Latest / Elon Musk Podcast / Innocent people jailed by faulty facial recognition

Transcript

Innocent people jailed by faulty facial recognition · Elon Musk Podcast

• 0:00
  Angela Lips, a 50 year old grandmother from Tennessee,
• 0:04
  spent nearly half a year in a North Dakota jail for an
• 0:07
  organized bank fraud scheme she had like, absolutely nothing to
• 0:10
  do with. Wow.
• 0:12
  Yeah, and this happened entirely because the West Fargo Police
• 0:15
  Department relied on a flawed facial recognition match
• 0:19
  generated by Clearview AI software.
• 0:22
  That is just, I mean, law enforcement agencies are
• 0:24
  treating mathematical probabilities and, you know,
• 0:27
  algorithmic suggestions as undeniable proof of guilt.
• 0:30
  Right now. They are skipping basic
• 0:32
  detective work entirely and directly targeting innocent
• 0:35
  people. And it really establishes how
• 0:37
  police reliance on artificial intelligence is quietly
• 0:40
  rewriting the rules of criminal justice for you and honestly,
• 0:44
  for everyone in your community. So how does a mathematical
• 0:47
  probability generated by a computer turn into a tactical
• 0:50
  team at an innocent person's front door?
• 0:52
  Well. The initial catalyst for all
• 0:53
  this was when investigators in West Fargo ran a low resolution
• 0:56
  image of a counterfeit military ID through Clearview A.
• 1:00
  There's a low res image. Right.
• 1:01
  And Clearview is a startup whose database is literally just
• 1:06
  scraped from the Internet and social media.
• 1:09
  So it's a massive pool of random images.
• 1:11
  Yeah, and the software flagged Lips as a potential suspect.
• 1:15
  And then West Fargo forwarded this intelligence to the Fargo
• 1:18
  Police Department, and Fargo detectives just assumed
• 1:21
  corroborating surveillance photos had also been submitted
• 1:24
  and verified. So they thought the homework was
• 1:27
  already done. Exactly.
• 1:28
  So instead of investigating further, an officer simply
• 1:31
  looked at lips social media profiles, decided her facial
• 1:35
  features and hairstyle matched the suspect, and secured A
• 1:38
  felony arrest warrant. We back up.
• 1:41
  They secured A felony warrant just from the computer match and
• 1:45
  a quick glance at Facebook. Yeah, they really did, because
• 1:49
  they failed to route the request through the certified North
• 1:52
  Dakota State and Local Intelligence Center.
• 1:54
  Oh, OK. They bypassed the established
• 1:56
  protocols completely. Which eliminates the
• 1:59
  foundational requirement of probable cause.
• 2:01
  It opens up a reality where you can be deemed A fugitive based
• 2:05
  entirely on a software guess without a single officer
• 2:08
  verifying your actual location. And the physical reality of that
• 2:12
  error is just, it's horrifying. Yeah, I mean, US marshals
• 2:15
  actually arrested lips at gunpoint while she was just
• 2:18
  sitting there watching her grandchildren.
• 2:19
  That is terrifying to even think about.
• 2:22
  And because she was classified as a fugitive from justice, she
• 2:25
  was held without bail. So over her months in jail, she
• 2:28
  lost her rental home, her car, her health insurance.
• 2:32
  Oh my. God and her elderly dog even had
• 2:34
  to be rehomed. You know, there is another
• 2:36
  victim we should talk about too. 61 year old Harvey Murphy
• 2:39
  Junior. Yeah, he was falsely identified
• 2:41
  by facial recognition tools used by Macy's and Sunglass Hut,
• 2:44
  which they're owned by Essler Luxottica, for an armed robbery
• 2:49
  in Texas while he was actually in California.
• 2:52
  And the outcome for Murphy was horrific.
• 2:54
  Yeah, After being jailed on that false warrant, he was sexually
• 2:57
  assaulted and beaten by three men while behind bars.
• 3:00
  We are talking about lines of code destroying a human life in
• 3:03
  a matter of seconds. Exactly.
• 3:05
  It changes the threat of artificial intelligence from
• 3:08
  abstract data privacy concerns into immediate physical danger
• 3:12
  and a total loss of liberty for you.
• 3:14
