Latest / Elon Musk Podcast / Trump administration sues states over AI laws

Transcript

Trump administration sues states over AI laws · Elon Musk Podcast

• 0:00
  The Trump administration is leveraging billions in broadband
• 0:03
  Internet funding and a specialized Department of
• 0:06
  Justice task force to sue individual U.S. states over
• 0:10
  their artificial intelligence laws.
• 0:12
  Yeah, the intensity of the federal government essentially
• 0:15
  taking its own states to court to protect tech innovation is,
• 0:19
  well, it's striking. We are watching a really high
• 0:22
  stakes collision. The White House is pushing
• 0:24
  aggressively for a single minimally burdensome national
• 0:29
  standard. But on the exact opposite side,
• 0:31
  you have individual states aggressively passing local
• 0:35
  protections for their citizens, covering, you know, everything
• 0:38
  from employment bias to child safety.
• 0:41
  So if the federal government successfully erases local tech
• 0:44
  roles, who actually decides how artificial intelligence behaves
• 0:48
  in your daily life? Well, the administration issued
• 0:50
  an executive order specifically targeting state laws that they
• 0:54
  view as a direct threat to US global tech dominance.
• 0:58
  And the strategy here is highly coordinated.
• 1:01
  I mean, it operates across multiple federal agencies.
• 1:04
  First, they directed the Attorney general to establish an
• 1:07
  AI Litigation Task Force within the Department of Justice.
• 1:11
  And their job is just to sue. Pretty much the sole
• 1:14
  responsibility of this group is to challenge state artificial
• 1:18
  intelligence laws in federal court.
• 1:21
  Their primary argument relies on the dormant Commerce Clause.
• 1:24
  They're claiming these local laws unconstitutionally burdened
• 1:27
  Interstate commerce. Hold on.
• 1:30
  For those of us who haven't been a law school, what does dormant
• 1:32
  commerce actually mean here? It is a constitutional principle
• 1:36
  that prevents individual states from passing laws that
• 1:39
  accidentally or intentionally build a toll booth on Interstate
• 1:43
  business. OK, so look, preventing trade
• 1:45
  wars between states. Exactly.
• 1:47
  The idea goes back to the Founders wanting to prevent
• 1:50
  trade wars between states. In this context, the Department
• 1:53
  of Justice is arguing that computer code and data, well,
• 1:56
  they naturally cross state lines constantly.
• 1:58
  Which is true. Right.
• 1:58
  So if a tech company in California has to alter its
• 2:01
  fundamental software architecture just to legally
• 2:04
  operate in a single state, the federal government argues that
• 2:07
  local law is illegally disrupting the national economy.
• 2:11
  So they are actively taking states to court to strike down
• 2:15
  laws those states voted. For yes, then, the Commerce
• 2:19
  Department is directed to publish a formal evaluation,
• 2:22
  which functions as a hit list identifying specific state tech
• 2:26
  laws they consider onerous. Onerous how?
• 2:28
  They're particularly looking for laws that force models to alter
• 2:31
  truthful outputs or mandate disclosures that violate the
• 2:35
  First Amendment. But the most aggressive
• 2:37
  mechanism is financial. The administration is
• 2:40
  threatening to withhold broadband equity access and
• 2:42
  deployment funds. Wait, the broadband money?
• 2:45
  Yeah, we're talking about $21 billion in undispersed
• 2:48
  infrastructure grants designed to expand high speed Internet
• 2:51
  access. They are threatening to withhold
• 2:53
  this money from states that refuse to align with the federal
• 2:56
  deregulatory agenda. Hold on, the federal government
• 2:58
  is using Internet funding to block local tech rules.
• 3:02
  So wait, are you saying they're using Hwy. repair tactics for
• 3:05
  the Internet? That's exactly what it.
• 3:07
  Is like we won't fix your digital roads unless you drop
• 3:10
  your local speed limits. Yeah, exactly.
• 3:12
  You have states desperately relying on this broadband money.
• 3:15
  We are talking about physically trenching dirt to lay fiber
• 3:19
  optic cables, connecting rural schools and hospitals that
• 3:23
  currently have terrible service, and now that physical
• 3:27
  infrastructure money is completely tied to how a state
• 3:30
  chooses to govern software development within its own
• 3:32
  borders. That severely limits local
• 3:35
  governments ability to govern technology.
