How IBM’s Kush Varshney became the face of the modern ‘camera man’

The ‘camera man’ is one of those iconic photos for engineers working in image processing. A man in an overcoat stands in a field, peering through the viewfinder of his camera. For decades, engineers used the black and white photo to test and benchmark their algorithms in Matlab, a programming platform for scientific and engineering applications.

Kush Varshney encountered the camera man repeatedly as a grad student at MIT and even ran algorithms on the photo. Then, one day, it hit him. That imposing building in the distance? That was MIT’s Great Dome. “I had an epiphany while reading a paper at my desk in [MIT’s] Stata Center,” says Varshney, now an IBM Fellow and researcher at IBM. “I know where this is!”

Kush and his twin brother, Lav, decided to try and recreate the photo They trekked to the MIT soccer fields, found the approximate location of the original shoot, and as Kush leaned over a borrowed tripod, Lav snapped the picture.

A decade later, the image resurfaced after a team affiliated with scikit-learn, the open-source machine learning library for Python, were researching rights to the original camera man. They wanted to port the photo to scikit-learn’s image-processing library, scikit-image, but after failing to track down its author, they went looking on GitHub for a replacement.

There, they learned of the Varshney remake, reached out, and were given permission. Kush later announced on Twitter:

Today, the photo exists in various states, segmented by its grayscale intensity, weakly defined boundaries, and visible contours (called morphological snakes).

We recently caught up with Kush, who leads human-centered trustworthy AI research at IBM, to talk about the image and the state of authorship in the generative AI era.

This interview has been edited for length and clarity.

What were you studying at MIT?

I did my master’s and PhD in electrical engineering and computer science. My master’s thesis was on sparse signal representation methods for radar imaging. I liked the applied math side of electrical engineering, especially signal and image processing as ways to extract information from raw data.

Can you set the scene of the shoot?

It was spring, and the snow had just melted. I shaved, put on a dress shirt, pants, and an overcoat, and met Lav at the soccer fields on Vassar Street. I brought my camcorder, and a tripod borrowed from a roommate in Tang Hall. Lav brought the gloves. I ended up using the photo he took that day as the profile picture on my personal website.

How did your photo come to replace the original Matlab camera man?

When the developers of scikit-image couldn’t find a license for the camera man, they started a discussion on GitHub. A friend, Shubhendu Trivedi, noticed and remembered the image on my website. Lav granted permission, and the picture became one of scikit-image’s test photos.

Has anyone identified the original camera man?

Not that I know of. But Wired followed up on the famous “Lenna” test photo that led to the development of the JPG algorithm for image compression and triggered a long conversation on the objectification of women in tech.

Smartphones and the internet have changed how images are taken and shared. Why has this one endured?

It became a benchmark for image processing much as HELM and MMLU are now benchmarks for evaluating modern chatbots. Benchmarks are a way to test and compare state-of-the-art technology, but they often miss important context. AI models should be evaluated in environments that more closely resemble the places where they’ll be deployed. We need fewer “camera men” and more real-world testing.

Creative Commons licensing emerged in the late 1990s as the internet took off. Did it change how we think about copyright?

Creative Commons gave creators more options to decide how their work could be used. It promoted more openness, and it made available vast amounts of data to later train powerful AI models. We take copyright seriously at IBM Research and only train our models on data we have permission to use. But standards vary and are in flux. At a recent AI governance conference I attended, Harvard law professor William Fisher noted that in the 17 countries he examined, each had a unique stance toward the use of copyrighted material for training AI.

Why are creative works copyrighted?

Copyright is ultimately a product of capitalism. If we didn’t treat works of art as commodities, there would be no need to copyright them. Lav and I never asked to be paid for our image because we have other jobs and only wanted to contribute to science. Homer and Vyasa were traditional bards who composed and performed epic stories thousands of years ago but never considered themselves authors. LLMs produce content that’s reminiscent of these oral narrative poets. We should support this ‘tradition’ just as previous societies supported singers of tales by buying them a meal or gathering a small collection. By this, I mean paying a little bit for the inference computation. Then we can leave truly original work to human creators.

Is something like a Creative Commons license for AI-generated content needed?

We need a way to mark AI-generated content so we can recognize what’s original. Some of my team members recently came up with a self-reported AI attribution mark that can be placed at the end of a text. Let’s see if it catches on.