It is redundant to point out how much progress AI applications made during the past few years. What is escaping the attention of many people, however, is that in most creative areas, there are already fully working consumer-grade tools based on generative AI that can produce output similar to that of a human with above-average capabilities, and there are many more on the way. These tools also happen to be rather affordable, making them accessible to a large amount of people.

The surprise is not that machines are able to perform such tasks but rather how fast we got here. After Google introduced its transformer deep learning model in 2017, the pace really picked up, with Natural Language Processing (NLP) projects becoming able to process data much faster, reducing training time and allowing for exponential scalability. It is hard to fathom that, in 2022 alone, there are already several competing initiatives for most creative activities.

Publishing a novel is a dream of sorts to many, but constraints such as time often make that difficult. However, a startup is already generating long-form creative writing, requiring just a few human prompts. Below, I generated the introduction of a novel about our international technology community, making use of the Euterpe AI. Sentences in bold are my prompts, while everything else is Euterpe’s work:

There was an international community who solved technology issues. They were called the ICTA, International Community for Technology Advancement. The goal of this group was to make all people in the world have equal access to information and communication technology, with a focus on those who were poor or disenfranchised.

Their efforts were carried out in different countries and united all sectors of society, such as NGOs, businesses, and governments. It was not only about giving away things; it also involved training people how to use new technologies.

The problem that the ICTA faced was the lack of funds to do what they did.

“I am sure you are aware that we are facing some challenges,” said the representative from the organization trying to help them amass funds. “We need more money to be able to provide services to everyone.”

“What kind of problems?” asked one of the senators from Brazil, whose country was one of the founders of the ICTA.

“For example, certain locations lack electricity, even.”

“We should just give everyone solar panels!”

“No, that would not be enough. Solar panels work fine in areas where there is ample sunlight, but what if the place has no sun for months at a time?”

While this sample novel might not be on its way to earning any prize (although a Japanese AI almost won one), it is hard to deny that these results are superior to those present in some published books written by humans. Euterpe’s ability to grasp the issues that the technology community faces is particularly impressive, given the lack of context provided. When faced with a new issue, such as that of the difficulty of providing electricity, it came up with both a solution and an impediment to its own solution.

Text generation has seen a fairly niche application so far, text-to-image generation has advanced at a break-neck pace, and in the span of 2 or 3 years, the technology went from producing impressive sunsets to being able to generate very complex scenes. The Midjourney AI has pulled ahead of the pack and, with every version, becomes better at its task.

Here is an image generated using version 4, with the prompt:

“imagine a multinational community reunited at the United Nations building in Geneva, a person is making a speech, there are exposed cables all around, in cinematic atmosphere, volumetric lighting”:

That image would have taken hours to be drawn by a professional artist, and yet the AI took around 30 seconds to generate it, plus another 30 seconds to upscale its final version. Midjourney promises this is only the beginning, with the generation of 3D models and scenes soon to enter testing. Other companies are also already exploring the generation and editing of entire videos, with shorter institutional ones already being quite viable.

Remember the DeepNude scandal from 2019? That might look like a walk in the park in comparison to what is on the way. While Midjourney itself frowns upon the generation of that type of content, several initiatives have forked the Stable Diffusion base and are already fairly advanced in terms of generating nude images more or less from scratch. None of these projects has achieved enough success to be disruptive… yet.

In other words, the debate around so-called deepfakes did not even reach the point in which much regulation was seen, but technology has already moved on towards a different set of conditions, in which often not even source video material will be necessary anymore. These will be purely synthetic productions. Considering how advanced text-to-speech has been for years, anything that currently exists in video form will soon be generated in a few clicks. There is certainly not a lack of sample videos on the Internet to perform the learning with.

An interesting reaction that seems to come up in discussions regarding these advancements in AI is that “I can tell it was a machine that generated this!” and that may well be true, but what this statement misses is exactly the understanding of how fast these models are evolving. It has come to a point in which even people who follow the field closely are losing track of its constant evolution.

It is hard to tell where this technology is going in the longer term. While some tasks are relatively simple for contemporary AI to perform, others are much more complex and still inaccessible. Artificial General Intelligence (parity with humans) seems to still be a few decades away, but when it arrives, it makes sense to believe we can expect an even more accelerated pace of development than the one we are currently observing.

It would be a fool’s errand to attempt to generate policy ahead of time in order to predict such scenarios, but we have already missed the boat in relation to the current generation of AIs. Debates around the IGF and other fora have been ongoing since last decade, but the global community never found much in the way of consensus, even over simpler matters, and this lack of direction has left the doors wide open to both innovation and potentially dangerous use of these technologies.

Asking ourselves what base limitations and principles we would like to see incorporated into these AI projects is not a subject for the future. It is a subject for today that is often being left on the sidelines by people other than the developers of the applications themselves (and a few governments). Maybe it is time we take a harder look into the situation before we are left behind once more.