THE INTERNET IS EATING ITSELF The Internet Is Now 52% AI-Generated Content And It's Training On Itself
New research just dropped numbers that should terrify anyone who cares about truth:
**52% of all internet content is now AI-generated.**¹ In 2022, it was 10%.
But here's where it gets insane (actually it's all insane TBH):
**74% of ALL NEW web pages contain AI-generated content.**² The internet added 3-5 billion pages monthly in 2024 most of it synthetic.³
The internet isn't just being eaten. It's being mass-produced by the thing that's eating it.
Why This Matters
Large Language Models aren't brains. They're giant stomachs. They consume everything. Digest nothing. Excrete more content, which gets consumed again an infinite feedback loop of synthetic regurgitation.
Here's what happens when AI trains on AI:
→ Model collapse: Recursive training causes "irreversible performance decay."⁴ → Narrowing of knowledge: Models reflect themselves, not reality → Death of originality: A hall of mirrors, each reflection dimmer than the last
We're replacing:
Human nuance Cultural context Real expertise Original thought The truth
With statistically probable simulations.
The Economics
Licensing real content costs billions. Synthetic data? Almost nothing. So they choose cheap scale over real knowledge. No regulation. No transparency. No tracking.
OpenAI doesn't ask permission to train on Anthropic's outputs. They just scrape the web.
Competition accelerates the collapse. Every AI company races to build bigger models.
They need more data. Synthetic data looks like a shortcut. Collectively, they're destroying the foundation their business depends on: real human knowledge.
We've Already Passed the Tipping Point
What happens when:
→ Medical information trains on synthetic medical papers? → Children learn history from recursive AI summaries? → Scientific research builds on fabricated datasets?
We don't just lose quality. We lose the ability to know what's real. The internet was humanity's collective memory. Now it's becoming humanity's collective hallucination.
The Bottom Line
LLMs are giant stomachs, not brains. They consume. They excrete. They consume again. ******************************************************************************** The trick with technology is to avoid spreading darkness at the speed of light Stephen Klein | Founder & CEO, Curiouser.AI | Teaches AI Ethics at UC Berkeley | Raising on WeFunder | Would your support (LINK IN COMMENTS)
Footnotes:
¹ Graphite Research (2025). Analysis of 65,000 URLs from Common Crawl. ² Ahrefs (2025). Analysis of 900,000 newly created web pages in April 2025. ³ Common Crawl Foundation. Database adds 3-5 billion pages monthly. ⁴ Shumailov et al. (2024). "AI models collapse when trained on recursively generated data." Nature, 631, 755-759. | 388 comments on LinkedIn
"If you like interesting snippets on all sorts of subjects relevant to academia, information, the world, highly recommended is @grip54 's collection:"
La curation de contenus, la mémoire partagée d'une veille scientifique et sociétale
Gilbert C FAURE's insight:
... designed to collect posts and informations I found and want to keep available but not relevant to the other topics I am curating on Scoop.it (on behalf of ASSIM):
because we have a long standing collaboration through a french speaking medical training program between Faculté de Médecine de Nancy and WuDA, Wuhan university medical school and Zhongnan Hospital
🧠📚L’un de mes premiers souvenirs marquants de contact avec le monde de la publication scientifique remonte à ma première année de licence. Un professeur nous avait montré une image frappante : si l’on empilait toutes les publications scientifiques produites chaque jour dans le monde, on atteindrait des distances astronomiques… jusqu’à la Lune, voire le Soleil. Cette métaphore m’a suivie tout au long de mon parcours. Pendant ma thèse, j’ai vite compris à quel point rester à jour sur la bibliographie de son propre domaine est un défi colossal. Et j’ai découvert une frustration bien connue des chercheurs : plus on lit, plus on réalise l’immensité de ce qu’il nous reste à lire et à comprendre.
📅 La semaine dernière, j’ai eu le plaisir de participer à une formation animée par Assia ASRIR, PhD sur l’outil Opscidia. Une vraie révélation.
✨ Opscidia, c’est une IA fiable, exhaustive, qui ne génère pas de sources fictives. Le rêve de toute personne ayant déjà tenté de réaliser un état de l’art rigoureux. Un outil modulable, qui permet de gagner du temps sans sacrifier la rigueur, tout en gardant la main sur le processus de rédaction. C’est aussi un excellent moyen de rester à jour sur les sujets que l'on maîtrise… mais surtout sur ceux que l'on rêve d’approfondir.