• 3:37
  But it opens up a streamlined single compliance track for tech
• 3:41
  companies, theoretically allowing them to innovate
• 3:43
  without navigating 50 different rulebooks.
• 3:46
  Which clears the path for American companies to compete
• 3:49
  globally. Look at the international
• 3:51
  pressure. The European Union has adopted
• 3:53
  binding legal frameworks imposing strict obligations on
• 3:56
  high risk models. Enforcing the rule that
• 3:59
  individuals have a right not to be subject to fully automated
• 4:02
  decisions. Right.
• 4:03
  So the federal government wants a frictionless runway here at
• 4:05
  home to compete with that kind of global regulatory power.
• 4:09
  The administration argues that tech startups simply cannot
• 4:11
  function if they have to navigate 50 different regulatory
• 4:14
  regimes. The states are not backing down,
• 4:16
  though, and this is entirely crossing party lines.
• 4:19
  Look at Utah, a state led by Republicans.
• 4:22
  Yeah. Representative Doug Fifia, who
• 4:24
  is actually a former Google employee, proposed A
• 4:27
  transparency bill. And this happened after a
• 4:29
  teenager named Adam used a chatbot to learn how to end his
• 4:33
  life. It did.
• 4:35
  The bill was designed to require developers of large frontier
• 4:38
  models to publish public safety and child protection plans, to
• 4:42
  report safety incidents to the state and to establish
• 4:45
  whistleblower protections. For employees who spot dangerous
• 4:48
  flaws in the code. Right.
• 4:50
  But the White House Office of Intergovernmental Affairs
• 4:52
  intervene directly. They sent A1 line memo to Utah
• 4:55
  lawmakers calling the bill unfixable and stating that it
• 5:00
  opposed the administration's agenda.
• 5:02
  Wait, backup? The federal framework explicitly
• 5:05
  claims to preserve state authority over child safety,
• 5:08
  right? The executive order even carved
• 5:11
  out exceptions, saying federal preemption would not apply to
• 5:14
  child protection laws. That is the exact contradiction.
• 5:17
  The federal administration actively opposed the bill
• 5:20
  directly aimed at child safety and transparency.
• 5:23
  Representative Fifia noted that the White House offered zero
• 5:27
  specific amendments to make the bill acceptable.
• 5:29
  So they didn't even try to fix it.
• 5:31
  They did not want to negotiate the text.
• 5:33
  They wanted total suppression of the state level effort.
• 5:36
  He argued that states have an absolute right to protect their
• 5:39
  constituents, especially when Congress is stuck in gridlock.
• 5:42
  Utah Governor Spencer Cox pushed back as well, explicitly stating
• 5:46
  that regulation should be left to the states.
• 5:48
  This shifts the conflict from a simple political disagreement
• 5:52
  into a fundamental constitutional power struggle
• 5:55
  regarding the 10th Amendment and state sovereignty.
• 5:58
  When the federal government intervenes in a heavily
• 6:00
  conservative state to block a child safety measure simply
• 6:03
  because it touches artificial intelligence, it signals that
• 6:07
  the push for federal primacy overrides traditional party
• 6:10
  platforms regarding states rights.
• 6:12
  Right. So if the federal government is
• 6:14
  stomping out state laws to protect truthful data, what
• 6:17
  happens when a state claims that data is already lying?
• 6:20
  Oh, wow. Yeah, that tension brings us to
• 6:23
  the Colorado AI Act, which is the first comprehensive statute
• 6:26
  in the country targeting high risk systems.
• 6:29
  OK, This law requires developers and deployers to use reasonable
• 6:32
  care to prevent algorithmic discrimination and disparate
• 6:36
  impact against protected groups. So if a system is making
• 6:39
  consequential decisions like who gets a job, who gets a loan, or
• 6:42
  who gets approved for housing. Exactly.
• 6:44
  The developers must ensure the outcomes do not proportionately
• 6:48
  harm protected classes, regardless of whether that harm
• 6:50
  was intentional. But the White House and AI czar
• 6:53
  David Sachs are specifically targeting this Colorado law.
