Shocked and amazed by image generators like DALL-E 2, this artist built her own AI model

It's a collage of nearly 200 images, arranged on a metre-high canvas. But even from across the room, something's decidedly unreal about the piece. 

On first glance, it's a crooked grid of photos: pictures of spiky succulents, for the most part. But there's something about the electric hue of the colours — the neon halo that radiates from every square. 

Up close, there's no denying it: these aren't photographs. Each aloe and echeveria in the collage was "grown" with AI, and the imagery's sci-fi paperback vibe is an instant giveaway; it's got to be the work of Midjourney.

Deep Fake (New Plants first imagined by AI) is one of several AI-assisted works appearing in Sanaz Mazinani's current exhibition, An Impossible Perspective (on through Nov. 4  at Stephen Bulger Gallery in Toronto). Whether working in photography or installation, Mazinani has long been interested in how we read images — news photos, for example, that are widely duplicated and published, reinforcing a particular way of seeing. So it's hardly surprising that the rise of generative AI, and its ability to conjure seemingly infinite visual content, was of enormous intrigue to the artist. 

Midjourney, Dall-E, Stable Diffusion: these text-to-image generators are barely two years old, and yet the world is already reading visual content in a whole new way, scanning for tell-tale signs of deep fakery: the aforementioned painterly glow of a Midjourney render, for example — or the presence of way too many fingers. 

We're at a turning point — "a generative turn," says Mazinani, citing AI scholar Kate Crawford. And as art-generating tools have become increasingly commonplace, Mazinani was left wondering how people will define creativity — not to mention truth and labour — when we're on the other side of this cultural shift. 

To get some answers, she embarked on an experiment, and her current exhibition reveals the results.

'Are these machines actually intelligent?'

Like anyone who's asked Dall-E to draw them a picture of Spider-Man eating spaghetti, Mazinani was floored by her early exposure to text-to-image generators. "I really wanted to understand it," says Mazinani, talking about AI. "Are these machines actually intelligent? Are they being creative?"

She decided the best way forward was to train an AI model herself. "When you look at something from the outside, it seems more impressive," she says. "But when you pull it apart and see how the sum of its parts are made, it becomes more realistic. I wanted to go through that process — and really understand the process."

Also, if she could learn to train her own AI, she wouldn't be left wondering how the data set might be skewed. "Other models that exist today have these biases in them," she says — racist bias, cultural bias — that reflect the data they've been trained on. 

How she trained an AI

Mazinani chose to focus her research on some relatively neutral subject matter: plants. And to train the AI, she collected a data set of more than 17,000 images: a mix of photographs (captured by Mazinani) and AI-generated "plants" (created via DALL-E 2 and Midjourney).

For guidance, Mazinani enlisted the help of Millan Singh Khurana, a computer programmer who's one of her art students at the University of Toronto. (Mazinani is an associate professor of studio art.) Together, they created custom code that would allow them to train a neural network. 

Mazinani was particularly interested in asking the AI to distinguish between real and "deepfake" images. "I wanted to really tackle what truth, or truthfulness, within an image means today … in this moment where everything could be faked."

But an AI doesn't inherently know the difference between true and false or natural and artificial. As the author of the experiment, it was up to her to define terms such as "real" and "deepfake" — and then teach the AI to file every image under categories that were based on a range of prescribed identifying characteristics. A "real" image, for example, might appear at a higher resolution than a "deepfake" one.

'Without the human element, it would be nothing'

"The more I work with [AI], the more I realize it's quite literally a sorting mechanism — a really sophisticated, multilayered sorting mechanism," says Mazinani. At the start of her experiment, she had a much different hypothesis. "I truly thought that this software could have creativity, and then by the end of it I realized that it just looks that way." Without transparent access to the extensive databases that tools such as DALL-E and Midjourney are trained on, there's a mysterious "black box" element to all of these technologies.

"Without the human element, it would be nothing," says Mazinani, and the work appearing now at Stephen Bulger Gallery highlights that fact. 

With the exception of one piece (All the Flowers), AI guided the composition of every work in the show. The elements of each collage are ordered based on a request; for example, Mazinani would ask the AI to rank the images from "most to least deepfake." 

Taking that information, Mazinani cut prints of each image by hand — snipping them into shapes that mimic the forms of a clothing pattern. And as another nod to tedious handiwork, each of the shapes has been fixed to its canvas with dressmaker's pins. "I really wanted the work to be about labour, because none of this would be possible without my labour of hundreds of hours," says Mazinani.

"I wanted this [project] to be a document of this specific time and place that we're in. I think this is a really interesting cusp," says Mazinani. 

So far, her experience with AI has left her with an optimistic attitude about the future. "The power is still ours, as humans, to decide how we want to use the software as a tool," says Mazinani.

"I think it's quite an amazing tool. Most photographers know Photoshop. They use it; that's a tool. But we never give credit to Photoshop for making the work," she says. "I think it's the same thing with AI and art creation. You can use it as a tool, as a way of working, but in the end, in order for the work to have meaning and merit, the lived experience of a human … needs to be part of it."

"I don't have another AI project on the go, but it is something that I'll be thinking through further," says Mazinani. "What I'm really interested in is the long-term impact of the technology, because it's going to get better slowly."

Sanaz Mazinani. An Impossible Perspective. To Nov. 4 at Stephen Bulger Gallery, Toronto. www.bulgergallery.com