In this article
Today we get into YouTube as a source for learning about artificial intelligence, and the short answer is qualified. Yes, you can learn. But most of what circulates under the label of "AI science communication on YouTube" isn't science communication; it's entertainment with a technical vocabulary. How do you tell one from the other without turning into the tired guy at the bar who criticises everything? There's a concrete test I'm about to lay out. My opinion is biased, because I've learned valuable things on YouTube and also lost afternoons to courses that teach nothing. Form your own: try the test on your favourite channel and judge.
For two years I've kept the YouTube tab open in parallel with the paper I'm reading. I've watched hundreds of AI science-communication videos in Spanish and English. The conclusion is that a very small percentage of that consumption leaves operational knowledge, and a large percentage produces the sense of knowing without knowing. The sense is real. The knowledge isn't.
The economics of the channel explains almost everything
Before talking about specific channels it's worth understanding why the vast majority fail, and the answer isn't individual. It's structural.
A YouTuber aiming to live off their channel — or to make it profitable as a second job — needs to publish at least one video a week, ideally two or three. Each video needs a hook in the first fifteen seconds: a striking claim, a provocative question, a promise of revelation. If the viewer doesn't get hooked in those fifteen seconds, they leave, and the retention metric drops. If retention drops, YouTube's algorithm reduces the reach of the next video. If reach drops, ad revenue drops and new subscriptions drop. The loop is clear.
The algorithm, after that, rewards sustained retention throughout the whole video. That means every minute must contain narrative tension, surprise or a small cliffhanger. The slow explanation, the nuance that doesn't resolve, the honest doubt, the technical caveat that doesn't fit the main thesis: all of that is the enemy of retention. The editor cuts it. Not out of bad faith; out of operational necessity.
The result is a format where the boring is systematically eliminated. But sometimes the boring is the important part. When the original paper a video makes eight minutes out of discusses, in its limitations section, that the method doesn't work in non-English languages, that footnote doesn't make it into the video. Not out of censorship; out of editing. And yet that footnote changed the meaning of the paper.
What happens to a thirty-page paper in eight minutes
Let's take a concrete case, without naming the YouTuber so as not to turn them into an individual case study. Imagine a typical DeepMind or Anthropic paper. Thirty pages. Three big tables. Fifteen relevant references. A method section taking up five pages. A results section with six benchmarks. A limitations section a page and a half long. An appendix of fourteen pages with technical details.
The eight-minute video has to produce a script of roughly 1,200 words spoken at standard speed. That allows you to cover, generously, a page and a half of the paper. The decisions about what gets in are brutal. Almost always in goes the most spectacular benchmark figure, a simplified explanation of the method, and a conclusion that reproduces the paper's own. Almost never in go the limitations, the counterfactuals, the comparison with prior work, or the details of the evaluation regime.
This isn't the YouTuber's defect. It's the format's defect. No human being can honestly explain a thirty-page paper in eight minutes. What gets done is selling a simplified version that keeps the paper's hits without the precision that frames them.
The serious part is that the viewer doesn't notice the loss. They leave the video believing they've understood the paper. Ask them two weeks later what the paper said, and they repeat the benchmark figure and little more. Ask them what limitations the paper itself declared, and they can't answer. That asymmetry between the sense of knowing and operational knowing is exactly what the format produces when applied to technical content.
The ones who do teach
There are notable exceptions and it's worth naming them, because they prove something else can be done within the format itself.
Carlos Santana, known on YouTube as Dot CSV, communicates in Spanish with technical judgement, data and code. His videos are long when they need to be — thirty, forty minutes, sometimes more — and they explain the method in detail, not just the result. When he reviews a paper, he reads it whole and shows the code if the code is available. When he talks about a benchmark, he explains the methodology before citing the figure. His videos on convolutional neural networks, on transformers or on the alignment problem are required references in Spanish. He isn't perfect and sometimes gets swept up by the news cycle, but the quality floor is well above the average.