🙏 Merci Assia pour cette découverte ! J’ai hâte de continuer à explorer les possibilités de cet outil.
La solution idéale pour le traitement des eaux usées de votre habitation principale. Une micro-station avec plus de 10 ans d'expérience et de satisfaction.
💙 🤍 ❤️ MedGPT a un mois ! Et on n’aurait jamais espéré un tel succès. Ce qui devait être une simple bêta est devenu un mouvement 🧑🧑🧒🧒
Quelques chiffres pour l'illustrer : ➡️ 615 professionnels de santé nous ont écrit spontanément pour notamment proposer de participer à son élaboration ! Une communauté se forme déjà.
➡️ 35.000 décisions cliniques accompagnées ! Ramené à la population de professionnels de santé, ce n'est "que" 15 fois moins que OpenEvidence sur la même période. A notre rythme de croissance actuel, on vise le même ratio d'utilisation national qu'OpenEvidence dans 5 mois. Les US ont pris une longueur d'avance, mais la France sait courir vite 🏃♂️➡️
➡️ Des centaines de retours positifs, des encouragements et plusieurs dizaines d'articles dans la presse.
Et MedGPT ne va pas s'arrêter là, des nouveautés arrivent : ) Pour tous ceux qui n'ont pas encore essayé le produit : c'est dispo et c'est gratuit, allez-y ! | 15 comments on LinkedIn
We talk about health literacy as if it lives inside people — as if the solution is to hand individuals better tools, clearer brochures, simpler language, and hope they can “navigate” the system. But that framing is fundamentally wrong. Health literacy is not an individual skill problem. It is a system design problem.
If a person struggles to understand, act, or make informed decisions, that is not a sign of their failure. It is a sign that the environment was not built to support them. It is an organizational failure, a policy failure, a leadership failure — a failure of design.
Health literacy is not about teaching people to try harder. It is about building systems that make health understanding, access, and action natural — not heroic. It is a matter of equity and power, not worksheets and pamphlets.
The true measure of a health-literate society is not how well individuals adapt to complexity — but how well institutions remove the complexity in the first place.
Until we shift the responsibility from people to systems, from coping to designing, from deficit to empowerment — we will keep treating symptoms while ignoring the root cause. | 11 comments on LinkedIn
Review articles used to be essential for scientific publishing - an important academic exercise, while reading them was important for anyone entering a field.
This week, I learned about Consensus App (thanks, Julian A. Serna), which seems to generate excellent, referenced summaries on any topic — often better than many “real” review papers.
So, my questions are: -Does it still make sense for humans to write review articles? Especially since no one today can realistically read and process all relevant papers in an active field. -If AI can already produce (and will soon perfect) summaries that are comprehensive, accurate, and continuously updated — what unique value does a traditional human-written (and probably with the use of AI anyway) review still add?
It’s also interesting what this shift means for publishers like Wiley, Elsevier, or Springer, whose journal impact factors often rely heavily on review articles.
My prediction is that the traditional concept of a “review paper” will soon lose its relevance.
Les plateformes sociales entrent dans une nouvelle ère : celle du désenchantement. La baisse mondiale du temps passé en ligne, combinée à l’arrivée massive de contenus générés par l’IA, révèle un tournant dans les usages.
1️⃣ L’ère du contenu ultra-transformé
Meta et OpenAI lancent leurs propres plateformes de vidéos générées par IA. Leur pari : que les utilisateurs aient encore envie de créer et surtout de consommer toujours plus de vidéos. Sauf que cette logique pousse à l’extrême ce que l’auteur du Financial Times appelle un « contenu ultra-transformé ». Dopamine garantie, valeur informationnelle quasi nulle…. Et perte de sens assurée.
2️⃣ Le reflux de l’attention
Les chiffres sont très intéressants. Selon une analyse du cabinet GWI, le temps passé sur les réseaux a atteint un pic en 2022 avant de chuter de près de 10 % fin 2024. Les 16 ans et plus y consacrent en moyenne deux heures et vingt minutes par jour. Et même les jeunes surconnectés décrochent les premiers.
💎En fait, les plateformes, devenues des « machines à capter l’attention », ne sont plus vraiment sociales.
