AI and the Onslaught of Deep Fakes
By: Michael Frendo, PhD, and Emmy Sobieski, CFA
We have all received them, emails purportedly from someone we know, but not quite right….
Emails that contain awkward phrasing, strange use of consonants, or just seemingly strange wording. Somehow we know they are fishy, and that they may be “phishing”. Unfortunately, technology may be here to change all that. Breakthroughs in our digital lives can make us more productive and more secure, but also more vulnerable.
Like many things in life, breakthroughs arrive slowly, and then all of a sudden.
Humans beat computers in chess for years, and then suddenly compute power and software capabilities led Big Blue to beat chess champion Gary Kasparov in 1997. This breakthrough did not mean the end of humans playing chess. In fact, computers trained humans to be better at chess, accelerating their improvement. Chess masters can now come from anywhere, not just big cities where they could play other top players. Technology made the world flatter for chess.
Humans create. We express our creativity through innovation, art, and words.
Natural language processing, NLP, a sub-segment of machine learning, aims to categorize words and form prediction models, and perform analysis. Prediction can be used to automate tasks like writing a blog, writing thank you notes, coding based on your instructions (or predicting what you want), building a database, writing emails. Analysis can be used to provide context, to understand what someone means by their words.
OpenAI led by Dr. Kenneth Stanley, former head of AI at Uber, delivered a major breakthrough in NLP prediction with the release of GPT-3.
OpenAI built GPT-3 by studying patterns of nearly a trillion words and 175 billion parameters across the internet and literature. This chart compares the scale of GPT-3’s data to previous breakthroughs and to its own predecessor, GPT-2. Breakthroughs arrive slowly, then suddenly…
Having learned using hundreds of billions of words, GPT-3’s breakthrough is to predict words using very little context or data.
You provide a small sample of writing or code and GPT-3 extrapolates! GPT-3 authors explain: “Humans can generally perform a new language task from only a few examples or from simple instructions – something which current NLP systems still largely struggle to do.” This is what GPT-3 was able to do in language creation although it still shows bias and has more work to do in contextual analysis and interpretation.
According to its creators, GPT-3 is strong in translation, question-answer, and fill-in-the-blank, on-the-fly reasoning, domain adaptation, unscrambling words, using a novel word in a sentence, and performing 3-digit arithmetic, all with very little new data.
GPT-3 was “released” at the end of May 2020, and users have to apply to use it (by getting whitelisted).
The excitement is palpable, with early users showing examples of starting to code or run a sequel query, and the GPT-3 completed their work. Another user showed GPT-3 a few news snippets about himself, and GPT-3 authored his obituary.
Using a few samples and prediction, GPT-3 can write your tweets, blogs, create code, and sequel queries!
“GPT-3 can generate samples of news articles which human evaluators have difficulty distinguishing from articles written by humans.” – GPT-3 OpenAI GitHub ReadMe
Imagine the productivity gains…
Imagine the opportunities for identity theft…
What does it mean for identity when AI can create a picture of you somewhere you never were, and write blogs, stories, and code that others would mistake for your own?
Will phishing still look fishy?
In the world of cybersecurity, this brings new meaning to the challenges around phishing. You could receive an email from your boss (but not really from your boss!) that sounds exactly like her. This can be even more challenging if it is not someone you know or trust, but is part of a trusted relationship, for instance, someone at a brokerage, bank, credit card company or utility.
How can we know who someone is when deep fake pictures and words can be created by AI?
How do you future-proof your business relationships to build trusted bridges with customers rather than chase down cybercriminals? It all comes down to identity and the ability to verify it. What is it?
Is it what you know about yourself, some unique device you have, or is it your unique living breathing self? Is it your picture or your living iris?
How can we leverage our true identity to build trust in a customer relationship?
Journey's breakthroughs in protecting our identity and securing our digital relationships are critical as we experience more breakthroughs in AI that can pretend to be us. These breakthroughs can make our lives easier or enable others to pretend to be us. It is being able to detect deep fakes, detect liveness vs video.
Journey built an inherently private trusted network that can use verifiers dynamically, in real-time, to dig deeper into identity when necessary.
It is not about something you know or something you have, it is about something you are.