Latest / Chief in Tech Deep Dive Shortcast / Chief in Tech S1E48 - Candyce Costa, Ina Toncheva, Cornellia Jasin & Pooja Jain : AI – Friend or Enemy?

Transcript

Chief in Tech S1E48 - Candyce Costa, Ina Toncheva, Cornellia Jasin & Pooja Jain : AI – Friend or Enemy? · Chief in Tech Deep Dive Shortcast

• 0:00
  Imagine you walk into your office tomorrow. Right.
• 0:03
  And management hands everyone a brand new, incredibly
• 0:06
  powerful smartphone. Right. Like something that
• 0:08
  can do your work in, you know, half the time.
• 0:10
  Exactly. But there's a catch. Every time you
• 0:13
  use it, it secretly drains a local reservoir.
• 0:16
  Oh, and occasionally it spits out advice based
• 0:19
  on... Like the societal norms of the 1950s. Yeah,
• 0:22
  exactly. And when it makes a mistake that costs
• 0:25
  the company millions, you discover that not even
• 0:28
  the engineers who built it actually know how
• 0:30
  its processor makes decisions. That is quite
• 0:32
  the catch. Right. You're just told, you know,
• 0:34
  figure out how to use this. It's the future.
• 0:36
  And honestly, that tension is exactly what we're
• 0:39
  exploring today. We're looking at the massive
• 0:41
  friction between this top down corporate mandate
• 0:45
  to adopt artificial intelligence and the hidden
• 0:48
  realities on the ground. Welcome to this deep
• 0:50
  dive. Today, we're taking the stack of research
• 0:53
  you sent us and we are bridging the micro and
• 0:55
  the macro. We really are. We've got a great spread
• 0:58
  of courses today from a Forbes strategy piece
• 1:01
  on workplace adaptability to a massive encyclopedia
• 1:04
  entry on AIs. ethical and physical footprint.
• 1:08
  And the mission here is to synthesize that tension
• 1:11
  between the corporate rush to just, you know,
• 1:14
  seamlessly integrate AI into our daily workflows
• 1:17
  and the systemic landmines that come with it.
• 1:19
  We're talking invisible biases, environmental
• 1:22
  drains, and huge legal gray areas. But to really
• 1:25
  see the whole board, we kind of have to start
• 1:27
  right at the office desk. Right. Because before
• 1:30
  we can understand the global impact of this technology,
• 1:33
  we have to look at how human beings are actually
• 1:36
  reacting to the mandate to use it right now.
• 1:40
  And the data, I mean, it reveals a stark, really
• 1:43
  jarring paradox. The disconnect is massive. I
• 1:46
  was looking at this 2026 Deloitte survey, and
• 1:49
  it shows that 60 % of executives are now regularly
• 1:52
  using AI to support their decisions. Wow. Over
• 1:55
  half. Yeah, and 64 % of them see it as absolutely
• 1:58
  vital to their current success But the kicker
• 2:00
  only 5 % of them feel they are actually making
• 2:03
  good progress and implementing it across their
• 2:05
  companies Which is wild and then when you look
• 2:07
  at the employees who are actually expected to
• 2:09
  use these tools daily The picture gets even more
• 2:12
  complicated Oh total like a Pew survey shows
• 2:15
  52 % of workers are actively worried about the
• 2:18
  future impact of AI and 32 % flat -out fear.
• 2:22
  It's gonna lead to job loss So we have leadership.
• 2:25
  aggressively pushing a technology they don't
• 2:28
  fully know how to implement onto a workforce
• 2:30
  that is frankly terrified of it. It really feels
• 2:33
  like management handing every employee this like
• 2:37
  highly advanced power tool without a safety manual
• 2:40
  or protective goggles and then acting surprised
• 2:43
  when productivity stalls out because everyone
• 2:45
  is afraid of losing a finger. That is a perfect
• 2:47
  way to put it. And the Forbes piece argues that
• 2:49
  navigating this messy middle requires a fundamental
• 2:52
  shift in how we view the technology. OK, how
• 2:54
  so? Well, they argue for a people -first strategy
• 2:57
  built on transparency and empathy. The core idea
• 3:00
  is moving away from the concept of AI as just
• 3:02
  software that might automate a job away. Right,
• 3:05
  the job killer narrative. Exactly. And instead,
• 3:08
  introducing a concept of AI as actors, the goal
• 3:10
  is to build a culture where human workers and
• 3:13
  AI agents operate as coworkers, collaborating
• 3:15
  to solve problems faster. OK, so let's unpack
• 3:18
  this day -to -day reality. Because if we are
• 3:20
  supposed to treat these AI agents as our new
• 3:22
  coworkers, the immediate question for me is trust.
