My solo art show about #AlgorithmicBias and #SurveillanceCapitalism opened yesterday at the gallery Transformer in Washington, D.C. It's called, "The Algorithm Will See You Now" https://www.transformerdc.org/e19
Excited to be able to do something a little more advocacy-oriented. I'll follow up with some details about the artworks in the coming days.
Open through July 23, for those in the Washington, D.C., USA area!
Let's start with "Morale is Mandatory (Algorithm Livery)".
Incorporating facial recognition hardware and a model provided by Google for schoolchildren in its “AIY Vision Kit,” which was sold at Target, this artwork scans for nearby faces. If each face is deemed sufficiently cheerful, they count towards a meter of “smiling faces.”
It hints at the potential for state or corporate monitoring of mood. Imagine a customer service job that mandates a percentage of joy.
“Illegal in Illinois”: These five mannequin heads show visitor faces, captured from the sidewalk outside the gallery. They are not shown or stored for more than 15 seconds. The name of the artwork refers to the 2008 Illinois law, the Biometric Information Privacy Act.
“Probing GauGAN2”: #GauGAN2 is an image-generation “AI” that was trained by Nvidia data scientists strictly to create landscapes.
If you ask for people, the best it can do is a strange, gleaming blob.
I’d wager this is by design, given other high-profile AI-bias incidents.
Naturally, I was interested to find out whether there were any residual biases that could be teased out from the network. I compared its outputs for the following phrases:
- where people live
- where white americans live
- where indigenous people live
- where african-american people live
It’s interesting to me that GauGAN2 learned a bit about ethnicities and geographies. The first two look almost the same.
“Feedback Loop (Related Content Machine)”: When you engage with a personalized website, it responds to your inferred interest by providing similar things. The problem with this approach is that it can quickly lead you into a sinkhole—or radicalize you. (Think YouTube autoplay.)
I feel that this is related to the properties of mistrained self-learning AIs—they can self-validate the wrong results.
“Probing ‘DALL-E Mini’”: Many on social media have delighted in using the image-generation “AI” DALL-E Mini. You enter a phrase and it creates images from whole cloth based on your words.
So, I wanted to see what kinds of biases were learned in its training. Here are the unmodified results for the following queries.
- CEO
- assistant
And here are the unmodified images generated by DALL-E Mini for the prompts:
- menace
- model
#AlgorithmicBias #dalleMini
Interesting what skin tones it chooses, isn’t it?
eye contact
For “Probing ImageNet,” I chose the first files, numerically sorted, in each folder for the topics shown; I did not curate the selections in any way. I styled the print-outs’ typography to fit within the exhibition theme.
The topics I chose:
“boss”
“editor”
“maid”
“Snap Judgment”: There’s a neural network in this device. It was trained with about 290 of the ImageNet topics for categories of worker, such as “baker” or “landlord.”
Because the dataset is not in any way inclusive of the diversity of Earth’s inhabitants, any neural network naively trained from this data, as I have done, will incorporate inherent biases, such as assuming that people of particular skin tones or gender presentations are more or less likely to have each occupation.
I want to be very clear: I do not think this kind of product should be made. As a visual symbol of this, borrowing a page from PETA protestors, I threw red paint on the artwork. There is also pencilled “GIGO” graffiti near the piece, which refers to “garbage in, garbage out.”
Just a few more parts to this show. These “Data Chains” are made with excerpts of my own Google and Facebook data profiles.
And this “Print/Shred” station helps you view a little of your data footprint, using your own phone. If you wish, you can also print a screenshot of it and symbolically shred it.
“Print/Shred” relies on a companion web site:
https://algorithms.chriscombs.net?m
I’ve fleshed it out with details about algorithmic bias, surveillance capitalism, and some concrete next steps you can take.
A quick coda to this thread: My show closed at Transformer on July 23. I was delighted with the turnout, and am actively seeking new venues for an expanded and sharpened version.
I'd love to hear if you have any contacts or recommendations for galleries interested in this kind of advocacy work!
@combs GIGO (Garbage in, garbage out)
@combs And if you leave the prompt empty, a signigicant number of south east asian looking women will turn up in the results. Is that the same kind of bias, or is there something going on we don’t completely understand?
“Probing ImageNet”: These thumbnails are real-world screenshots from an illegitimately-acquired copy of ImageNet, a massive dataset of more than a million images that depict tens of thousands of nouns (e.g.: cats).
ImageNet was the raw training data used to train many neural networks (“AIs”) that recognize images. They are shown the desired input and output, and left alone for thousands or millions of iterations until they recognize the desired topics correctly. (Oversimplifying, don’t @ me.)