Audio Logs 48: The Blur

Julian Pankratz

This is Audio Logs. I’m Julian Pankratz. Alizebeth Bequin directs Frothly’s privacy programs, including Scramble, which builds tools to protect people from facial recognition. Alizebeth, welcome.

Alizebeth Bequin

Thank you. It’s nice to be remembered long enough to be invited.

Julian Pankratz

That sounds like a joke with a story behind it.

Alizebeth Bequin

People have always found my face hard to remember. Colleagues walk past me in corridors. Waiters forget my order. For years I thought of it as a small curse. These days I think of it as a professional qualification.

Julian Pankratz

Give me the worst example.

Alizebeth Bequin

A man I’d worked beside for six years once introduced himself to me at a party. Twice. Same party. The second time he said, “I feel like we’ve met.” And I said, “We have. About forty minutes ago.”

Julian Pankratz

Ha! Oh, that’s brutal.

Alizebeth Bequin

He was mortified. I wasn’t. I’d had a lot of practice. But that’s only half the story.

Julian Pankratz

What’s the other half?

Alizebeth Bequin

Airports. About ten years ago I walked up to one of the first automated passport gates, and it knew me in under a second. Then, in the same terminal, a colleague I’d known for years walked straight past me. People had always forgotten me. The machine never would. And I remember thinking, the curse is over, and not just for me. That gift of being nobody in particular was being taken away from everyone, and most of them never knew they’d had it.

Julian Pankratz

Is that what brought you to this work?

Alizebeth Bequin

That, and what came before. I’d worked alongside Gregor Eisenhorn for a long time, in the kind of work where being forgettable was useful. You learn that the person nobody remembers can go almost anywhere. I didn’t want that to become something only a few people with forgettable faces get to keep.

Julian Pankratz

Scramble itself started in 2017.

Alizebeth Bequin

With a few reporters whose sources had been identified from footage. The first toolkit was built with them, more or less around their kitchen tables. Back then it was almost all about faces. Within a few years the systems had learned to follow people by their clothes and the way they walk, so we had to follow the systems.

Julian Pankratz

Scramble’s tagline is that it makes people “a blur.” What does that mean in practice?

Alizebeth Bequin

It means a recognition system looks at you and fails to settle on an identity. Not that you disappear, which would draw attention, but that you become nobody in particular. Our tools work on several fronts at once: the face, the way you walk, the patterns in how you move through a city. People are recognized by much more than their faces now.

Julian Pankratz

Let’s take those one at a time. But first, why isn’t disappearing the goal? Wouldn’t that be safer?

Alizebeth Bequin

I wrote a working paper about this a couple of years ago called “Null Signatures,” and the short version is that perfect invisibility is extremely conspicuous. A hole in the data is a shape. If a hundred people walk past a camera, and ninety-nine of them come back with a name, and one comes back as a shimmering question mark, guess who gets looked at by a human being. There’s a police shroud on our own Registry, licensed for undercover officers, that projects a constantly shifting composite of faces and bodies. Beautiful engineering. Completely useless at a protest, because everyone would stare at you.

Julian Pankratz

So the trick is to be dull.

Alizebeth Bequin

Uninteresting beats unreadable. Every time. We learned that the hard way. Early on we had a printed pattern, a sort of scarf, that defeated three commercial face detectors in the lab. We gave it to field testers in two cities, and within a week one of them wrote back, “I look like I’m on my way to a costume party. A police officer stopped me for the scarf alone.” She was right. We shelved it.

Julian Pankratz

So what happens now, before something leaves the lab?

Alizebeth Bequin

Nothing leaves the lab until the people who’ll actually use it have worn it on an ordinary Tuesday and told us nobody looked twice. Every tool we make is built with the people it’s for. They’re much better critics than we are.

Julian Pankratz

The face, I think most people understand. What about the way you walk?

Alizebeth Bequin

Gait. Your stride, your cadence, the way your shoulders move. It’s surprisingly distinctive, and it works from a long way off, from behind, at low resolution, with your face covered. A mask does nothing about it. What helps is mostly unglamorous. Different shoes change your cadence more than you’d expect. A bag carried on one shoulder changes your posture. We publish guidance on all of that. It isn’t science fiction. It’s mostly shoes.

Julian Pankratz

It’s mostly shoes. I love that.

Alizebeth Bequin

I’m afraid so. Nobody ever wants to hear it.

Julian Pankratz

Which is bad news for my boots.

Alizebeth Bequin

Julian, your boots could be picked out from orbit.

Julian Pankratz

They were a gift.

Alizebeth Bequin

From someone who wanted to keep track of you, presumably.

