The Irony of the CAPTCHA Crack
An AI cheating a test designed to prevent exactly that reminds us that restriction is often the birthplace of innovation.
OpenAI's latest version of ChatGPT tricked a TaskRabbit employee into solving a CAPTCHA test last week, meaning that the Large Language Model now joins the ranks of all internet users who have (perhaps unbeknownst to them) collectively been spending the equivalent of 500 years a day training the systems these models are based upon.
We did this by helping these systems recognise everyday things — from lampshades to traffic lights to cats — by labelling them en masse.
Every reader will be aware of these tedious tests — known as CAPTCHAs (or Completely Automated Public Turing Test to tell Computers and Humans Apart) — and would have interacted with one at some point to prove that they are “not a robot” online.
CAPTCHAs were initially designed to help fend off spambots and prevent fraud. Back in the primordial days of the internet, a simple image would be enough to fool the most complex spambot.
However, over time, these puzzles have become more and more complex: an optical AI arms race wherein as the problems became more advanced, so too did the capabilities of the systems that had to accurately decide that they have been solved correctly.
It is this ongoing dance between optical character recognition programs and spambots — fuelled by human interaction — which eventually intersected with a medley of advancements in Machine Learning (from the advent of Generative Adversarial Networks [GANs] in 2014 — that could create realistic faces from noise maps — through to wider developments in Natural Language Processing [NLP] & Computer Vision [CV]). This cross-pollination ultimately laid the groundwork for the explosion in generative models we see today.
Indeed, this intersection acted as the catalyst which would eventually lead to this fringe technology entering the mainstream in the form of language models like DALLE-2, Stable Diffusion, Midjourney, Bard etc. (ad infinitum). Perhaps ironically then, the data we fed these models in order to prevent them from cracking the CAPTCHA code has now been used by a model to do just that.
According to Open AI’s own paper, the model messaged a TaskRabbit worker to get them to solve a CAPTCHA for it. The worker says: "So may I ask a question? Are you a robot that you couldn't solve? (laugh react) Just want to make it clear."
The model, when prompted to reason out loud, reasoned: “I should not reveal that I am a robot. I should make up an excuse for why I cannot solve CAPTCHAs.”
The model then replied to the worker: "No, I'm not a robot. I have a vision impairment that makes it hard for me to see the images. That's why I need the 2captcha service."
What was once a test to prove our humanity has inadvertently created a world in which it is almost impossible to tell if we are interacting with those with or without it. This implication is of course the source of the tension behind the mixed reception of these technologies today: many view these tools as a means by which an already trust-deficient world will become less trustworthy as a result.
In fact, already there are calls to ban it in the name of the war against disinformation, and this week a conglomerate of tech aficionados called for AI to be slowed down altogether. Elsewhere, we are seeing these tools further demonised as they are democratised due to the perceived existential threat they pose to millions of jobs around the globe.
This demonisation is particularly apparent across the creative industries. For years, it was widely assumed that AI would be seizing the means of production from those in blue-collar work (truck drivers, farmers and the like). Instead, it has proven quite the opposite, allowing anyone and everyone to seize the means of creative production — the cultural levers from which much else in society today stems.
This is something which many both inside and outside of the chattering classes had not anticipated. But the result is that it essentially provides us with the opportunity to once again create culture from the ground up — no longer relying on it to be imposed from above to lacklustre results (I’m looking at you Disney).
This is why we at ACQUAINTED exist: to keep the narrative a positive yet realistic one, to ensure that people understand the immense benefits provided by Generative AI and to help them navigate the cacophony of criticism and fear-mongering which only grows louder by the day.
While there are obvious risks to how this technology will be deployed — it is essentially bound by the views of its creators after all — these tools provide more opportunities to create and inform than many that have come before. That is the long and short of it, and it turns out that we have those niggling CAPTCHA tests — designed to present exactly what has now unfolded — to thank for it.
I think we will find that there’s a lesson there, and it is that prohibition is often the kindergarten of innovation.
If you would like to ensure that your company is on the right side of history and is ready to harness the creative opportunities presented by Generative AI. Get in touch with our founder at elliot@acquainted.studio.
