Latest / Elon Musk Podcast / Meta sacrifices human oversight for AI

Transcript

Meta sacrifices human oversight for AI · Elon Musk Podcast

• 0:00
  Meta has discussed cutting off all funding for its independent
• 0:03
  oversight board by the end of their current commitment.
• 0:06
  That is pulling the plug on a corporate governance experiment
• 0:09
  designed specifically to take the heat off Mark Zuckerberg.
• 0:12
  Yeah, I mean, that board was built to handle high stakes
• 0:15
  speech decisions. They were basically the Supreme
• 0:17
  Court for the platform. But now Meta is aggressively
• 0:21
  restructuring. They are cutting 15,000 jobs and
• 0:25
  pivoting entirely toward artificial general intelligence.
• 0:29
  So if a global platform dismantles its human safety net
• 0:32
  to fund automated intelligence, how does it govern the flow of
• 0:35
  information for billions of users?
• 0:37
  Well, the engine driving this internal restructuring is a
• 0:40
  really rigid performance review system.
• 0:43
  Managers are actually forced to place 15 to 20% of their staff
• 0:47
  into a meets most or lower category.
• 0:49
  Wait, hold on back up. Yeah.
• 0:51
  Are you saying that even if a team consists entirely of high
• 0:53
  performers, a manager is forced to penalize a set percentage of
• 0:56
  them just to meet a quota? Exactly.
• 0:59
  The system forces a distribution curve.
• 1:01
  Historically, the performance cycle was a mechanism for
• 1:05
  growth, You know, getting feedback, calculating bonuses,
• 1:08
  right? But during workforce reductions
• 1:11
  it becomes the primary mechanism for contraction.
• 1:14
  It is like grading on a curve in a class of valedictorians.
• 1:17
  Even if everyone scores a 99%, someone still has to fail just
• 1:21
  to satisfy the math. It basically turns teammates
• 1:25
  into competitors overnight. And that completely destroys the
• 1:29
  concept of a meritocracy. The internal culture changes
• 1:33
  into a survival game. An employee who gets that meets
• 1:36
  most rating faces a severe bonus cut, and they land on a stealth
• 1:39
  list for the next round of layoffs.
• 1:42
  I mean, the phrase meets most sounds perfectly fine in a
• 1:44
  normal corporate environment. It implies you are doing your
• 1:47
  job. Right.
• 1:47
  Usually it does. But in this specific
• 1:50
  mathematical curve, it is a death sentence for your career.
• 1:53
  If I am an engineer working there, I am terrified.
• 1:56
  You are essentially sacrificing your own crew just to maintain
• 1:59
  momentum. You.
• 1:59
  Really are. So you adapt your behavior to
• 2:02
  stay on board. Surviving crew members are
• 2:04
  abandoning legacy projects entirely.
• 2:07
  Whole teams are fleeing the virtual reality division Reality
• 2:10
  Labs, and they're scrambling to join the artificial intelligence
• 2:14
  divisions. Because the AI teams are the
• 2:16
  ones receiving the blank checks for computing power and
• 2:18
  headcount. Exactly.
• 2:19
  Survival requires high visibility.
• 2:22
  Internal employee accounts actually reveal that workers are
• 2:26
  explicitly instructed to demand specific written feedback from
• 2:29
  managers about their promotion track.
• 2:32
  So this connects directly to what you experience when you
• 2:34
  open the app on your phone. If you are a quiet heads down
• 2:38
  worker like the silent rinder, maintaining the basic
• 2:40
  infrastructure or working on trust and safety, you get fired.
• 2:44
  Yeah, the loud employee who aggressively aligns with the new
• 2:46
  AI mandates gets retained. So the bugs do not get fixed
• 2:50
  because the engineer who used to fix them is trying to rebrand
• 2:52
  themselves as an AI specialist. The internal focus is entirely
• 2:56
  on the new mandate. I mean the internal review
• 2:59
  system even factors in AI driven impact.
• 3:02
  It judges employees on how effectively they use internal
• 3:05
  automated tools to increase their output.
