Tuesday, July 25, 2017


There seems to be a little spurt of interest around automated photography, of using Artificial Intelligence (AI) to replace the human here, in one way or another. All part of the current trend of Rich Idiots in Silicon Valley being interested in AI, I assume.

A little background, speaking as a recovering technologist.

AI is a blanket term for a set of technologies that allow computers to simulate, with varying degrees of crudity, things humans do. Recognize spoken speech, recognize objects in pictures, play chess, solve logic puzzles, and so on. Despite the name, none of these things are "intelligence" in anything like human terms. These are all things I would characterize as remarkably simple, even stupid, algorithms (pieces of software) that produce startlingly, unreasonably, sophisticated results. In this context, "remarkably simple" doesn't mean "simple", it means "a lot less mind-bogglingly complicated than one might think."

Important notes in no particular order: There's no "person" in there in any meaningful sense, and in fact nobody knows how to even get started making a software "person." This does not appear to be a current area of study. The software that turns your spoken words into text has no relationship with the software that plays chess at a Grand Master level. The phrase "neural network" does not mean that the software resembles a brain in any meaningful way, the phrase just means another remarkably simple algorithm (inspired by real neurons) which can produce useful results.

Onwards. Colberg pointed out on twitter this project: Computed Curation. Philppe Schmitt (whoever that is) has bolted together some of the aforementioned AI technologies to make a piece of software that can sift through some photos and sequence up a photobook.

It does a tolerable, if pedestrian, job. The connections appear to be both simplistic and entirely linear. The only points of real interest are places where I suspect the computer has mis-read a photograph. We should also keep in mind that this might be entirely a scam.

Anyways, let us suppose that the software works as advertised. So what?

The point of a creative endeavor is that it is the output of a person, it is the result of intent driven by a lifetime of experience, experience remembered by the unreliable mechanisms of the brain. Without a person, there can be no intent. Without intent, there is arguably no creative output.

Schmitt's book is sequenced, as I noted, but there's no intent. It's obvious in a moment that this is a random string of pictures connected together with superficial visual similarities. The important point here, though, is that this is not a problem that can be fixed. This isn't simply an early version of the software, this is basic to what the software does. There's no place for an "intent" to be generated by the thing, and to get it right you first have to make a software "person" to have the intent.

This is not to say that some clever Johnny won't take a crack at simulating intent with, no doubt, a remarkably simple piece of software. This pseudo-intent, while potentially interesting, isn't actual intent. Remember, there is no ghost in the machine, there's no person in there in any meaningful sense. If it's good enough, I suppose it would say something or other about the nature of Art. If I can establish a genuine-feeling connection with a simulation, then what? It doesn't matter much, I am not much interested in yet another way to Trick The Brain.

An alternative is to provide a way for an actual person to encode an intention, and to have the software produce results based on that intention combined with the usual collection from the AI parts bin. In that case, what we have is another tool for artists to use. Maybe it's interesting, maybe it's not, but at the end of the day it's not much different from premixed oil paints.

Art is distinguished from much of the rest of human endeavor by lacking a well defined endpoint. I can define exactly what it means to win a game of chess, I can measure accurately how precise a speech recognition algorithm is. I cannot define precisely when a piece of Art works, I cannot define precisely when it is Good. Without a well-defined goal, software inherently has problems, and is left to thresh around a bit wildly. Art's result, Art's endpoint, is itself. The goal is in the thing itself, in the process by which it was made, in the human who made it, an an inextricable ball of entangled relationships. This presents a real problem for software.

Our friend Lewis Bush, writing his usual maze of misplaced apostrophes and abused prepositions (or is "A is different to B" actual British usage?) for World Press Photo over here is on the trail of something more interesting.

Journalism isn't Art, and there's no particular reason that algorithms can't do it just fine that I can see. Lewis is just throwing a bunch of stuff against the wall, a random selection of technological works in progress (that are mostly going nowhere) that he's stumbled over in his reading. Lewis, as usual, is fetishizing technology without much understanding of it.

There is real trouble here, though, which as far as I can tell, he's missed completely. Journalism will inevitably shift to accommodate automation. Lewis needs to recall his Sontag: we don't photograph the important things, we make some things important by photographing them.

Automated journalism will reshape the news. If the algorithms can't be taught to reliably make national elections an Important Subject Of News (to pick an example at random) then national elections will cease to be important news. We see this constantly in all walks of life now, the flexible human bends to accommodate the inflexible and stupid computer.

I don't know if we're there yet, but if we're not, it's very very close: Facebook's robots will select for us which pictures are interesting, and those will be the ones we see. A combination of algorithms and crowd-sourcing will select for us what to look at. It is literally inevitable.

News and journalism do suffer from the lack of endpoint problem that Art has, the difference is that it doesn't matter (in the current zeitgeist, anyways). Anything will do for news, now, as long as it generates the clicks. So, in fact, it has always been. News was whatever got the troubadour paid, news was whatever shifted newspapers, news was whatever got you to watch the TV, and now news is whatever gets the clicks. Still, there were humans in the loop, and ideas about what journalism "ought to be" which now and then got a little play. Algorithms will remove that entirely. The engineers will get it good enough to pull the clicks, and then stop, because that's what engineers do.

You can argue, I think, that Art currently also has to "nobody cares" problem, and that therefore software that does a shabby job of simulating artistic intent will do just fine. You might be right. Not sure the Art Market will put up with that for very long. They don't put up with anything for very long. The relentless pursuit of the novel will push the computer aside in due course, in favor of some new variant of the facile fast-talking young turk with a new story and a new My Sad Project project.

For journalism, perhaps there is space, suddenly, for a revival of Life magazine in some new, contemporary form. People are, thankfully, a cutely aware of the fact that Facebook is running their lives, that Facebook is skewing what they see in order to sell them stuff.

Prepare for some backlash! I don't know what it will look like! I hope it will be fun!


  1. Yes, UK English is different to US English.

  2. I can't decide if Facebook is playing some sort of double bluff with it's targeted ads - pretending that it really is hopelessly wild of the mark (wallless homes with yellow mattresses? Yes please) in order that I don't notice some subtle, under-the-counter processes that have me buying stuff I didn't even know existed, let alone wanted.

  3. If AI is going to select and sequence photos after the fact, then I wonder how long before it will be used to decide which photos are taken in the first place?

    Cameras already have autofocus, autoexposure, face recognition, and image stabilization built-in, so having them also arrange compositions certainly appears to be a logical next step.

    1. I think there's been a bunch of work there already, and it kind of works some of the time.

      There are applications for this, to be sure. Still, without a person there isn't any intention. There is at best simulated intention, and probably not even that for a good long time.

  4. AI developers are simply the most recent and virulent symptom of our species' inherent difficulty with distinguishing the "price" and the "value" of an activity, whether it be chopping wood, going to work, or taking photographs. By endlessly reducing the price (by pursuing "efficiency"), we also continually reduce the value (the point of being alive in the first place). The absurd end-point of AI is the extinction (most likely through terminal boredom) of messy, inefficient, pointless old humanity altogether...

    Over my working life I have witnessed several waves of good, ordinary people being rendered purposeless by clever technology. At some time soon it has to stop: people are the point, not the problem!


  5. For now, it's just baby steps, but AI is definitely coming to video and photography: https://www.youtube.com/watch?v=3r_Z5AYJJd4