THE INTERNET IS EATING ITSELF The Internet Is Now 52% AI-Generated Content And It's Training On Itself
New research just dropped numbers that should terrify anyone who cares about truth:
**52% of all internet content is now AI-generated.**¹ In 2022, it was 10%.
But here's where it gets insane (actually it's all insane TBH):
**74% of ALL NEW web pages contain AI-generated content.**² The internet added 3-5 billion pages monthly in 2024 most of it synthetic.³
The internet isn't just being eaten. It's being mass-produced by the thing that's eating it.
Why This Matters
Large Language Models aren't brains. They're giant stomachs. They consume everything. Digest nothing. Excrete more content, which gets consumed again an infinite feedback loop of synthetic regurgitation.
Here's what happens when AI trains on AI:
→ Model collapse: Recursive training causes "irreversible performance decay."⁴ → Narrowing of knowledge: Models reflect themselves, not reality → Death of originality: A hall of mirrors, each reflection dimmer than the last
We're replacing:
Human nuance Cultural context Real expertise Original thought The truth
With statistically probable simulations.
The Economics
Licensing real content costs billions. Synthetic data? Almost nothing. So they choose cheap scale over real knowledge. No regulation. No transparency. No tracking.
OpenAI doesn't ask permission to train on Anthropic's outputs. They just scrape the web.
Competition accelerates the collapse. Every AI company races to build bigger models.
They need more data. Synthetic data looks like a shortcut. Collectively, they're destroying the foundation their business depends on: real human knowledge.
We've Already Passed the Tipping Point
What happens when:
→ Medical information trains on synthetic medical papers? → Children learn history from recursive AI summaries? → Scientific research builds on fabricated datasets?
We don't just lose quality. We lose the ability to know what's real. The internet was humanity's collective memory. Now it's becoming humanity's collective hallucination.
The Bottom Line
LLMs are giant stomachs, not brains. They consume. They excrete. They consume again. ******************************************************************************** The trick with technology is to avoid spreading darkness at the speed of light Stephen Klein | Founder & CEO, Curiouser.AI | Teaches AI Ethics at UC Berkeley | Raising on WeFunder | Would your support (LINK IN COMMENTS)
Footnotes:
¹ Graphite Research (2025). Analysis of 65,000 URLs from Common Crawl. ² Ahrefs (2025). Analysis of 900,000 newly created web pages in April 2025. ³ Common Crawl Foundation. Database adds 3-5 billion pages monthly. ⁴ Shumailov et al. (2024). "AI models collapse when trained on recursively generated data." Nature, 631, 755-759. | 388 comments on LinkedIn
I discovered that a reading pack for my doctoral leadership subject contained fabricated references. Almost everything listed was AI-generated citations that either didn’t exist or linked to the wrong papers.
When I raised it, the provider confirmed that AI had been used and that the material was shared before human review. They also reminded me that doctoral candidates should be able to verify their own sources.
That response was so disappointing. Doctoral candidates are expected to build on verified scholarship, not correct institutional errors. I’ve asked to withdraw from the course because the university doesn’t seem to understand why this is a serious concern and has pushed the responsibility back on me.
Distributing unverified academic material in a foundation subject is a breach of academic integrity and sets entirely the wrong ethical tone for the course.
Am I overreacting? Or is this yet another symptom of the wider issues that are undermining confidence in the sector? | 306 comments on LinkedIn
My first visit to Tbilisi, Georgia for the International Conference on Medical Education has been incredible and filled with thoughtful discussions, engaged learners, and the perfect mix of local and international perspectives. Thanks to Salome Voronovi for the invitation, and always nice to see David Taylor.
A concept that really struck a chord is what I’ve started calling the Suitcase Paradox:
In lifelong learning or curriculum design, just like when packing for a trip, you can’t keep adding new things unless you take something out first. And you have to fit it into the overhead compartment on a plane and in our anatomical overhead compartment, the brain!
Healthcare professionals must continually unlearn outdated practices to make room for new evidence, new technologies, and new ways of thinking.
That’s what lifelong learning, and particularly continuing professional development (CPD), is all about.
But to make it work, educators must evolve into learning facilitators, helping learners curate, adapt, and apply knowledge depending on where they are on the learning continuum.
And because healthcare doesn’t happen in silos, neither should learning. Interprofessional education (IPE) brings students from different health professions together.
Interprofessional continuing education (IPCE) extends that collaboration into practice. And when it’s done right, it leads to interprofessional collaborative practice (IPCP), where the ultimate outcome is better patient care.
I even got in a mention of the Donald & Barbara Zucker School of Medicine curriculum!
Plenty more to come: I’ve still got a wine tour 🍷 ahead and a masterclass on lifelong learning and CPD on Monday!
