Habsburg AI
noun

- An AI model degraded through repeated training on AI-generated outputs.

- An generative AI system suffering from reduced diversity and originality because it learns from synthetic data.

- An metaphor describing AI "inbreeding."

- An AI system whose outputs become increasingly generic, distorted, or repetitive over successive generations.

- Jathan Sadowski (original coinage), Medium, Futurism, Making Science Public

—

PHRASE ORIGIN


The phrase combines: Habsburg — the famous European ruling dynasty, the House of Habsburg, whose extensive intermarriage over centuries became associated with inherited physical traits such as the "Habsburg jaw" and AI (Artificial Intelligence).

The metaphor suggests that AI models repeatedly trained on the outputs of other AI models resemble a royal family repeatedly marrying within its own bloodline: variation decreases, weaknesses accumulate, and distinctive traits become exaggerated. The term was coined by Jathan Sadowski in 2023. (X (formerly Twitter))

—

WHEN IDEAS INTERMARRY


For centuries, the Habsburgs ruled much of Europe. Their strategy was famously summed up as: "Let others wage war; you, fortunate Austria, marry."

The policy worked brilliantly politically. Genetically, not so much. Generation after generation, marriages occurred within a remarkably small pool of aristocratic families. Over time, distinctive traits became more pronounced. Historians and medical researchers still discuss the famous "Habsburg jaw" that appeared repeatedly among members of the dynasty.

Something similar happens to ideas,… early internet content was created by millions of people. Journalists, scientists, novelists, teachers, hobbyists, engineers, conspiracy theorists, teenagers, grandmothers, comedians, and bored office workers all contributed their tiny pieces to the world's knowledge. Then came generative AI.

Initially, AI learned from humans. But increasingly, humans are reading AI-generated articles, summaries, social media posts, product descriptions, reviews, and even academic drafts. Future AI systems may therefore end up learning from content created by earlier AI systems. In other words, the descendants begin marrying the descendants.

What's fascinating is that humans suffer from a similar problem. A company that only hires people who think alike becomes intellectually inbred. A friendship circle that shares identical opinions gradually loses perspective. A manager who only listens to people who agree with them becomes vulnerable to blind spots.

Whether we’re training an AI system, leading a team, building a company, or simply trying to keep our own thinking sharp, we need to regularly seek information from outside our usual sources. Because once ideas start marrying their relatives,... things can get a little jaw-dropping :-)



Helga & Paul Smith

—

SYNONYMS

AI inbreeding, AI self-training degradation, autophagic/coprophagic AI, data inbreeding/incest, feedback-loop AI, generative model decay, HABSBURG AI, inbred model, model autophagy disorder (MAD), model collapse, recursive AI training failure, recursive degradation, self-consuming AI, synthetic-data contamination, synthetic-data collapse/degradation


—

SMUGGLE OWAD into a conversation today, start with something like:



“If a future AI learns mainly from AI-generated content, and humans increasingly learn from AI-generated summaries, are we creating a HABSBURG AI ? … where nobody remembers what the original view looked like!”

—



THANKS to Madeleine for suggesting today’s OWAD.

-----------------------
PLEASE SUPPORT US

On evenings and weekends, I research and write your daily OWAD newsletter together with Helga (my lovely wife and business partner) and our insightful daughter Jennifer. It remains FREE, AD-FREE, and ALIVE thanks to voluntary donations from appreciative readers.

If you aren't already, please consider supporting us — even a small donation, equivalent to just 1-cup-of-coffee a month, would help us in covering expenses for mailing, site-hosting, maintenance, and service.

Just head over to DonorBox:
https://donorbox.org/owad-q4-2023-5

or

Bank transfer:
Paul Smith
IBAN: DE75 7316 0000 0002 5477 40

Important: please state as 'Verwendungszweck': "OWAD donation" and the email address used to subscribe to OWAD.

Thanks so much,

Paul Smith
(OWAD Founder)