3Blue1Brown — Grant Sanderson's channel, in English — is probably the best maths channel on YouTube. His series on neural networks — available at youtube.com/@3blue1brown — explains the maths behind backpropagation, transformers and deep learning with carefully designed visualisations. Sanderson doesn't do fast science communication; he takes weeks to produce each video, uses his own mathematical animation tool called Manim, and builds intuitions that stick. It isn't entertainment. It's teaching.
Andrej Karpathy — former OpenAI researcher and former director of AI at Tesla — has published on his personal YouTube channel a series of free classes on how to build large language models from scratch. The series includes two-to-four-hour episodes where Karpathy builds, live and line by line of code, a GPT-type model. Whoever endures those hours comes out understanding the method operationally. It's the modern equivalent of a free university course.
There are other serious channels. Yannic Kilcher reads papers in English with some detail. Two Minute Papers — Károly Zsolnai-Fehér — is communicative but honest about the simplified nature of its format. Robert Miles AI Safety covers the alignment problem with rigour. Welch Labs explores applied maths with visual care. There are also smaller but honest efforts in Spanish, on smaller channels worth following even if they don't compete on algorithm.
What unites all these channels isn't the language or the size. It's the willingness to sacrifice retention for precision when the topic demands it. That willingness is exactly what the algorithm penalises. That's why the honest channels grow more slowly than the spectacular ones. And that's why the serious viewer has to seek them out actively, not wait for the feed to recommend them.
The problematic ones
There are three channel patterns worth recognising and avoiding.
The first is the channel that sells paid AI courses with affiliations to companies or other courses. The typical scheme is: short free videos that generate subscribers, a call at the end to enrol in a paid course of 200 to 1,500 euros, positioning as a sector authority. The operational problem is that the paid course content, in the vast majority of cases, replicates free material available elsewhere — including, ironically, the channels of Dot CSV, 3Blue1Brown and Karpathy. What you're paying for isn't unique content; it's access to the instructor and a community. That can be worth it for some profiles, but it's worth knowing before you pay.
The second pattern is the channel that lives off spectacular demos with no context. "Look what this new model does." "You won't believe what GPT-5 just generated." "This prompt broke the internet." Each video is a demo, with no discussion of the method, no independent verification, no context on what that task is and what degree of success or failure it represents. The viewer accumulates a sense of progress without accumulating knowledge. It's the audiovisual equivalent of the headline with no paragraph: pretty, empty, forgettable.
The third pattern is the channel of permanent news. "The new model X is the best." "Anthropic has launched this." "What Sam Altman said this week." One video a week, always about the most recent launch, with no perspective, no cross-checking, no memory of what was said three months ago. It's the equivalent of the agency wire translated and voiced over, with the difference that it takes up ten minutes of the viewer's time instead of one of reading. Informational benefit per unit of time: very low.
The concrete test
There's a simple routine to judge, in the cold light, whether a YouTube video taught you something or just entertained you. Four questions, thirty seconds.
First question: can I name the original paper, dataset or specific source being discussed? If the video is about a new model and at the end I don't know who the first author of the paper is, what conference it was published at or what repository the code is in, the video told me a headline, not a piece of work.
Second question: can I list the limitations declared by the work's own authors? Almost every serious paper includes a limitations section and discussing them is part of understanding the result. If the video doesn't mention them, it sold the clean result without the fine print.
Third question: can I formulate a counterfactual? That is, what would have happened if the experiment had been done another way, with another dataset, with another base model? If the only hypothesis I can articulate is "this works," I haven't understood anything operational; I've memorised a slogan.
Fourth question: can I name relevant prior work this one compares with? Almost all serious work in AI continues or criticises a previous research line. If the video doesn't place it, it presents it as if it had appeared in a vacuum. And in science, nothing appears in a vacuum.
If the video fails all four questions, it didn't teach me. It entertained me. That isn't bad in itself — entertaining yourself is legitimate — but it's worth not confusing the two. If it passes one or two, it's soft science communication. If it passes three or four, it's serious science communication, and it deserves time and a subscription.
The operational rule
Three habits change the YouTube diet from crushing to useful.
The first is to turn off the feed's automatic recommendation. YouTube pushes toward high-retention content, and as we've just seen, high retention is inversely correlated with depth. The diet based on what the algorithm recommends tends to degrade over time. Better to go directly to Dot CSV's channel, 3Blue1Brown's or Karpathy's, watch what's new and leave.
