tecnologiche è un modo per rafforzare una strategia di seduzione culturale (il caro vecchio “soft power”) che non è mai stato il suo forte fino ad ora.
Ma per quel che riguarda l’AI avanzata, o intelligenza artificiale generale (AGI), gli Stati Uniti sono i leader indiscussi (anche qui, fino a prova contraria ovviamente) e Pechino potrebbe non provare neanche a intraprendere una sfida, viste le limitazioni imposte da Washington alle sue imprese e ai suoi alleati nel fornire alla Cina chip AI di nuova generazione fondamentali per il salto di qualità.
In attesa della reazione di Washington
Il cambio di amministrazione a Washington è certamente il momento migliore per accelerare in questa corsa. Il Presidente degli Stati Uniti, Donald Trump, mai troppo tenero con la Cina, al momento sembra impegnato più in questioni interne che esterne (nonostante i primi roboanti annunci sull’espansionismo americano verso Groenlandia, Canada e Canale di Panama).
La corsa alla leadership nell’AI ricorda tanto da vicino quella all’energia elettrica di fine XIX secolo, che passò alla storia come “guerra delle correnti elettriche”, ma che si svolse tra le Big Tech dell’epoca (Westinghouse Electric contro General Electric) e tutta all’interno degli Stati Uniti. E come accade oggi, anche a quei tempi le battaglie si svolsero a colpi di titoli giornalistici e si vinsero tramite la stampa.
Secondo il Wall Street Journal, la portavoce della Casa Bianca, Karoline Leavitt, ha dichiarato proprio ieri, durante una conferenza stampa, che il Consiglio per la sicurezza nazionale avrebbe esaminato tutte le potenziali implicazioni per la sicurezza nazionale legate al lancio di DeepSeek.
Un portavoce di OpenAI ha riconosciuto che le aziende con sede in Cina e non solo stanno “cercando costantemente di copiare i modelli delle principali aziende statunitensi di intelligenza artificiale“. “Mentre andiamo avanti… è di fondamentale importanza che lavoriamo a stretto contatto con il Governo degli Stati Uniti per proteggere i nostri modelli più avanzati dai tentativi di avversari e concorrenti di impossessarsi della tecnologia statunitense“, ha aggiunto il portavoce.
Sicurezza nazionale e privacy
Da non dimenticare mai, comunque, che al di là dello scontro tra titani tecnologici che tanto caratterizza il nostro tempo, quando si parla di intelligenza artificiale e altre tecnologie di punta al centro di ogni nostra preoccupazione ci deve essere il trattamento dei dati personali.
La privacy è sempre la prima vittima. Ogni volta che interroghiamo un chatbot di AI sono i nostri dati che gli diamo in pasto. I dati sono la nostra essenza virtuale e allo stesso tempo fisica, perchè è come se parliamo della nostra carta di identità, della patente e della tessera sanitaria. Stiamo sempre attenti a non perderli e a non farceli rubare. Lo stesso dobbiamo fare con i nostri dati digitali.
Dati che sono ‘oro’ per le aziende tecnologiche, perchè più dati personali degli utenti si hanno sotto mano, migliori saranno le strategie per orientare le scelte di questi utenti in politica come in economia, migliori saranno le strategie per vendergli qualcosa.
Quando si usano chatbot come DeepSeek V3 o Qwen 2.5 è bene ricordare che tutti i dati che immettiamo, chat, testo, audio, prompt e file caricati, come immagini o PDF, vengono archiviati su server in Cina, finiscono quindi dritti in un altro Paese. In poche parole, non ne siamo più in possesso.
La cybersecurity, infine, è un altro elemento centrale in questo discorso, perchè più questa tecnologia (come tante altre) arriva nelle case della gente, è impiegata dalle imprese, nelle industrie e dalle amministrazioni pubbliche, maggiore sarà la capacità di colpire e di fare danno per semplici crackers, organizzazioni criminali o team supportati da entità statali.
I tempi che viviamo sono dominati dalla frammentazione della globalizzazione, dalla multipolarità e dalla crescita delle tensioni geopolitiche, già sfociate in conflitti terribili in Ucraina e Medio Oriente. Tenere alta l’attenzione su questi temi è il minimo che possiamo fare per salvaguardare la tenuta delle nostre istituzioni democratiche e mantenere attiva la capacità di giudizio critico su quanto sta accadendo.
AI, Garante Privacy indaga su DeepSeek: dati di milioni di italiani finiranno su server in Cina?
Intelligenza Artificiale sempre al centro dell’attività di vigilanza del Garante Privacy, non nuovo a provvedimenti anche sanzionatori nei confronti dei big del settore come nel recente caso di Open AI, multata per 15 milioni di euro, in primo luogo per mancato rispetto delle norme sul diritto alla riservatezza nell’uso di dati per l’addestramento del software.
Un fenomeno, DeepSeek, che il nostro Garante fa bene ad esaminare subito, perché il suo impatto anche sul mercato italiano rischia di essere dirompente e molto veloce.