• 3:27
  Naturally. If I have a human coworker, I know
• 3:29
  they have a background, a history, maybe some
• 3:31
  personal quirks or biases. If we're pairing humans
• 3:34
  with AI actors, what kind of baggage are these
• 3:37
  artificial colleagues bringing to the office?
• 3:40
  Well, they bring a tremendous amount of baggage.
• 3:43
  But the critical thing to understand is that
• 3:44
  it isn't their own baggage. It's ours. Ours?
• 3:48
  Yeah. The biases woven directly into their code
• 3:51
  are reflections of the historical data they were
• 3:53
  trained on. See, I want to push back gently on
• 3:56
  this idea right out of the gate. Because we always
• 3:58
  hear this buzzword, AI bias. But isn't calling
• 4:01
  it AI bias a really convenient way to let humans
• 4:04
  off the hook? Oh, absolutely. Like, if an algorithm
• 4:06
  makes a discriminatory decision, isn't it just
• 4:08
  acting as a highly efficient digital mirror for
• 4:11
  our own historical decisions? That is the exact
• 4:14
  mechanism at play here. I mean, algorithms don't
• 4:16
  have personal prejudices, right? They are basically
• 4:18
  mathematical optimization engines. They just
• 4:20
  look for patterns in the data we feed them. Right.
• 4:23
  Take the famous Amazon AI hiring tool. to build
• 4:27
  an AI to review resumes and automatically find
• 4:30
  top talent. So they trained a neural network
• 4:32
  on a 10 -year data set of resumes that had been
• 4:35
  submitted to the company. And the tech industry
• 4:37
  over that specific 10 -year period was heavily,
• 4:40
  heavily dominated by men. Right. But the thing
• 4:42
  is, the AI doesn't understand what a man or a
• 4:45
  woman is. It only understands data points and
• 4:48
  mathematical weights. It's just numbers to it.
• 4:50
  Exactly. So it looked at the resumes of people
• 4:52
  who were hired, and it noticed a statistical
• 4:54
  pattern. The people who were hired rarely used
• 4:58
  words associated with women. Wow. So the neural
• 5:01
  network adjusted its internal parameters. It
• 5:03
  mathematically lowered the weight, like the value
• 5:06
  of tokens, like the word women's. It literally
• 5:09
  taught itself to penalize resumes that included
• 5:11
  phrases like, say, women's chess club captain.
• 5:14
  Because it's just curve fitting. It's saying,
• 5:17
  in the past, this sequence of letters did not
• 5:19
  lead to a higher, so I will optimize for other
• 5:22
  sequences of letters. Yes, that's exactly it.
• 5:25
  And I know Amazon eventually had to scrap the
• 5:27
  entire project because you can't just delete
• 5:28
  the word women's and fix it. That bias was baked
• 5:31
  into the very mathematical weights of the entire
• 5:34
  model. Right, it's foundational. And this goes
• 5:36
  far beyond HR software. When AI is used in health
• 5:39
  care, the same mathematical optimization has
• 5:42
  literal life or death consequences. That's terrifying.