• 6:57
  Yes, they are. Sachs has publicly argued that
• 7:00
  state's regulatory frenzy is encouraging what he calls woke
• 7:04
  AI. The administration argues that
• 7:06
  forcing developers to alter raw data outputs to achieve equity
• 7:10
  introduces ideological bias. They categorize this as
• 7:14
  deceptive under Federal Trade Commission rules.
• 7:17
  We have to look at both sides of this carefully.
• 7:19
  The state view is entirely focused on protecting citizens
• 7:22
  from baked in bias. They argue that historical data
• 7:25
  is naturally flawed. Right, like if a resume
• 7:27
  screening tool is trained on 20 years of hiring data from a male
• 7:31
  dominated industry. The math will naturally penalize
• 7:33
  female applicants because it associates success with male
• 7:37
  profiles. Colorado argues that if you do
• 7:39
  not actively correct for that discrimination, the software
• 7:42
  just automates past prejudices. But the federal view is focused
• 7:45
  on preventing the forced manipulation of raw data.
• 7:48
  They argue that if you force a mathematical model to alter its
• 7:51
  truthful statistical outputs just to avoid a differential
• 7:54
  impact, you're forcing the system to produce false results.
• 7:57
  To the federal government, applying artificial weights to
• 8:01
  achieve a specific social outcome is inherently deceptive
• 8:05
  to the consumer using the product.
• 8:07
  He's like, the technology is just a mirror, right?
• 8:09
  Yeah. Do you want the mirror to show
• 8:11
  exactly what is there, even if the reflection contains
• 8:14
  historical societal flaws? Or do you adjust the mirror so
• 8:18
  it reflects more balanced equitable picture?
• 8:21
  Exactly. And the federal government say
• 8:23
  don't touch the mirror. If society is ugly or uneven,
• 8:26
  the mirror has to show exactly what is there.
• 8:28
  But Colorado is saying no. This mirror is being used to
• 8:31
  approve mortgages. If the reflection is warped by
• 8:33
  historical redlining, you have a legal obligation to straighten
• 8:37
  the glass so it reflects A balanced, equitable picture.
• 8:39
  Which limits states tools to fight algorithmic discrimination
• 8:43
  while opening up a massive legal debate over who defines truthful
• 8:47
  data. The executive order explicitly
• 8:50
  directs the Federal Trade Commission to issue a policy
• 8:53
  statement clarifying when state laws mandating alterations to
• 8:56
  truthful outputs cross the line into unfair and deceptive acts.
• 9:00
  If the Federal Trade Commission officially issues guidance
• 9:03
  classifying bias mitigation as a deceptive trade practice, it
• 9:06
  creates a direct legal collision.
• 9:09
  Developers will be caught in an impossible position.
• 9:12
  Violate state law by failing to mitigate bias or violate federal
• 9:17
  directives by allegedly altering truthful outputs.
• 9:20
  We have two completely different visions for the rules of the
• 9:23
  road. The White House released a
• 9:25
  legislative blueprint that is a very light touch policy
• 9:29
  framework. Yeah, they want to rely on
• 9:30
  existing agencies and industry LED standards.
• 9:34
  They explicitly ask Congress not to create a new federal
• 9:37
  rulemaking body. Their plan focuses heavily on
• 9:40
  infrastructure and economic security.
• 9:43
  For instance, they urge Congress to codify the ratepayer
• 9:46
  protection pledge. Right.
• 9:47
  What does a ratepayer pledge have to do with language models?
• 9:50
  It comes down to the physical electricity required to run
• 9:53
  these systems. Massive new data centers consume
• 9:56
  staggering amounts of power, which strains local electrical
• 9:59
  grids. Oh, I see.
• 10:00
  The pledge would shield residential consumers from the
• 10:03
  soaring electricity costs driven by those power demands.
• 10:07
  The administration also wants to streamline federal permitting
• 10:10
  for these facilities so they can be built faster.
• 10:13
  Ensuring the underlying infrastructure is there to
• 10:15
  support the computing power needed.