• 3:08
  If you are not utilizing the new tools, you are penalized.
• 3:11
  Wow. And this internal shift has
• 3:13
  direct external consequences. Meta is ending its third party
• 3:17
  fact checking program in the United States and replacing it
• 3:19
  with a crowdsourced community note system kind of similar to
• 3:22
  what X uses. The causal connection here is
• 3:25
  pretty direct. Mehta is slashing its human
• 3:28
  workforce to fund enormous artificial intelligence data
• 3:31
  centers, right? So to balance the books, they're
• 3:34
  offloading the expensive human intensive work of moderation
• 3:38
  onto the users themselves. Wait, let me make sure I
• 3:40
  understand the mechanics of this.
• 3:41
  They're replacing professional journalists and paid researchers
• 3:44
  with anonymous users on the Internet.
• 3:46
  Yes. Researchers classify the act of
• 3:48
  writing and rating these notes as unpaid data labor extracted
• 3:52
  from the user base. You're relying on individuals to
• 3:55
  perform complex verification work for free.
• 3:57
  That is wild. Historically, content moderation
• 4:01
  is incredibly expensive because humans suffer psychological
• 4:05
  tolls when reviewing difficult content.
• 4:08
  Exactly. And the platform is bypassing
• 4:10
  that cost entirely. But that opens up serious
• 4:13
  vulnerabilities. Trusting random users to
• 4:16
  moderate a global platform feels incredibly risky.
• 4:19
  It is the oversight board explicitly warned about this.
• 4:23
  Relying on consensus from users is incredibly dangerous in
• 4:27
  countries with a history of coordinated disinformation
• 4:29
  networks. Because the community notes
• 4:31
  system operates on a specific mathematical assumption, right?
• 4:35
  Yeah, it assumes A sufficiently diverse and independent set of
• 4:38
  contributors will evaluate content in good faith.
• 4:41
  What happens when the bad actors have more resources than the
• 4:44
  good faith users? Authoritarian regimes possess
• 4:47
  the technical sophistication to coordinate massive numbers of
• 4:50
  accounts. Right, so the algorithm
• 4:51
  calculates a score based on whether contributors who usually
• 4:55
  disagree with each other find a note helpful.
• 4:57
  He looks for a bridge between divided groups.
• 4:59
  But if a state sponsored troll farm floods the system with
• 5:02
  coordinated accounts, they can artificially simulate that
• 5:05
  historical disagreement. Oh I see, they can build fake
• 5:08
  profiles that appear to sit on opposite sides of the political
• 5:11
  spectrum and then have those accounts agree on a deceptive
• 5:14
  note. Exactly.
• 5:16
  And when they do that, the assumption of good faith
• 5:18
  consensus collapses. They can exploit the feature to
• 5:21
  manipulate the information ecosystem.
• 5:23
  They are essentially using the platform's new safety feature as
• 5:27
  a weapon. If a coordinated network
• 5:29
  downvotes accurate context or upvotes deceptive context, the
• 5:33
  algorithm is mathematically twicked into believing A
• 5:36
  consensus has been reached. And that risk becomes vastly
• 5:39
  more acute as artificial intelligence facilitates the
• 5:41
  scaled creation and operation of these fake networks.
• 5:45
  You no longer need human operators sitting in a warehouse
• 5:48
  to run these sick profiles. Yeah, the oversight board
• 5:50
  recommended that Meta omit countries with a historical
• 5:53
  pattern of intentional large scale disinformation networks
• 5:56
  from this program entirely. Right.
• 5:58
  Eventually, inclusion should require rigorous testing like
• 6:02
  red teaming program vulnerabilities just to prove
• 6:05
  the safeguards actually work. But we are already seeing the
• 6:08
  friction points during active conflicts.
• 6:11
  During the Israel Iran war, a completely deceptive AI
• 6:15
  generated video showing extensive damage to buildings in
• 6:18
  Haifa received over 700,000 views on Facebook before being
• 6:22
  caught. With professional fact checkers
• 6:24
  removed and Crowdsource notes vulnerable to manipulation,
• 6:28
  platforms are exposed to the exact artificial intelligence
• 6:31
  tools they are currently spending billions to develop.