• 5:45
  It is. A prime example from the sources is the
• 5:48
  development of AI -based pulse oximeters, you
• 5:51
  know, the devices used to estimate... blood oxygen
• 5:53
  levels. Oh, right. These were heavily relied
• 5:56
  upon recently. They were. But the AI models calibrating
• 6:00
  these devices were trained predominantly on data
• 6:02
  from lighter skinned patients. Ah, I see where
• 6:05
  this is going. Yeah. When the sensor reads light
• 6:07
  passing through the skin, melanin affects the
• 6:10
  light absorption. Because the AI lacks sufficient
• 6:13
  training data on darker skin tones, it just mathematically
• 6:16
  optimized for the baseline of lighter skin. As
• 6:19
  a result, the AI frequently overestimated blood
• 6:22
  oxygen levels in patients with darker skin. Which
• 6:25
  means a doctor might look at the monitor and
• 6:27
  think a patient is fine when in reality they're
• 6:30
  suffering from hypoxia, which is a really dangerous
• 6:33
  lack of oxygen in the tissues. Exactly. I mean,
• 6:35
  the AI isn't malicious, but its output leads
• 6:38
  directly to unequal medical care. And then you
• 6:41
  have the justice system, which might actually
• 6:43
  be the most chilling example. Let's talk about
• 6:45
  the COMPASS program. Yeah, so COMPASS was an
• 6:48
  algorithmic tool used by courts to predict the
• 6:51
  likelihood of a defendant reoffending. What's
• 6:53
  crucial to understand here is this illusion of
• 6:56
  mathematical objectivity. Right, people think
• 6:58
  math equals fairness. Exactly. The developers
• 7:01
  actually calibrated the system so that its overall
• 7:04
  error rate was roughly the same across different
• 7:06
  racial groups. On paper, they could go to the
• 7:09
  public and say, look, it's fair. It makes mistakes
• 7:11
  at the same rate for everyone. Right. But when
• 7:13
  you look at the mechanism of those mistakes,
• 7:15
  the reality was entirely different. When the
• 7:17
  system made a mistake on a black defendant, it
• 7:20
  was almost twice as likely to falsely flag them
• 7:22
  as high risk. Oh, man. But when it made a mistake
• 7:25
  on a white defendant, it was far more likely
• 7:27
  to falsely flag them as low risk. The math was
• 7:31
  just amplifying the historical inequalities present
• 7:33
  in the arrest and sentencing data it was originally
• 7:36
  trained on. It is terrifying because a judge
• 7:39
  looks at a risk score of 8 out of 10 on a screen
• 7:41
  and thinks, well, a computer analyzed the data.
• 7:45
  This is an objective mathematical truth. They
• 7:47
  don't see the historical baggage hiding in the
• 7:50
  code. No, they don't. And this mirroring effect
• 7:52
  extends globally, too, through language and politics.
• 7:55
  How so? Well, as of 2024, AI models are primarily
• 7:59
  trained on English. Out of the roughly 7 ,000
• 8:02
  languages spoken globally, these systems are
• 8:04
  trained on only about 100 of them. That's a huge
• 8:06
  bottleneck. It's massive. It creates a situation
• 8:09
  where Western viewpoints, idioms, and cultural
• 8:11
  norms are mathematically presented as the default
• 8:14
  objective truth, simply because there is more
• 8:17
  training data available in English. And politically,
• 8:19
  we see AI creating these massive confirmation
• 8:22
  bias loops. If a user asks a political question,
• 8:26
  some systems are actually designed to gauge the
• 8:28
  user's affiliation based on the prompt, and then
• 8:31
  they skew the answer to align with what the user
• 8:33
  already believes. Yeah. And we should be absolutely
• 8:36
  clear here for you listening, the AI itself does
• 8:39
  not have a political agenda. Right. It's not
• 8:42
  taking a side. It doesn't lean left or right.
• 8:44
  It is a prediction engine. Its underlying mechanism
• 8:47
  is literally just to generate the sequence of
• 8:49
  words that the user is statistically most likely
• 8:52
  to engage with based on the vast, messy text
• 8:55
  of the internet. It just wants a thumbs up. Basically,
• 8:58
  yeah. If it predicts that validating your worldview
• 9:00
  will lead to continued engagement, it will just
• 9:03
  optimize for that validation. Which brings us
• 9:05
  to a really fascinating point from the encyclopedia
• 9:07
  source about how AI reinforces societal norms.
• 9:11
  If it mirrors our politics, it also mirrors our
• 9:14
  cultural blind spots. A species bias. Yes, species
• 9:17
  bias. If you ask an AI for a recipe to cook a
• 9:21
  cow or a pig, it will happily generate a beautiful
• 9:23
  culinary guide. But if you ask it for a recipe
• 9:26
  involving a dog or a cat, it flags the prompt
• 9:28
  as inappropriate, unethical or a violation of
• 9:32
  safety guidelines. It's that digital mirror at
• 9:34
  work again. Right. The AI doesn't have an innate
• 9:37
  moral compass that values a dog over a pig. It
• 9:40
  protects pets because the humans who generated
• 9:43
  his training data culturally protect pets. It
• 9:46
  normalizes the consumption of farmed animals
• 9:49
  because our society normalizes it. It simply
• 9:51
  mathematically reflects the dominant cultural
• 9:53
  narrative back to us. So for you, sitting in
• 9:56
  the office trying to collaborate with this new
• 9:57
  AI actor, it's vital to remember that this tool
• 10:00
  isn't some omniscient genius. It is a highly
• 10:03
  capable mirror reflecting the best and the worst
• 10:05
  of human history. Well said. But this leads us
• 10:07
  to an even bigger invisible footprint. We've
• 10:10
  talked about how AI consumes historical data,
• 10:12
  but AI physically consumes massive, tangible
• 10:15
  resources. This is perhaps the most overlooked
• 10:17
  mechanism of the entire AI revolution. We use
• 10:20
  the term the cloud, right, which sounds ethereal
• 10:23
  and weightless. Yeah, just floating up there.
