How I get started on AI Privacy and Security
Being scammed, hacked and unable to verify IDs.
Act 1: Lucks
I always feel lucky as a frontier AI researcher.
Fifteen year ago. I was an early user of social media.
And even more as a graphical designer to do cool “starry” art. I freelanced in several projects. Mostly end up on television, stage, public spaces, and social media. Three years made me become a manager, eventually managed managers where I could see an end-to-end process from the inside. It was strange with my science and pure math background, but experience did not count education quite well.
Social media is quite opposite to science. People do not count from 0 to 1, again and again as if I do projects in academic contexts. They count from big numbers first as default numbers. Most of engagement is completely made up from a team of 5-7 people, and even an entire company have to do the social job to support a team with low engagement. Likes, shares, comments. Or simply they just buy ads.
I thought it was normal. Being proud of content, my main responsibility. The only problem was a competition between us and several social media outlets trying to engage with the same audience. However, this was my naive belief. The CEO of the company wanna make money as fast as possible with dark patterns to trick people’s behavior and affirm pseudo-belief. I resigned and returned to academia for a PhD in information system. Accidentally jumped on privacy and security that filled with all favorite math. Later, found an unknown article on open archive website with a title Generative Adversarial Networks. And then I had a flash of the future with my past as a graphical design, content creator, and video maker with … some fancy graduate math.
I described that vision to grad student telling me that it was similar to very well-known research as a death end of all. The reputation was so bad that almost nobody could not publish any paper, have grants or recruit students. In 2019, it was still ridiculously insane. Mid 2022, some first early successful experimentation shows super interesting result on the internet. Late 2022, I became an early user of a new research project that have a mega viral hit with 100 millions of users after two months. The rest is a history of an entire human kind. And that insane number keep rising non stop. More and lots of more GPUs. Finally, the world agree with us.
Act 2: AI4Good
During years working on social media, I used to work on a personality model that we called Big Five. By assessing the language model built on top of it hundred times, I noticed that there were some highly correlated words hitting people perception more than the other word. I proposed a language model on top to analyze data that we collect from users to create an effective viral social media campaign, but then left without completion for grad school. Late 2016, Donald Trump become the president of the US. My instinct quickly points to that research I did early starting from a psychometric center of Cambridge university. In 2018, an investigation exposed a company called Cambridge Analytica. They extracted data users from a social media to create a campaign and make influence on the US presidential election.
To understand how language form our perception based on experience. Let’s use AI as an example. During my first five years of research, if you talk to any scientist with experiences, they will give you a long lecture about “AI winter” with bitterness, despair, and isolation. For the audience without any knowledge or shallow experience, they mostly refer AI with questions based on some Hollywood sci-fi movies about "doom days” or “Will AI kill us?” or “Will AI take jobs from us?” or just a pitch to make you look fancy by floating hot airs in an entrepreneurship contest. Everywhere from my home country to South East Asia, Canada, US, Europe. Nothing more than just a marketing phrase to sell ideas, products, and services based on “fantasies”.
Early 2019, with a scrappy proof of concepts, and a green signal from the first demo made by Nvidia, I confirmed the future definitely coming in the next 5-10 years. I started to build a company working on GenAI with a project called Smart Brush right in San Francisco. And wrote a book about Augmented Intelligence with a vision incorporated my past experience from IBM Watson. But then crisis, lots of crisis happened. That forced me to pose some fundamental questions about human vulnerabilities and how to have healthy mechanism to cope with our negative self. Honestly, this is a topic of religion, faith and social science more than natural science and engineering. I eventually shifted healthcare as the only direction of exploration from creativity and design during a trip in Switzerland.
The pandemic happened. My home country was one of the top ten right next to China. News spread super fast even to a rural village days before the biggest celebration of a year with dozen of millions migration around the country. My healthcare project was detecting early cancer with hope to prevent, protect and mitigate risks. If we could make it happen, we could save up to 30% of world population and a promote a better healthy lifestyle that prioritize sustainable environment. I estimated that it would take 10 years for this project to hit a final goal. However, a new unknown disease appeared. My country suspended travels, canceled almost all public activities, shut down borders, practiced social distancing, and several months of severe lock down.
Physically we were safe, but mentally our health were not tremendously. I was alive but barely feel happy with constant burn out and exhaustion. Days passed passively without any experiences as if you forever walk on a dessert. Nothing seemed matters. At this moment, I was able to enter the level of faith and found Gratitude. It also mean nothing matters, but here is a state indicating an existing of a meta class. Similar to zero as an integer number or natural numbers. It is still a number similar to others. I found resonate with a teaching of Buddha, i.e. Heart sutra, and my pure math education with a definition of topological spaces. Now you see that human language and perception are so tricky as if scientist first learn about quantum theory.
