#bot ethics

Google Assistant

I was a bit surprised when the notification popped up to tell me that my new panorama was ready. I’d shot the photos for it, of course, but I hadn’t gotten around to putting them together. And then it followed it up with an animation, assembled from a burst of photos, and put a filter on yet more photos.

I’m not sure when my Google account started doing this. I think I might have turned Google Assistant on accidentally. It’s ability to detect what I intended to do with the photos is a bit unsettling. 

This kind of AI-assistant automation is becoming more prevalent, no doubt inching towards the Star Trek computer we always wanted. But, for me personally, this is the first time an AI I didn’t know has given me a present.

I’ve had bots give me presents before, of course: Appropriate Tributes sent a tweet to me the other day, for example. But this is like getting a surprise present from a stranger.

It’s not an entirely bad feeling: there is some surprise and delight at the results. But it is unpleasant to have a stranger show up with a present, and it feels socially off.

I assume that I technically did opt-in, or at least I think I remember enabling some new Google service without realizing that it did this. And I have no idea what else it’s doing: what if it decides that it doesn’t like my photos or that I don’t need to own a copy of my own files and deletes them?

Don’t laugh: this actually happened with iTunes.

So there are many ethical and social issues with AI, particularly when it starts directly managing our things and inserting itself into our lives without waiting for explicit permission.

Ethical imperatives in AI and generative art

Procedural generation is about creating things that aren’t under your complete control, and Liza Daly has put some thought into the ethical implications of that. Drawing from the examples of the bot community, this essay extends the guidelines to cover all kinds of generative media.

Ethics have come up here a few times before, though mostly in the context of bots. This stuff is important to think about if you have any intention of releasing your generative projects into the wild.

Liza uses a number of examples I was previously unaware of, and puts forth three principles: anticipate deliberate misuse, consider how code is a powerful amplifier, and show your work.

I think that the importance of showing your work is the one that have been the most overlooked. Despite things like the EU’s right to explanation, most algorithms are still opaque. And, as Liza points out, most people have vast misconceptions about artificial intelligence.

How will you think about asking for an explanation when you don’t realize there’s a question? Can you tell what was made by a computer and what wasn’t?

I share Liza’s concerns about the implications of the intersection of all three guidelines. Your news streams are already being polluted by maliciously-spawned fabrications. I’ve pointed out many ways that images can be manipulated. Now picture all those clickbait articles automated.

A politician said something inflammatory: here’s a manipulated video, with the dialog remapped. A celebrity laughed at the solemn funeral or the racist joke: here’s a smile-vector-altered photo. Dial it up to 11, automate the entire pipeline, and flood YouTube with “proof” of all kinds of things.

People with unusual or unpopular problems will suffer more: people who don’t want to listen will have a new excuse to claim everything was Photoshopped, even when they can’t really tell the difference themselves. We believe evidence that supports our beliefs and discard evidence that makes us feel uncomfortable.

So, paradoxically, we’ll have people believing all kinds of things that fit their biases while rejecting things that are true but discomfiting. If you thought Flat Earthers and Moon-Landing denialists were bad, wait until they can claim that everything is manipulated while understanding none of the science behind it. Truth becomes what we already believe.

I share Liza’s belief that generative artists can help slow this. By exposing more people to how the systems actually work, countering abuses of the tools before they happen, and taking steps to deal with the ability of computers to scale problems up to massive levels, we can make our own corners of the world a little better.

How should we design intelligent learning machines that minimize undesirable behavior?

  The widespread conversations about AI took a new turn in March 2016 when Microsoft launched, then quickly unplugged, Tay, its artificial intelligence chat robot. Within 24 hours, interactions on Twitter had turned the bot, modeled after a teenage girl, into a “Hitler-loving sex robot.”

This controversy, on the heels of the Feb. 14, 2016, accident of Google’s self-driving robot car, has ignited a new debate over artificial intelligence. How should we design intelligent learning machines that minimize undesirable behavior?

While both of the aforementioned incidents were relatively minor, they highlight a broader concern, namely, that it is very difficult to control adaptive learning machines in complex environments. The famous cybernetician Norbert Wiener warned us about this dilemma more than 50 years ago: “If we use, to achieve our purposes, a mechanical agency with whose operation we cannot interfere effectively … we had better be quite sure that the purpose put into the machine is the purpose we really desire and not merely a colorful imitation of it.”

The difficulty with learning machines arises because they train themselves on what they see, the “training data,” that cannot completely represent situations the machine will see in the future. These are sometimes called “edge cases” or “Rumsfeldian unknowns,” in that they are unknowable in advance. Humans typically deal with such edge cases remarkably well with no prior “training” by applying common sense or finding analogies and learning from these samples of one.

There is no clear answer to this vexing issue at the moment. However, it is essential for designers of such systems to assess worst-case implications in terms of the risks associated with future edge cases.