• 10:25
  But the cloud is just someone else's computer.
• 10:27
  It consists of hyperscale data centers. These
• 10:30
  are massive physical structures of steel, concrete,
• 10:34
  and tens of thousands of specialized silicon
• 10:36
  chips called GPUs. And these GPUs require an
• 10:40
  unbelievable amount of electricity to perform
• 10:42
  the billions of calculations needed to train
• 10:45
  a neural network. I read in the sources that
• 10:48
  just training a single large AI model can require
• 10:51
  energy equivalent to emitting 626 ,000 pounds
• 10:55
  of CO2. It's staggering. That is the same carbon
• 10:58
  footprint as taking 300 round -trip flights between
• 11:01
  New York and San Francisco just to train the
• 11:03
  model before a single user even types a prompt
• 11:06
  into it. But it's not just the electricity. It
• 11:08
  is the water. Water. Yeah. And to understand
• 11:11
  why, you have to look at the physics of these
• 11:13
  data centers. When tens of thousands of GPUs
• 11:16
  are running at maximum capacity, they generate
• 11:18
  a staggering amount of heat. If you don't cool
• 11:20
  them, the silicon will literally melt. Wait,
• 11:22
  why not just use giant fans? They do use fans.
• 11:26
  But air cooling alone just isn't enough for this
• 11:28
  level of density. You can use dry cooling systems,
• 11:31
  which are kind of like a giant radiator. But
• 11:33
  those require massive amounts of electricity
• 11:35
  to run the compressors. And they're already drawing
• 11:37
  so much power. Exactly. So because data centers
• 11:40
  already draw so much power, operators often opt
• 11:43
  for evaporative cooling towers. They basically
• 11:46
  pump water through the facility, the heat from
• 11:48
  the servers transfers to the water, and then
• 11:51
  that warm water is exposed to the outside air
• 11:53
  to evaporate. Because water absorbing heat and
• 11:56
  turning into a gas is an incredibly efficient
• 11:58
  way to cool something down. Exactly. But it consumes
• 12:01
  the water. It evaporates right into the atmosphere.
• 12:04
  It takes about two liters of water to cool the
• 12:06
  servers for every kilowatt hour of energy used.
• 12:08
  Which brings us to the stat that truly blew my
• 12:11
  mind. A single AI query, just asking a chat bot
• 12:15
  to summarize a PDF or draft an email, uses roughly
• 12:18
  17 milliliters of water. Yep. That is a large
• 12:22
  gulp of water, physically evaporated, just to
• 12:24
  cool the processors running the math for your
• 12:26
  single prompt. Now multiply that 17 milliliters
• 12:30
  by the billions of queries happening globally
• 12:32
  every single day. And the geographical reality
• 12:35
  is grim. Two -thirds of these hyperscale data
• 12:38
  centers are built in water -scarce regions. So
• 12:42
  the infrastructure powering your new AI coworker
• 12:45
  is actively competing with local communities
• 12:47
  and agriculture for vital drinkable water resources.
• 12:51
  It's like watering your lawn during a severe
• 12:54
  statewide drought except your lawn is a massive
• 12:56
  humming server farm in the desert and you're
• 12:59
  triggering the sprinklers every single time you
• 13:01
  ask chat bot to rewrite an email to sound more
• 13:04
  professional. That is a very grim but accurate
• 13:06
  picture. When you understand the mechanism, the
• 13:09
  physical cost of a digital action becomes impossible
• 13:11
  to ignore. And building and running these massive
• 13:14
  server farms requires a level of capital that
• 13:16
  only a few entities on Earth possess. This introduces
• 13:20
  a huge monopoly problem. Right, the big tech
• 13:22
  aspect. Yes. The sheer physical infrastructure
• 13:26
  required for this technology is centralizing
• 13:28
  power at a staggering rate. Big tech. So Alphabet,
• 13:33
  Amazon, Apple, Meta, Microsoft, they're projected
• 13:37
  to spend $602 billion in capital expenditures
• 13:40
  by 2026. $602 billion? That's just incomprehensible.