Julian Pankratz

Ha! Moving swiftly on. And the patterns in how you move through a city?

Alizebeth Bequin

That’s the one that worries me most. People call it pattern of life. Your phone checks in with cell towers and Wi-Fi networks all day. Your transit card logs every gate you pass through. Cameras read your number plate. None of those know your face, and they don’t need to. Put them together and you can work out where someone sleeps, where they work and who they visit on Thursday evenings. The face is just the thing that ties it all to a name. Our colleagues on Glossia protect what you say. We’re trying to protect the part encryption can’t reach, which is your body, standing in a public place.

Julian Pankratz

So, for listeners keeping score: the face, the walk and the pattern of your days. And the face is almost the least of it.

Alizebeth Bequin

The face is the label on the file. The rest is the file.

Julian Pankratz

Who uses it?

Alizebeth Bequin

Journalists meeting sources. Organizers who need to attend a protest without being catalogued. People leaving abusive relationships whose former partners have access to camera networks. Around a hundred and ninety thousand people in thirty-eight countries, as of this year.

Julian Pankratz

Can you tell me a story? Without anything that could identify anyone, obviously.

Alizebeth Bequin

A composite, with the details changed. A reporter meets a source in a café, twice. She leaves her phone at home both times. She’s careful. Six months later the source is identified anyway. How? His transit card showed the same station on both days, and the café had a camera over the door that fed into the building’s security system. Nobody ever recognized anybody’s face. They matched a coat.

Julian Pankratz

A coat.

Alizebeth Bequin

A very good coat, in fairness. Distinctive. We tell people now that the most identifying thing they own is probably hanging in their wardrobe.

Julian Pankratz

And the people leaving abusive relationships? That feels like a very different kind of threat.

Alizebeth Bequin

It is, and in some ways it’s the hardest, because the person looking for you knows you. They know your routines, your friends, your old passwords. And very often what finds someone isn’t a sophisticated camera network at all. It’s a family photo album still syncing to a shared account, with location tags switched on. It’s the app for the car. It’s a tracker sewn into the lining of a bag. And yes, sometimes it’s an ex who works security for a building, or knows someone who does, and can pull up footage. So that work starts with an audit of everything that’s still connected to the old life. It’s slow and tedious, and it saves lives.

Julian Pankratz

What do people in that situation actually want from you?

Alizebeth Bequin

Almost never to vanish. The first thing most of them ask is how to keep their life. They want to take the kids to school, keep their job, go to their sister’s wedding. So our job isn’t to make anyone disappear. It’s to let them go to the wedding without being found there.

Julian Pankratz

How do those people find you?

Alizebeth Bequin

Mostly they don’t, and that’s fine. We train the people they already trust: shelters, legal aid clinics, newsrooms, unions. Then those people train others. Most of the people Scramble protects have never heard of Frothly. Which, given what we do, is the whole point.

Julian Pankratz

Critics say tools like this help criminals evade the police.

Alizebeth Bequin

I understand the concern. But being unrecognized in public used to be the normal condition of life. Nobody had to justify it. Mass recognition is the new thing, and it’s the new thing that should have to make its case. We are trying to preserve something people had for all of history until about ten years ago.

Julian Pankratz

What changed, ten years ago?

Alizebeth Bequin

Three things arrived at once. Cameras got cheap, storage got cheaper, and software got good enough to search the footage. A camera used to be a witness nobody ever asked. The tapes were recorded over after a week, because nobody could possibly watch them all. Now it’s a witness that will answer any question, about anyone, forever. That isn’t a change of degree. It’s a different world.

Julian Pankratz

Let me push a little, though. There must be cases where recognizing a face in a crowd catches someone genuinely dangerous.

Alizebeth Bequin

There are, and I won’t pretend otherwise. But the police have warrants, witnesses, investigations, all the ordinary tools of a free society. What I object to is the default: recognizing everyone, everywhere, all the time, just in case. And here’s the uncomfortable part. Serious criminals and intelligence services already know how to stay off camera. They have budgets for it. The people who use our tools are the ones who can’t afford a professional.

Julian Pankratz

And the old line, that if you’ve done nothing wrong, you’ve nothing to hide?

Alizebeth Bequin

I’d ask that person where they were last Tuesday at four, and with whom, and whether they’d like it published. Everyone walks into buildings they’d rather choose when to mention. A clinic. A lawyer’s office. A job interview while they’re still employed. A meeting in a church basement. Being seen going in is information. Privacy was never about hiding wrongdoing. It’s about deciding who gets to know your life.

Julian Pankratz

Your colleague Robert Arctor has a paper out this year on how recognition systems explain their decisions. Your name’s on it too.

