Loading video…
All breakouts
13.0Talking HeadWarning hookAI Tools
Hook
You're not going to like this one.
@sarainwondertech
12.5K followers · usually 1.4K views a post
Views
14.4K
Likes
579
Comments
26
Shares
134
Why it worked
The video leverages a high-stakes, fear-based hook regarding personal privacy that is immediately relevant to almost every internet user, combined with clear, simplified data visualizations of a complex research paper.
Summary
The creator discusses a research paper from ETH Zurich demonstrating how AI models like Gemini and ChatGPT can deanonymize pseudonymous online accounts. The video explains the methodology used by researchers to link anonymous profiles to real-world identities and warns about the implications for online privacy.
Structure
- 1Hook about AI deanonymization
- 2Explanation of the research methodology
- 3Results of the two experiments
- 4Implications for online privacy
- 5Call to action to share the video
Call to action
share this video
On-screen text
You're not going to like this one.
Plot twist: can match your anonymous online accounts to your LinkedIn
$3
FAKE OR ANONYMOUS ACCOUNTS ARE NOT GOING TO HELP YOU
67% identified
1000 LINKEDIN PROFILES
EXPERIMENT 2 - LinkedIn to 89,000 HN accounts
45% matched correctly
99% precision at top accuracy
48% OF THE TIME
SAME PERSON - 94% confidence
NO PROTECTION EXCEPT DON'T POST ONLINE
WHAT CAN GO WRONG, RIGHT?
I WILL FIND YOU
Original caption
Researchers from ETH Zurich just published a paper showing that AI can deanonymize pseudonymous online accounts at scale. Your online privacy is officially gone. Using a pipeline combining Gemini embeddings and GPT-5.2 reasoning, they tested whether AI could match anonymous profiles back to real identities across three experiments. In the first, they took 338 Hacker News accounts, stripped all identifying information, and let an AI agent search the open internet. It correctly identified 67% of users — with 90% precision. That means when it guessed, it was right 9 times out of 10. In the second, they took 987 real LinkedIn profiles and asked AI to find their anonymous Hacker News accounts hidden inside a pool of 89,000 users. It matched the right person 45% of the time at 99% precision. In the third, they split a single Reddit user's comment history across time — one year gap — and asked AI to recognize the same person in both halves. It succeeded 38% of the time. The cost of running this attack: $1 to $4 per person. The data needed: nothing but your public posts. According to the authors, there is no meaningful protection except not posting online at all — because the very content that makes online communities valuable is the same content that makes you identifiable. This is not a niche security finding. It affects anyone who has ever posted online under a username. Paper: "Large-scale online deanonymization with LLMs" — Lermen, Paleka, Swanson, Aerni, Carlini, Tramèr. ETH Zurich & Anthropic, February 2026. #airesearch #aiprivacy #greenscreen
More from @sarainwondertech
More breakout videos from this creator.