The second is to prefer long videos to short ones when the topic is technical. A forty-minute video on transformers is five times more likely to teach you something than five eight-minute videos on assorted topics. Popular intuition says the opposite; it's wrong.
The third is to always complement with text. After watching a serious video on a paper, go to arXiv, open the paper, read at least the introduction and the limitations section. Thirty minutes. That reading solidifies what the video left in suspension.
The political question
This is personal opinion, but the economics of the medium hold it up. The share of video time devoted to serious AI science communication is marginal against the volume of spectacular content. As long as the audience keeps rewarding the short sensationalist format — and it rewards it with its time, not its principles — that format will remain dominant. The aggregate quality of the ecosystem won't improve through the goodwill of the honest YouTubers; it will change only when the viewer's diet changes.
What the platforms could do, but won't, is adjust their algorithms not to penalise length when the content is verifiably educational. What the advertisers could do, but won't, is shift budget from the short format to the long when the long teaches. What the viewer can do, and depends only on them, is choose consciously.
The concrete figure that closes it: according to the Reuters Institute Digital News Report 2025, 31% of users aged 18 to 24 in Spain name YouTube as one of their main sources of information about technology, against 14% who cite a general-interest press outlet. That figure describes the shift. Whoever ignores that AI coverage has moved to YouTube ignores where the next generation's opinion is formed today. But whoever doesn't distinguish, within YouTube, what teaches from what entertains, stays at the same starting point they're complaining about.
Definitions
Hook: in YouTube jargon, the first fifteen to thirty seconds of a video, designed to stop the viewer from leaving. Its existence conditions the structure of the rest of the video.
Retention: the percentage of viewers who keep watching the video at each minute. It's the metric that weighs most in YouTube's recommendation algorithm, alongside the click-through rate on the thumbnail.
Science communication: the practice of conveying technical content to a non-specialist audience with the aim of transmitting operational knowledge, not just a sense of comprehension. Good science communication admits complexity without hiding it.
Paid affiliation course: a training product whose business model depends partly on capture through free content on social media. Its quality varies a lot; it's worth comparing before buying.
References
Carlos Santana, Dot CSV (youtube.com/@DotCSV). A serious science-communication channel in Spanish with an emphasis on code and method.
Grant Sanderson, 3Blue1Brown (youtube.com/@3blue1brown). An applied-maths channel with a visual emphasis; an accessible and rigorous series on neural networks.
Andrej Karpathy, personal channel (youtube.com/@AndrejKarpathy). Free classes on building GPT models from scratch.
Reuters Institute for the Study of Journalism, Digital News Report 2025 (Oxford University, June 2025). Figures on news consumption by source and by age bracket.
Yannic Kilcher, critical paper-reading channel (youtube.com/@YannicKilcher). An English-language reference for analysing technical papers.
Robert Miles AI Safety, channel on alignment (youtube.com/@RobertMilesAI). Rigorous science communication on the problem of controlling AI systems.
Welch Labs, applied-maths channel (youtube.com/@WelchLabsVideo). Careful visual support for technical concepts.
Going deeper
Nicholas Carr, The Shallows: What the Internet Is Doing to Our Brains (W. W. Norton, 2010). A frame for understanding how fragmented attention forms with continuous audiovisual consumption.
Cal Newport, Deep Work (Grand Central Publishing, 2016). An operational argument for why prolonged attention isn't optional for learning complex subjects.
Neil Postman, Amusing Ourselves to Death (Viking, 1985). A classic text on how the audiovisual medium transforms the substance of the content transmitted.
Maryanne Wolf, Reader, Come Home: The Reading Brain in a Digital World (Harper, 2018). Neurocognitive research on the difference between learning by video and learning by long text.
You might also like
- Three months without reading headlines to understand AI
- Stratechery, Platformer and 404 Media, AI journalism done seriously
- Ten reasons never to read another listicle about AI in your life
- Why AI master's degrees dominate Google's results
Comments0
No comments yet.
Leave a comment