• 6:34
  It creates a massive blind spot where deceptive output Garner's
• 6:37
  huge numbers of views in a soft war.
• 6:40
  Exactly the The algorithms that recommend content prioritize
• 6:43
  engagement, and artificially generated conflict videos are
• 6:46
  perfectly designed to maximize that engagement.
• 6:49
  As a consequence, Meta has been forced to apply AI info labels
• 6:53
  to manipulated media rather than deleting it.
• 6:55
  This limits the platform's burden regarding free speech
• 6:58
  moderation. Right, but it places the
• 6:59
  responsibility entirely on you, the listener, to discern
• 7:03
  reality. The Oversight Board actually
• 7:05
  recommended labeling as a less restrictive alternative to
• 7:08
  deletion. The premise is that not all
• 7:11
  manipulated media is harmful. Think of political satire or
• 7:15
  harmless parodies. The idea is to provide context
• 7:18
  without unduly restricting expression, but I have to push
• 7:22
  back here. Is labeling actually better than
• 7:25
  deletion when the automated systems making these decisions
• 7:28
  are incredibly flawed? That is a fair question.
• 7:31
  I mean, Meta's algorithms once flagged a photograph of onions
• 7:34
  as sexual content. Oh yeah, that is a perfect
• 7:36
  example of how these moderation algorithms fail.
• 7:39
  They rely on things called word embeddings and image recognition
• 7:42
  patterns. OK, what exactly is a word
• 7:44
  embedding? How does the algorithm confuse a
• 7:47
  vegetable with explicit material?
• 7:50
  Think of word embeddings as a multi dimensional mathematical
• 7:53
  map of concepts. The algorithm plots words based
• 7:56
  on how often they appear together in training data.
• 7:59
  So it does not actually understand what an onion is.
• 8:01
  No, not at all. It just calculates distance
• 8:04
  between data points. If the word spicy or certain
• 8:08
  curved shapes are mapped closely to restricted content, and the
• 8:11
  photo of the onion triggers those specific mathematical
• 8:14
  coordinates, the system fails. The data set used to train these
• 8:18
  models often reproduces strange associations, so we are trusting
• 8:23
  the math that banned an Onion to accurately label sophisticated
• 8:27
  political deepfakes in a war zone.
• 8:29
  Which is exactly why researchers are exploring alternative
• 8:32
  methods to fix these inherent biases, like paraphrasing
• 8:36
  technology. I was reading about this in the
• 8:37
  sources and frankly it sounds terrifying.
• 8:40
  It is a smart filter that automatically rewrites a user's
• 8:43
  hateful post into something polite before the recipient sees
• 8:46
  it. Instead of deleting the speech
• 8:47
  or labeling it, the system alters the semantic value of the
• 8:51
  message entirely. That is a real time digital
• 8:54
  ventriloquist that crosses an ethical line regarding speaker
• 8:57
  autonomy. A user's speech would be
• 8:59
  secretly altered by an algorithm without their knowledge.
• 9:03
  Imagine if you post a sarcastic criticism of an authoritarian
• 9:06
  regime and the smart filter decides your tone is too
• 9:09
  hostile. It rewrites your post to be
• 9:12
  polite, effectively making it look like you are endorsing the
• 9:15
  regime. You are completely distorting
• 9:17
  the author's intent to maintain a sanitized environment.
• 9:20
  Under international human rights law, restrictions on freedom of
• 9:24
  expression must be clearly articulated so that speakers
• 9:28
  know what the rules are. If a smart filter is silently
• 9:31
  rewriting your text, the interest protected by freedom of
• 9:34
  speech, like individual self definition and democratic
• 9:37
  engagement, are seriously compromised.
• 9:40
  The legal frameworks we rely on assume that the words published
• 9:43
  under your name are actually the words you wrote.
• 9:46
  And if Meta's automated systems are struggling to tell the
• 9:48
  difference between onions and deepfakes, and they are
• 9:51
  replacing human moderators with unpaid users, who is holding
• 9:55
  them accountable? Because depending on where you
• 9:56
  live, the government's answer is entirely contradictory.