  So to understand why the software makes these specific
• 3:17
  errors, we have to look at MI TS Gender shade study by Joy
• 3:20
  Bulimwini and research from the National Institute of Standards
• 3:24
  and Technology. And the data there is striking
• 3:27
  the error rates for facial recognition identifying light
• 3:29
  skinned men are less than 1%. Super low.
• 3:32
  But for darker skinned women, the error rate balloons to
• 3:35
  nearly 35%. And age is another massive
• 3:38
  failure point. The algorithm has really
• 3:40
  struggled to read the changing facial structures of the elderly
• 3:44
  and children. So trusting this software is
• 3:47
  like it's like relying on a star eyewitness who has perfect 2020
• 3:50
  vision for one specific demographic but needs incredibly
• 3:54
  thick prescription glasses to identify anyone else.
• 3:57
  So. That is a perfect analogy, and
• 3:58
  that limits the reliability of the software drastically.
• 4:01
  It exposes marginalized groups, as well as older citizens like
• 4:05
  Lips and Murphy, to a disproportionately high risk of
• 4:08
  false arrest. So mechanically, facial
• 4:12
  recognition basically converts pixels into mathematical vectors
• 4:15
  called face prints. Right face prints.
• 4:17
  And low quality surveillance footage creates really weak face
• 4:21
  prints, generating a low similarity score that's just
• 4:25
  filled with uncertainty. Hold on.
• 4:27
  If the score is low, why do the police act on it?
• 4:30
  Well, it all comes down to thresholds, which is essentially
• 4:34
  the acceptable margin of error. Without federal regulations,
• 4:37
  police departments lack mandatory second review
• 4:39
  protocols, so they just act on these shaky similarity figures
• 4:44
  without even recording the algorithm's confidence
• 4:46
  percentage in their reports. See, I feel like the primary
• 4:48
  fault lies with the tech companies here.
• 4:50
  They're the ones selling defective, unverified tools to
• 4:53
  local municipalities. I hear that, I really do.
• 4:56
  But I have to push back. I think law enforcement alone is
• 4:59
  to blame for actively choosing to treat a probable
• 5:01
  investigative tool as an absolute truth just to speed up
• 5:05
  their casework. Well, regardless of who is more
• 5:07
  at fault, this lack of standardized guardrails changes
• 5:11
  local police departments into beta testers literally gambling
• 5:15
  with your freedom to streamline their paperwork.
• 5:17
  And the resolution of the Lips case is just wild.
• 5:21
  She was only cleared after a public defender, Jay Greenwood,
• 5:24
  pulled her bank records. Right, the actual financial
• 5:27
  receipts. Because while the fraud occurred
• 5:29
  in North Dakota, Lips was 1200 miles away in Tennessee, buying
• 5:33
  pizza, picking up cigarettes at a gas station and ordering Uber
• 5:37
  Eats. In the aftermath of all that,
• 5:39
  Fargo Police Chief Dave Zabalski implemented a temporary
• 5:43
  directive restricting facial recognition use to the Criminal
• 5:46
  Investigations Division and banning the use of neighboring
• 5:49
  department software. But despite the drop charges,
• 5:52
  the police refused to directly apologize to Lips, claiming the
• 5:56
  investigation was still active and she hadn't been fully ruled
• 5:59
  out of a large conspiratorial operation.
• 6:01
  Which is just crazy because that shifts the burden of proof
• 6:04
  entirely on to the accused. The system now requires you to
• 6:07
  definitively prove a computer wrong with literal receipts just
• 6:11
  to win your life back. So treating statistical software
• 6:13
  as a substitute for human investigation bypasses your
• 6:17
  constitutional protections and destroys innocent lives.
• 6:21
  It really makes you wonder how many other algorithmic errors
• 6:24
  are currently sitting in jail cells right now, completely
• 6:27
  unable to access the simple receipts that could save them.
• 6:30
  If you're not subscribed yet, take a second and hit follow on
• 6:33
  whatever app you're using. It helps us keep making this.
• 6:35
  We appreciate you being here.