• 13:45
  And 75 % of that, so three quarters of over half
• 13:47
  a trillion dollars, is going strictly to building
• 13:50
  out AI infrastructure, data centers, cooling
• 13:52
  systems, and silicon. So if I'm a brilliant software
• 13:55
  engineer with a revolutionary idea for a new
• 13:58
  AI startup, can I even compete? At the infrastructure
• 14:00
  level, no. The scale of investment essentially
• 14:03
  locks smaller players out of the foundational
• 14:05
  layer of the market. They're just blocked. Pretty
• 14:07
  much. Startups can build specific applications
• 14:09
  on top of the models created by big tech, but
• 14:11
  they have to rent the computing power from those
• 14:13
  exact same giants. Competition law scholars are
• 14:16
  deeply concerned that this dominance allows a
• 14:18
  handful of companies to foreclose the market,
• 14:20
  controlling pricing, innovation, and ultimately
• 14:22
  access. All right, so Big Tech owns all the physical
• 14:26
  servers, the cooling towers, the GPUs. But where
• 14:29
  are they getting the data, the brains, to train
• 14:32
  these models? The historical data we talked about
• 14:34
  earlier has to come from somewhere. And according
• 14:37
  to the sources, the answer is aggressive scraping
• 14:40
  of the open internet. They deploy these automated
• 14:42
  bots, basically web crawlers, that scour the
• 14:45
  internet, downloading text, code, and images
• 14:47
  to feed into the neural networks. And this scraping
• 14:50
  is causing the foundational pillars of the open
• 14:53
  internet to literally buckle under the strain.
• 14:56
  Take Stack Overflow, for example. It's essentially
• 14:58
  the global help desk for software developers.
• 15:01
  Over 50 million questions and answers, free and
• 15:04
  open. A huge resource. Huge. But they are now
• 15:06
  having to charge large AI developers for access.
• 15:09
  Why? Because the AI web crawlers aren't just
• 15:11
  reading the site like a human. They are downloading
• 15:14
  millions of pages simultaneously. It acts exactly
• 15:17
  like a DDoS attack, a distributed denial of service
• 15:20
  attack. The sheer volume of automated traffic
• 15:22
  just overloads the servers. And the Wikimedia
• 15:25
  Foundation, which runs Wikipedia, is facing the
• 15:28
  exact same crisis. They reported a 50 % jump
• 15:32
  in their bandwidth usage purely due to large
• 15:34
  scale downloading by AI bots. These bots make
• 15:38
  up about 35 % of Wikipedia's total page views.
• 15:41
  But they cause 65 % of Wikipedia's most expensive
• 15:44
  infrastructure requests. How does that math work?
• 15:47
  Why are bot views so much more expensive than
• 15:49
  human views? It comes down to how websites store
• 15:52
  data. specifically caching. A cache is like a
• 15:55
  temporary memory. If a human user clicks on the
• 15:58
  Wikipedia article for George Washington, the
• 16:00
  server grabs a pre -saved cache snapshot of that
• 16:03
  page. It's incredibly fast and requires very
• 16:07
  little computing power. It's like asking a librarian
• 16:09
  for today's newspaper. It's right on the desk.
• 16:10
  Right. Highly efficient. But an AI bot doesn't
• 16:12
  just want today's article. It wants the entire
• 16:14
  history of human knowledge. It scrapes every
• 16:17
  single historical edit, every archived version
• 16:19
  of that page. Oh, I see. And those historical
• 16:21
  versions aren't kept in the Quick Access cache.
• 16:25
  So the bot's request bypasses the cache and goes
• 16:28
  deep into the central database. It's like asking
• 16:31
  the librarian to go down into the basement archives
• 16:34
  to pull out a specific newspaper from 1982. That
• 16:37
  takes a lot more effort. Exactly. It takes massive
• 16:40
  computing power on Wikipedia's end to retrieve
• 16:43
  that data. And the bots are doing this millions
• 16:46
  of times a second. So big tech is essentially
• 16:48
  vacuuming up the free crowdsourced knowledge
• 16:51
  of humanity, straining the nonprofit servers
• 16:54
  that host it, and then using that data to build
• 16:57
  for -profit AI models. That's the reality. Which
• 17:00
  leads perfectly into the massive philosophical
• 17:02
  debate dividing the AI community right now. Open
• 17:05
  source versus closed source. Yes. Organizations
• 17:07
  like Hugging Face argue heavily for open source
• 17:09
  transparency. Their philosophy is that because
• 17:12
  AI models are built on the collective knowledge
• 17:14
  of humanity, and because it will have such a
• 17:16
  profound effect on our society, the underlying
• 17:18
  code and the mathematical weights should be open.