• 10:00
  Global governments are stepping in to mandate how the Internet
• 10:03
  should be policed. The European Union's Digital
• 10:05
  Services Act legally forces platforms to maintain strict
• 10:08
  moderation. While a federal appeals court in
• 10:11
  the United States upheld a Texas law prohibiting platforms from
• 10:15
  censoring user viewpoints. Exactly.
• 10:18
  That is like driving a car where the passenger in Europe is
• 10:21
  yanking the steering wheel left and the passenger in Texas is
• 10:24
  yanking it right. This severely limits Metas
• 10:27
  ability to operate a unified global platform.
• 10:30
  In Europe, the Digital Services Act implements systemic risk
• 10:33
  assessments. It creates trusted flagger
• 10:36
  programs and allows users to legally challenge platforms
• 10:39
  through independent, out of court dispute settlements.
• 10:41
  Let me clarify the Trusted Flagger program.
• 10:44
  This gives specific organizations A specialized
• 10:47
  status where their reports of illegal content are processed
• 10:49
  faster than a normal user's report.
• 10:51
  Correct. And platforms face immense
• 10:54
  financial penalties for non compliance with the European
• 10:56
  Framework, including fines of up to 6% of their global turnover.
• 11:00
  We are talking about billions of dollars in potential penalties.
• 11:04
  Yeah, Meanwhile, the Texas law treats social media platforms
• 11:07
  like common carriers, similar to a telephone company, restricting
• 11:10
  their ability to remove content based on viewpoint.
• 11:13
  A phone company just connects a wire, but a social network
• 11:16
  curates a feed. Those are completely different
• 11:19
  functions. If you treat a social network
• 11:22
  like a public utility, they lose the ability to stop spam or
• 11:25
  organized harassment because they have to serve everyone
• 11:28
  equally. The legal reasoning behind the
• 11:31
  Texas law argues that platforms merely engage in viewpoint based
• 11:34
  censorship with respect to expression they have already
• 11:37
  disseminated. They argue these companies are
• 11:40
  the modern public square. But this directly conflicts with
• 11:43
  European standards, where public incitement to hatred or violence
• 11:47
  is strictly outlawed, requiring proactive moderation.
• 11:50
  Right. They are legally required to
• 11:51
  leave content up in one jurisdiction and legally
• 11:54
  required to take it down in another.
• 11:56
  And when platforms face immense pressure to moderate, they
• 11:59
  quickly retreat when legally challenged.
• 12:01
  Meta tried to implement APG 13 rating for teen accounts to
• 12:05
  signal safety. The intention was to borrow the
• 12:08
  cultural familiarity of a movie rating to show parents they were
• 12:11
  taking teen safety seriously. Why did they cave and abandoned
• 12:14
  the idea? Well, the Motion Picture
• 12:15
  Association complained about the trademark.
• 12:18
  Trademark litigation is immediate and highly costly,
• 12:21
  whereas platform safety is abstract.
• 12:24
  It proves that despite creating these elaborate systems for
• 12:27
  safety, corporate policy is incredibly fragile When
• 12:31
  confronted with external legal friction.
• 12:33
  They will abandoned a safety feature the moment it threatens
• 12:36
  their legal standing with another powerful corporation.
• 12:39
  Exactly. Let me summarize the reality of
• 12:42
  what we are looking at. Meta is funding its artificial
• 12:45
  intelligence ambitions by stripping away its human
• 12:47
  workforce, professional fact checkers and independent
• 12:50
  oversight. The systems designed to protect
• 12:52
  the user base are being automated or outsourced to the
• 12:56
  users themselves. It leaves you wondering, when
• 12:58
  the next global crisis floods the Internet with targeted
• 13:01
  misinformation, will the remaining algorithms and unpaid
• 13:04
  users be enough to hold the line?
• 13:06
  Or are we entirely on our own? If you're not subscribed yet,
• 13:09
  take a second and hit follow on whatever app you're using.
• 13:12
  It helps us keep making this. We appreciate you being here.