• 10:18
  Well, in response to that framework, Senator Marsha
• 10:21
  Blackburn introduced a massive 291 page bill called the Trump
• 10:26
  America AI Act. And the strictness of
• 10:28
  Blackburn's bill is genuinely surprising given the broader
• 10:32
  deregulatory push. It is highly prescriptive.
• 10:34
  It establishes a federal duty of care for chatbot developers.
• 10:38
  In legal terms, that means developers would be legally
• 10:41
  required to exercise reasonable care to prevent foreseeable
• 10:44
  harms. If a user gets hurt because a
• 10:47
  chatbot gave them dangerous instructions, the developer
• 10:50
  could be held liable in court. The bill also mandates annual
• 10:53
  independent audits for high risk systems, specifically to detect
• 10:56
  political viewpoint and affiliation discrimination.
• 10:59
  It does. It also includes the End No
• 11:01
  Fakes Act to protect individuals from unauthorized digital
• 11:05
  replicas of their voice or likeness, and it tackles labor
• 11:09
  impacts. Directly covered entities,
• 11:11
  including publicly traded companies and federal agencies,
• 11:14
  would be required to submit quarterly reports to the
• 11:17
  Department of Labor detailing layoffs, hiring and retraining
• 11:20
  attributable to artificial intelligence adoption.
• 11:23
  But most aggressively, it pushes for a repeal of Section 230 of
• 11:27
  the Communications Decency Act for Artificial intelligence.
• 11:30
  I hear Section 230 brought up constantly in tech policy.
• 11:33
  What does repealing it actually do to the average user or
• 11:37
  developer? Section 230 is a law from 1996.
• 11:40
  The base says an online platform is not legally responsible for
• 11:44
  the content its users post. It treats platforms like a
• 11:47
  newsstand rather than a publisher.
• 11:49
  If you run a social media site and someone post something
• 11:51
  defamatory, the person who posted it is liable, not the
• 11:54
  site itself. If Congress repeals that
• 11:56
  protection for artificial intelligence, it completely
• 11:59
  eliminates that long standing legal immunity.
• 12:01
  So that chatbot you use to write emails at work?
• 12:04
  The company that made it suddenly becomes legally
• 12:07
  responsible for every single typo, hallucinated fact, or
• 12:11
  copyright infringement it generates.
• 12:12
  Exactly. Now, meanwhile, Democrats in
• 12:15
  Congress introduced the Guardrails Act, right?
• 12:18
  Their bill is designed to repeal the executive order entirely and
• 12:22
  protect state power. Representatives driving this
• 12:25
  bill argue that a federal moratorium on state regulations
• 12:29
  would create a lawless environment.
• 12:31
  Exposing the public to unchecked risks without setting any
• 12:33
  acceptable federal standards in their place.
• 12:36
  They view the administration's strategy as an attempt to kill
• 12:39
  state safeguards while leaving a regulatory vacuum.
• 12:43
  So Washington is completely divided on exactly how to
• 12:45
  preempt the states. This opens up the possibility of
• 12:48
  a massive new federal liability regime that completely redefines
• 12:52
  the legal risks for tech developers.
• 12:54
  Between the proposed duty of care, the strength liability for
• 12:58
  defective systems, and the potential repeal of Section 230,
• 13:02
  companies could face a drastically altered legal
• 13:05
  environment where they are directly responsible for the
• 13:07
  outputs and downstream uses of their software.
• 13:10
  A single federal statute could rewrite the entire business
• 13:14
  model of the Internet. But we need to look at the on
• 13:17
  the ground reality for businesses right now.
• 13:19
  An executive order cannot automatically void a state law.
• 13:22
  No. Only an act of Congress or
• 13:24
  federal court ruling can do that.
• 13:26
  Therefore, every single state tech law remains fully
• 13:29
  enforceable until a court formally strikes it down.
• 13:33
  And state attorneys general are actively hunting for violations.
• 13:36
  We have seen a 42 state attorney general coalition send a joint
• 13:40
  letter demanding additional safeguards, specifically warning
• 13:43
  about outputs that harm children.
• 13:45
  They are not just going after the tech giants who build the
• 13:47
  foundational models, right? Right.
