When AI makes art, what can go wrong?

How does AI manage to create art?

In 2014, a research scientist at Alphabet-owned DeepMind created a new form of algorithms, called Generative Adversarial Networks (GANs). GANs use two neural networks, called “generators” and “discriminators” which are ‘adversarial’ to each other. While the former creates fake images, the latter catches them. An output is generated when the generator beats the discriminator. Eight years since the first GAN, researchers and AI firms have made these algorithms available to the general public, allowing anyone to create art with them, in return increasing the training data that they get.

Are GANs used solely to create art?

No. GANs can be used to generate video, text and even audio. For instance, last Diwali, Mondelez International, the owner of the Cadbury brand, created an advertisement called ‘Not just a Cadbury ad’. In the ad, actor Shah Rukh Khan is seen promoting neighbourhood stores in India, except that Khan never actually spelt out the store names. To do this, Mondelez used a GANs program from startup Rephrase.ai, to create the videos. GANs have also been used to create ‘artificial humans’, which can replace actors in movies, and be used as personal assistants, receptionists for offices, and more.

Is it possible for someone to misuse GANs?

The most infamous misuse of GANs is in creating ‘deepfakes’ — realistic fake images and videos of personalities. An example was the demo AI deepfake of ex-US president Barack Obama abusing Donald Trump. Patrick Hillman, CCO of Binance, said scammers used his video footage from news interviews to create AI-generated fakes, and scam crypto users.

What are the options consumers can use?

While the highest quality tools are paid and available only to a limited few, there are plenty of free options that anyone can explore. For example, DeepAI, a US-based AI startup, offers both image and text generation to users for free. Others include Hotpot AI, Pixray and Midjourney. The last one is likely the most popular right now, and allows users to create images through its Discord group. It has also been likened to Dall-E, a limited-access algorithm created by Elon Musk-backed AI research firm OpenAI.

How does one get to use these systems?

For most of these tools, it is as simple as opening your email or replying to a tweet. DeepAI, Hotpot and Pixray have websites with text boxes where you can type what you please, and wait for the image to be generated. Pixray has advanced features, where if you are a developer and have built an AI model using Python, a paid service lets you try your AI model to create an image. Midjourney creates four images on a text prompt, from which users can select one and upgrade to high resolution.

First article