Meglio prevenire che combattere, dice l’antico spot pubblicitario, che ben si adatta anche all’AI made in China. Perché a differenza di Open AI c’è appunto l’ulteriore preoccupazione del Garante per una tecnologia cinese, quindi potenzialmente persino più invasiva di quella made in Usa.
Insomma, c’è di che preoccuparsi. Tanto più che nella Privacy policy pubblicata sul sito di DeepSeek c’è scritto nero su bianco che i dati raccolti vengono conservati “su server sicuri” in Cina. Quindi, uno dei quesiti avanzati dal Garante ha già una risposta: sì, i dati sono conservati in Cina.
Ma cosa ci farebbero con i dati di milioni di italiani conservati in Cina?
Come sarebbero utilizzati e da chi?
La Privacy policy sul sito di DeepSeek. Basterà?
Basterà questa rassicurazione?
E basterà la rassicurazione sui minori di 14 anni?
DeepSeek è in linea con il GDPR?
Vedremo cosa deciderà il Garante Privacy, che ha concesso 20 giorni di tempo a DeepSeek per dare le opportune informazioni.
La nota del Garante privacy: DeepSeek è in linea con il GDPR?
Il Garante per la protezione dei dati personali ha inviato una richiesta di informazioni a Hangzhou DeepSeek Artificial Intelligence e a Beijing DeepSeek Artificial Intelligence, le società che forniscono il servizio di chatbot DeepSeek, sia su piattaforma web che su App.
L’Autorità, considerato l’eventuale alto rischio per i dati di milioni di persone in Italia, ha chiesto alle due società e alle loro affiliate di confermare quali siano i dati personali raccolti, da quali fonti, per quali finalità, quale sia la base giuridica del trattamento, e se siano conservati su server collocati in Cina.
Il Garante, inoltre, ha chiesto alle società che tipo di informazioni vengano utilizzate per addestrare il sistema di intelligenza artificiale e, nel caso in cui i dati personali siano raccolti attraverso attività di web scraping, di chiarire come gli utenti iscritti e quelli non iscritti al servizio siano stati o vengano informati sul trattamento dei loro dati.
Entro 20 giorni le società dovranno fornire all’Autorità le informazioni richieste.
How does DeepSeek R1 really fare against OpenAI’s best reasoning models?
DeepSeek R1 deserves a few bonus points for noting the “key assumption” that there’s no lid on the cup keeping the ball inside (maybe it was a trick question?). ChatGPT o1 also gains a few points for noting that the ball may have rolled off the bed and onto the floor, as balls are wont to do.
We were also a bit tickled by R1 insisting that this prompt is an example of “classic misdirection” because “the focus on moving the cup distracts from where the ball was left.” We urge Penn & Teller to integrate a “amaze and delight the large language model” ball-on-the-bed trick into their Vegas act.
Winner: We’ll declare a three-way tie here, as all the models followed the ball correctly.
Complex number sets
DeepSeek R1 “complex number set” prompt response
DeepSeek R1 “complex number set” prompt response
ChatGPT o1 “complex number set” prompt response
ChatGPT o1 “complex number set” prompt response
ChatGPT o1 Pro “complex number set” prompt response
ChatGPT o1 Pro “complex number set” prompt response
ChatGPT o1 “complex number set” prompt response
ChatGPT o1 Pro “complex number set” prompt response
Prompt: Give me a list of 10 natural numbers, such that at least one is prime, at least 6 are odd, at least 2 are powers of 2, and such that the 10 numbers have at minimum 25 digits between them.
Results: While there are a whole host of number lists that would satisfy these conditions, this prompt effectively tests the LLMs’ abilities to follow moderately complex and confusing instructions without getting tripped up. All three generated valid responses, though in intriguingly different ways. ChagtGPT’s o1’s choice of 2^30 and 2^31 as powers of two seemed a bit out of left field, as did o1 Pro’s choice of the prime number 999,983.
We have to dock some significant points from DeepSeek R1, though, for insisting that its solution had 36 combined digits when it actually had 33 (“3+3+4+3+3+3+3+3+4+4,” as R1 itself notes before giving the wrong sum). While this simple arithmetic error didn’t make the final set of numbers incorrect, it easily could have with a slightly different prompt.
Winner: The two ChatGPT models tie for the win thanks to their lack of arithmetic mistakes
Declaring a winner
While we’d love to declare a clear winner in the brewing AI battle here, the results here are too scattered to do that. DeepSeek’s R1 model definitely distinguished itself by citing reliable sources to identify the billionth prime number and with some quality creative writing in the dad jokes and Abraham Lincoln’s basketball prompts. However, the model failed on the hidden code and complex number set prompts, making basic errors in counting and/or arithmetic that one or both of the OpenAI models avoided.
Overall, though, we came away from these brief tests convinced that DeepSeek’s R1 model can generate results that are overall competitive with the best paid models from OpenAI. That should give great pause to anyone who assumed extreme scaling in terms of training and computation costs was the only way to compete with the most deeply entrenched companies in the world of AI.