J’ai le plaisir de vous annoncer la tenue de la journée d’étude « Penser l’éducation avec John Dewey : une approche pragmatiste et pluridisciplinaire », qui se déroulera à Paris le 1er décembre 2025: https://lnkd.in/ehh5MbTx
Cet événement, co-organisé par Annabelle Cara, Anne BARRERE et moi-même, avec le soutien du CERLIS, de l’EDA et du Centre de recherche sur les médiations (Crem), proposera un regard croisé sur l’œuvre de Dewey et ses apports pour penser les questions et pratiques éducatives. Avec la participation de Renaud Hétier, Anne Lehmans, Samuel Renier, Sébastien-Akira Alix, Céline Robillard et Arthur Ancelin.
Une étude universitaire américaine met en lumière les limites de l’intelligence artificielle dans le domaine de l’éducation. Menée par les chercheurs Torrey Trust et Robert Maloy de l’Université du Massachusetts Amherst, cette analyse de plus de 300 plans de cours générés par ChatGPT, Gemini et Copilot conclut que l’IA, dans sa forme actuelle, échoue à…
This video is the FULL interview of the Business of Colorado segment featured on Studio Twelve.
From Studio Twelve: Business of Colorado, Frannie Matthews interviews Nicholas Sly of the Federal Reserve on Colorado’s economic trends, inflation challenges, and the evolving job market in the AI era.
étude, 1er semestre 2025 sur réponses à des questions d'actualité de 4 assistants d'IA : Copilot, ChatGPT, Perplexity et Gemini. Bilan : un % massif de reponses contenant des erreurs ! IA n’est pas info !
AI slop–low-quality, often fake AI-generated content – is proliferating at a staggering rate. So what do you and your students need to know to combat it?
"AI slop is the low-quality, often fake content, such as text, images, or videos, that is generated by AI. It’s currently overwhelming social media and the internet"
New paper just published with Shaydanay Urbani and Eric Wang. We wanted to understand how people searched for health information on AI-powered technologies (specifically ChatGPT, Alexa and Gemini Overviews on Google Search results), so we interviewed 27 people while watching their behavior and asking for additional information about why they were doing certain things and what they would do with the results. https://lnkd.in/eMfZUcxD
tl;dr : Participants integrated AI tools into their broader search routines rather than using them in isolation. ChatGPT was valued for its clarity, speed, and ability to generate keywords or summarize complex topics, even by users skeptical of its accuracy. Trust and utility did not always align; participants often used ChatGPT despite concerns about sourcing and bias. Google’s AI Overviews were met with caution—participants frequently skipped them to review traditional search results. Alexa was viewed as convenient but limited, particularly for in-depth health queries. Platform choice was influenced by the seriousness of the health issue, context of use, and prior experience. One-third of participants were multilingual, and they identified challenges with voice recognition, cultural relevance, and data provenance. Overall, users exhibited sophisticated “mix-and-match” behaviors, drawing on multiple tools depending on context, urgency, and familiarity.
Fascinating project, and as ever, people's behavior tends to be much more complex and nuanced than headlines would suggest.
Nous avons un cadeau pour vous : un tout nouveau MOOC gratuit sur l’intelligence artificielle ! 🎁
Quel que soit votre métier – industrie, services, médecine, art – l’IA peut devenir votre meilleur allié.
Ce MOOC a été conçu pour vous aider à la pratiquer concrètement, selon 3 parcours : > Débutants : pour découvrir pas à pas l’IA. > Professionnels : pour ceux qui ont déjà testé et veulent aller plus loin. > Experts : pour créer et utiliser des agents IA.
💡 Vous y trouverez des contenus exclusifs issus d’expériences du monde entier, bien au-delà des classiques formations des GAFAM ou des YouTubers.
L’objectif ? Vous permettre de transformer vos compétences et votre métier grâce à l’IA.
Réalisé au Cnam (Conservatoire national des arts et métiers) avec le #LearningLabHumanChange
📅 Rendez-vous le 13 octobre 2025 à 18h pour le lancement !
Conservatoire National des Arts et Métiers CAIRE | 52 comments on LinkedIn
To get content containing either thought or leadership enter:
To get content containing both thought and leadership enter:
To get content containing the expression thought leadership enter:
You can enter several keywords and you can refine them whenever you want. Our suggestion engine uses more signals but entering a few keywords here will rapidly give you great content to curate.
Your new post is loading...