When you call something by a name, despite of how it shows up, it’s already exist inside your head regardless of how true it is in the physical form. Example: Don’t think of the elephant =) Bet you don’t think of the elephant when it is called out in the sentence. More interestingly, even it’s just a concept or an idea, the mind will try to find an evidence to confirm its existence in the shared world as long as it believe in it. This is an affirmative system, our human nature. The more you believe in something, the more you will find those in your lifetime even if it does not have any direct inference or statistical correlation at all. Human tend to repeat ourselves and group believes in social contexts, so we have cult(ure) and religion.
Act 3: Education
Switzerland is where people will yell at you about privacy. The only country where people aggressively explicitly demand for respecting the culture if you violate. I would compare this to how well the Swiss train is on time and their tax everything system. On my first day in country, I had to pay for a fine with a super long bill just because of buying a 1-day train ticket on the internet. Luckily they refund me later, but an unforgettable memoir.
Five years in, I made some early social proof to promote social well-being by helping people re-frame “Nothing matters”. We reach 12M people world wide and up to 1M per post without any ads. Started from 0 to 1 every single day that I hand-pick a prompt to help motivate people and practice gratitude. Looking at this there were 1000+++ zero to one, then turn into a different numbers. This story was much matter to me than searching for GPUs to train a fancy GenAI model.
But then, I lost it. As a frontier Gen AI researcher, data scientist and former social media person, I was always confident to by-pass an algorithm that constantly selling me drama shit and bot of scammers, spammers, phishers, impersonators, or just content farmers. It was nearly decades I have left social media for academia since 2015. However, I still fall victim to a scam. Additionally, the first time ever, I could not prove that I am an actual owner of a backup account.
In the past two years, I constantly see an update of new human verification on all websites or a browser every time I use the service. I took 30 basic tests to prove that I am a human on X before requesting for a support. After an SEC compromised account one day before the Bitcoin ETF, they even enhanced user management via the login process. I was one of several not able to log-in, but lucky to get it back after a few days. Similar things happened to iCloud and many other services. But for this time, despite of showing up all kind of legal documents and send in-person request to a Security team, I could not get an access back.
After that, I move around to examine situation in a broader context. Scam is everywhere. There was a day ten out of ten people I contact are all scammers. A “victim” user story were ridiculous for job seekers, tax payers, and all kind of sh!t that you could name that ask you for “an investment” before getting paid or refunded, or a scam of scam via social posts or social ads telling you they know how to get your money or account back. The worst part of this is those ads following my digital footprint to all of other social apps and e-commerce platforms w.t.r the local ads and content database. Similar incidents happened to me earlier during trips around the world, but it was still human made content despite of how outdated they were.
No doubt that GenAI has been used for social engineering to hallucinate more scenarios that actively hunt victims without borders. Indeed, it is a wide spectrum, spam, scam, phish, gaslight, abuse, manipulation. Looks simple. No fancy tech and code, just words and images, but super effective. Indeed, last year, prompt injection ranks top 1 in the a list of top 10 vulnerabilities published by OWASP.
After 30+ days of being scammed and hacked, I complete a 6-month-Google-Course on Cybersecurity, and working on another course of IBM. It’s not the end of education, but the beginning of something BIG coming. If 2023 was year 1 of Gen AI explosion worldwide, year 2 would be a massive online scam and spam across borders. Don’t know how much is enough for this, or it will be never enough for an exponential growth of Gen AI hallucinations. Not sure if technologies or our government regulation could save us from bad actors. I pray for our good faith but definitely it will be better if you prepare for yourself first.
Honestly, I constantly see loopholes and dark patterns that could be exploited by bad actors. Not hard core engineering, but just pure human language text and photo on social media or any platform. They collect your data “with your consent” than selling it to a broker to fuel their recommendation system or any third party to personalize “strong emotionally revoking” message sent to you. A pure form of manipulation.
This has been highlighted in EU AI act in the prohibited part for consumers and individuals, however it does not apply for a broader context outside EU where laws are too weak to protect users. The first UN AI resolution touch on level of nations, mostly about military use and warfare. Personally I am looking for social well-being and data privacy/ protection that regulate activities of the data collector and brokers industry without our consent.
We might only (a hash name) row in their database. Nevertheless, the negative impact are real and significant to our daily life quality, so the fabric of human relationships and our society. For all of these reasons, I would like to invite you all in my journey of life long learning for AI security and privacy through this substack.
