OpenAI recognises this, but keeping an AI on the straight and narrow is difficult. Machines have no sense of what’s wrong, and to stop its ChatGPT from becoming a horror show, OpenAI outsourced the clean-up to Kenya.
A Time investigation showed the workers who sifted through and filtered out text-nasties in ChatGPT were paid less than US$2 an hour.
Facebook used human moderators in a similar capacity, a nightmare job if there ever was one. Ironically, ChatGPT labelled the Kenyan outsourcing as economic injustice and a legacy of colonial violence.
Training data is all important for AI and paradoxically, it needs to be human-created because the tech is meant to mimic people. In his Four Battlegrounds book on AI, author Paul Scharre recounts how the United Marine Corps tested a robot with an algorithm which was able to detect people.
The human-detection algorithm proved easy to defeat for the Marines who used tricks like somersaulting towards the robot, hiding in a cardboard box and masquerading as a fir tree. None of that was covered by the training data.
“These simple tricks, which a human would have easily seen through, were sufficient to break the algorithm,” as Scharre wrote.
In other words, without human input, AI is neither here nor there. Getty Images is taking legal action against Stability AI, the creators of the Stable Diffusion text-to-images, which is similar to OpenAI’s DALL-E 2, for intellectual property rights infringements.
Stability AI didn’t obtain a licence for the millions of Getty pictures it used to train the machine that creates images from text input, allegedly.
That raises the question if AI really is “generative” as claimed or simply a computer tool that produces amusing (and weirdly distorted) facsimiles of human-created images, video and writing?
One area where it doesn’t matter at all if what AI creates is outlandish and not anchored in reality is disinformation and propaganda posted on social media sites.
The experience of past years has shown that you can claim just about anything, like mRNA vaccines against COVID-19 cause people to shed viruses and even worse nonsense.
Some people are guaranteed to believe the strangest of things. Now, take a tireless AI trained on reams of rubbish posts, and combine it with the huge reach social media has with billions of users and see what it’ll create.
That’s problematic to say the least, and there’s no guarantee a “counter AI” will be able to work out what’s machine and what’s human-generated and filter out the former. OpenAI’s manual filtering in Kenya suggests it’s not possible.
What does the future hold for AI? That’s difficult to say, but it’s almost as if there’s an immutable law which states that good ideas will invariably become corrupted at Internet scale.
While Google started as a way to better index web pages based on their relevance, it has evolved into a privacy menace that undermines democracy as it consumes the advertising money that once paid for diverse media.
Cryptocurrencies were supposed to bring financial freedom to people with computers, but turned out to be giant pyramid schemes that fuelled the ransomware industry.
Even smaller things, like using email addresses and passwords for account logins, has become an unmanageable mess that is now an effective attack vector for cyber criminals.
We’ve learnt all that over the years, and signs are that AI will be no different as it grows and reaches more areas of society. To not repeat the mistakes of the past would seem to be the intelligent thing to do here - but it’s probably too late already.