• 13:49
  They are going after traditional businesses deploying these
• 13:52
  tools. The message from the states is
• 13:54
  clear. Claiming you just bought the
• 13:56
  software from a vendor is not a valid defense if that software
• 14:00
  produces harmful outcomes for your customers.
• 14:02
  If you are working at a company deploying these systems, you
• 14:05
  cannot wait for the courts to figure this out.
• 14:07
  You have laws taking effect right now.
• 14:09
  Absolutely. California requires training
• 14:12
  data transparency. That means developers have to
• 14:14
  publish exact documentation on what data types they use to
• 14:18
  train their models and explicitly state whether
• 14:21
  copyrighted material is included in that data.
• 14:24
  Texas passed the Responsible AI Governance Act, prohibiting
• 14:28
  systems designed for restricted purposes like unlawful
• 14:31
  discrimination. Illinois mandates explicit
• 14:34
  notice when automated systems are used in employment
• 14:36
  decisions. If your company uses a tool to
• 14:38
  scan resumes or evaluate video interviews, you have to tell the
• 14:42
  applicant. Furthermore, cyber insurance
• 14:44
  carriers are already requiring specific security controls.
• 14:48
  They are adding security writers to their policies.
• 14:50
  If you do not have documented risk management practices, you
• 14:53
  face coverage denials or prohibitive premiums.
• 14:56
  And the technology is creating new vulnerabilities.
• 15:00
  Threat actors are leveraging these systems to orchestrate
• 15:02
  attacks at unprecedented speeds. Identity based attacks are
• 15:06
  rising sharply. Employees are creating severe
• 15:09
  risks through shadow artificial intelligence too.
• 15:12
  This is when an employee uses unsanctioned third party
• 15:16
  generative tools to do their daily work.
• 15:18
  Like they might copy and paste a confidential client contract
• 15:21
  into a public chat bot to summarize it.
• 15:24
  Exactly which inadvertently leaks sensitive corporate data
• 15:28
  right into the public training models.
• 15:30
  Organizations have to look at federal agency enforcement.
• 15:33
  That is already happening under existing laws too.
• 15:35
  The Securities and Exchange Commission is actively targeting
• 15:38
  AI washing. This happens when companies make
• 15:40
  misleading or exaggerated claims about their technological
• 15:43
  capabilities to attract investment, slapping the label
• 15:46
  on basic automation just to boost their stock price.
• 15:49
  The Equal Employment Opportunity Commission is monitoring how
• 15:52
  these tools are used in hiring and promotion to ensure they do
• 15:55
  not violate existing civil rights protections.
• 15:59
  The Food and Drug Administration is actively regulating software
• 16:02
  as a medical device. This limits the immediate
• 16:04
  practical impact of the federal preemption push, forcing
• 16:08
  organizations to build massive compliance programs based on the
• 16:11
  strictest state laws anyway. A company cannot simply ignore
• 16:15
  Colorado or California's requirements based on the hope
• 16:19
  that a Department of Justice task force might eventually
• 16:22
  succeed in striking them down in federal court.
• 16:25
  The compliance architecture has to be built.
• 16:26
  Today, some state laws actually provide safe harbors.
• 16:30
  For instance, compliance with the voluntary National Institute
• 16:33
  of Standards and Technology Risk Management Framework provides an
• 16:37
  affirmative defense under the new Texas law.
• 16:39
  This means the localized rules and voluntary frameworks are
• 16:43
  currently dictating corporate behavior across the entire
• 16:45
  country, regardless of the executive orders coming out of
• 16:48
  Washington. The battle for control over
• 16:50
  artificial intelligence is currently a constitutional tug
• 16:53
  of war. The federal government is trying
• 16:55
  to clear the runway for total global dominance, while states
• 16:58
  are scrambling to build protective barriers for their
• 17:01
  citizens. If artificial intelligence truly
• 17:03
  is a borderless technology, can a system built on localized
• 17:08
  state by state laws ever really govern it?
• 17:10
  Or does it inherently require a single global rulebook?
• 17:14
  If you're not subscribed yet, take a second and hit follow on
• 17:16
  whatever app you're using. It helps us keep making this.
• 17:18
  We appreciate you being here.