Alizebeth Bequin

It’s an important paper, and it’s really Robert’s. I mostly argued with him until he proved me wrong. When a system flags someone, it produces an explanation: a heat map, a list of features. Robert’s work shows those explanations are often wrong. The system says it looked at your face; in fact it looked at your coat. If explanations like that are going to be used as evidence, people deserve to know how unreliable they are.

Julian Pankratz

There’s that coat again.

Alizebeth Bequin

There’s that coat again. We tested six commercial systems. Every one scored above eighty-nine percent on the standard benchmarks. Not one gave a faithful explanation more than about half the time. And when a system pointed at someone’s face, the decision actually depended on the face less than four times in ten. The rest was clothing, the background, the lighting, even which camera took the picture.

Julian Pankratz

So the system gets the answer right, but it can’t tell you why.

Alizebeth Bequin

Or it’s learned what a good explanation is supposed to look like. Reviewers expect to see a face lit up, so the face gets lit up. It’s telling people what they want to hear.

Julian Pankratz

Robert has an unusual background for a researcher, doesn’t he?

Alizebeth Bequin

He spent several years doing undercover field work before he came to us, and he’s watched more surveillance footage than anyone in the building. During the testing, he reviewed a batch of flagged clips without knowing that some of them were of him. He didn’t recognize himself in three out of five.

Julian Pankratz

Ha! How did he take that?

Alizebeth Bequin

He said it was the most disorienting afternoon of his career. I told him, welcome to my life. But there’s a serious point under it. When a person reviews a machine’s decision, they only see what the machine puts in front of them. So we’ve changed how we test our own tools. We don’t care any more whether a tool changes the explanation a system gives. We care whether it changes the decision. Only the decision has consequences.

Julian Pankratz

So what should change, if the explanations can’t be trusted?

Alizebeth Bequin

The simplest thing is that an explanation shouldn’t count as evidence until the system has passed independent testing of whether its explanations are faithful. Reviewers should see the raw footage, not the system’s highlighted version of it. We’ve asked our own Registry to make that testing a condition of licensing surveillance systems, and our colleagues in Zürich are considering it carefully, which is how Zürich considers everything. And if a system is ever used against you, ask for its decision to be tested before anyone treats its explanation as fact.

Julian Pankratz

Do the tools keep working? I’d imagine these systems are updated constantly.

Alizebeth Bequin

Nothing we make works forever. The systems are retrained all the time, and something that worked in March may not work in September. Robert’s team retests against the systems people are actually facing, and when a tool stops working, we say so publicly, even when it’s embarrassing. A tool that gives you false confidence is worse than no tool at all.

Julian Pankratz

Do you use them yourself?

Alizebeth Bequin

I test all of them, although I hardly need them, which makes me a terrible test subject. Robert calls me the control group. But yes. And I own a very boring coat.

Julian Pankratz

Listeners who visit our staff pages will notice that all of our portraits are shown through Scramble. Including mine, about which I’ll admit I lodged a complaint.

Alizebeth Bequin

You did. In writing. In verse, as I recall.

Julian Pankratz

It was a short verse.

Alizebeth Bequin

It rhymed “blur” with “connoisseur.” I’ve kept it.

Julian Pankratz

As you should. Say I’m going to a march this weekend. Peaceful, perfectly legal. What would you tell me?

Alizebeth Bequin

I’d start by asking who, specifically, might be looking for you. Not “the government.” A name, if you can. The answer changes everything. For most people at a peaceful march, the useful things are boring. Wear what everyone else is wearing. Leave your phone at home, or switch it off, not just to silent. Agree a meeting point with your friends beforehand, so you don’t need it. And the big one: don’t post photos of other people’s faces.

Julian Pankratz

That last one’s interesting. It’s about everyone else.

Alizebeth Bequin

Privacy at a protest is a group project. You can do everything right and still be identified because the person next to you posted a selfie with you in the background. One of our tools blurs every face in a photo and strips out the location before you share it. It’s the most popular thing we make, and it mostly protects people other than the person using it. I find that very moving, actually.

Julian Pankratz

What do you most want listeners to understand?

Alizebeth Bequin

That privacy in public isn’t a contradiction. You can be seen without being identified. You can walk down a street without leaving a permanent record of having done so. That used to be called freedom. It still should be.

Julian Pankratz

Alizebeth Bequin, thank you.

Alizebeth Bequin

Thank you. I hope you’ll remember me.

Julian Pankratz

I’ve written it down. Twice. That was Audio Logs. Scramble’s tools and its digital safety curriculum are free to anyone at risk, and Robert and Alizebeth’s paper, “Through a Glass, Darkly,” is in our research library. We’ll be back in two weeks.

The Blur

Transcript