Amazon heads off facial recognition rebellion

By Leo Kelion
Technology desk editor

Image source, Amazon
Image caption,
Rekognition can match photos to databases holding millions of people's faces

Shareholders seeking to halt Amazon's sale of its facial recognition technology to US police forces have been defeated in two votes that sought to pressure the company into a rethink.

Civil rights campaigners had said it was "perhaps the most dangerous surveillance technology ever developed".

But investors rejected the proposals at the company's annual general meeting.

That meant less than 50% voted for either of the measures.

A breakdown of the results has yet to be disclosed.

The first vote had proposed that the company should stop offering its Rekognition system to government agencies.

The second had called on it to commission an independent study into whether the tech threatened people's civil rights.

The ballot in Seattle would have been non-binding, meaning executives would not have had to take specific action had either been passed.

Amazon had tried to block the votes but was told by the Securities and Exchange Commission that it did not have the right to do so.

Image source, Amazon
Image caption,
Rekognition gives a confidence score as to whether a person's face is a match

"We will see what the tally is, but one of our primary objectives was to bring this before shareholders and the board, and we succeeded in doing that," Mary Beth Gallagher from the Tri-State Coalition for Responsible Investment told the BBC.

"This is just the beginning of this movement for us and this campaign will continue. We have built links to civil rights groups, employees and other stakeholders.

"And the most important thing is that regardless of the result, we still want the board to halt sales of Rekognition to governments, and it has the capacity to do that."

The American Civil Liberties Union added that the very fact there had been a vote was "an embarrassment to Amazon" and should serve as a "wake-up call for the company to reckon with the real harms of face surveillance".

Amazon has yet to comment.

But ahead of the votes it said it had not received a single report of the system being used in a harmful manner.

"[Rekognition is] a powerful tool... for law enforcement and government agencies to catch criminals, prevent crime, and find missing people," its AGM notes state.

"New technology should not be banned or condemned because of its potential misuse."

Image source, Amazon
Image caption,
Amazon has promoted its tech as a tool to fight crime

Face matches

Rekognition is an online tool that works with both video and still images and allows users to match faces to pre-scanned subjects in a database containing up to 20 million people provided by the client.

In doing so, it gives a confidence score as to whether the ID is accurate.

In addition, it can be used to:

  • detect "unsafe content" such as whether there is nudity or "revealing clothes" on display
  • suggest whether a subject is male or female
  • deduce a person's mood
  • spot text in images and transcribe it for analysis

Amazon recommends that law enforcement agents should only use the facility if there is a 99% or higher confidence rating of a match and says they should be transparent about its usage.

Image source, Amazon
Image caption,
Rekognition can be used to flag "suggestive content"

But one force that has used the tech - Washington County Sheriff's Office in Hillsboro, Oregon, - told the Washington Post that it had done so without enforcing a minimum confidence threshold, and had run black-and-white police sketches through the system in addition to photos.

A second force in Orlando, Florida has also tested the system. But Amazon has not disclosed how many other public authorities have done so.

Biased algorithms?

Part of Rekognition's appeal is that it is cheaper to use than several rival facial recognition technologies.

But a study published in January by researchers at Massachusetts Institute of Technology and the University of Toronto suggested Amazon's algorithms suffered greater gender and racial bias than four competing products.

It said that Rekognition had a 0% error rate at classifying lighter-skinned males as such within a test, but a 31.4% error rate at categorising darker-skinned females.

Image source, Amazon
Image caption,
Amazon said that it ran tests against images of men and women from six ethnicities to check for signs of bias

Amazon has disputed the findings saying that the researchers had used "an outdated version" of its tool and that its own checks had found "no difference" in gender-classification across ethnicities.

Even so, opposition to Rekognition has also been voiced by civil liberties groups and hundreds of Amazon's own workers.

Ms Gallagher said that shareholders were concerned that continued sales of Rekognition to the police risked damaging Amazon's status as "one of the most trusted institutions in the United States".

"We don't want it used by law enforcement because of the impact that will have on society - it might limit people's willingness to go in public spaces where they think they might be tracked," she said.

But one of the directors from Amazon Web Services - the division responsible - had told the BBC that it should be up to politicians to decide if restrictions should be put in place.

Image source, Amazon
Image caption,
Rekognition can give a confidence score for several different features

"The right organisations to handle the issue are policymakers in government," Ian Massingham explained.

"The one thing I would say about deep learning technology generally is that much of the technology is based on publicly available academic research, so you can't really put the genie back in the bottle.

"Once the research is published, it's kind of hard to 'uninvent' something.

"So, our focus is on making sure the right governance and legislative controls are in place."