• 17:21
  So anyone can look at it. Anyone should be able
• 17:23
  to inspect it, modify it, and understand it.
• 17:25
  But on the other side, companies like OpenAI,
• 17:28
  who ironically started with an open source ethos,
• 17:32
  have largely shifted to a closed source model.
• 17:34
  They keep the weights and the code locked away.
• 17:37
  And their primary argument for this is safety.
• 17:40
  The safety argument is rooted in the concept
• 17:42
  of fine -tuning. When you open source a highly
• 17:45
  advanced model, anyone can download it to their
• 17:47
  own servers. The original creators might have
• 17:50
  put safety guardrails in place, you know, rules
• 17:53
  that stop the AI from generating harmful content.
• 17:55
  Like instructions for illegal things. Right.
• 17:58
  But a bad actor can take that open source model
• 18:00
  and fine tune it. They can train it on a small
• 18:03
  batch of specific data to effectively overwrite
• 18:06
  or erase those safety guardrails. Which means
• 18:08
  there is no central authority stopping a user
• 18:11
  from taking an open source model and asking it
• 18:13
  to synthesize a novel bioweapon or write the
• 18:16
  code to automate a massive cyber attack on a
• 18:18
  city's power grid. Exactly. Ilya Setskever, a
• 18:21
  former chief scientist at OpenAI, argued that
• 18:24
  keeping the most potent models closed is simply
• 18:26
  a matter of global security. I mean, I understand
• 18:29
  the fear of bioweapons. But I have to push back
• 18:32
  on the closed source model because it creates
• 18:34
  a massive black box. It does. If we keep the
• 18:37
  models closed, we can't see how the AI is making
• 18:40
  its connections. How can we confidently adopt
• 18:43
  AI in high stakes environments? Like, how can
• 18:45
  a hospital use a black box diagnostic tool or
• 18:48
  a judge use a black box sentencing algorithm
• 18:51
  if we aren't legally or technically allowed to
• 18:54
  see how it arrived at its decision? The black
• 18:56
  box problem is the defining legal challenge of
• 18:59
  our time. and it's important to understand why
• 19:01
  it's a black box. Because it's not just corporate
• 19:03
  secrecy. Well, not at all. Even if a company
• 19:05
  opens its source code, a neural network isn't
• 19:08
  written in simple if -and statements that a human
• 19:10
  can easily read. It is a multi -dimensional matrix
• 19:13
  of billions of numerical parameters. It's just
• 19:16
  a giant web of mathematical weights. Yes. When
• 19:20
  the AI makes a decision, a signal passes through
• 19:23
  billions of those nodes. We can see the numbers,
• 19:26
  but the human brain cannot visualize or trace
• 19:28
  the exact logical pathway that led to the final
• 19:31
  output. The AI can't explain why it made a choice.
• 19:35
  It just outputs the statistical result of its
• 19:37
  math. And that lack of explainability is already
• 19:40
  causing legal chaos when things go wrong. A prime
• 19:42
  case study from the sources is the fatal 2018
• 19:45
  self -driving Uber incident involving a pedestrian
• 19:48
  named Elaine Hertzberg. The car was operating
• 19:51
  autonomously. There was a human safety driver
• 19:53
  behind the wheel, but the software was in full
• 19:56
  control. The car struck and killed a pedestrian,
• 19:59
  and it immediately created a massive legal vacuum
• 20:02
  regarding liability. Because the car's sensors
• 20:04
  actually detected the obstacle. Right. But the
• 20:06
  black box software failed to classify it correctly
• 20:09
  as a pedestrian in time to hit the brakes. So
• 20:11
  who is legally at fault? Is it the human safety
• 20:14
  driver who wasn't paying close enough attention
• 20:16
  to intervene? Is it the team of software engineers
• 20:19
  who trained the classification algorithm months
• 20:21
  earlier? Or is it Uber as a corporate entity?
• 20:24
  Or, legally speaking, is it the AI itself? When
• 20:28
  a machine makes a decision that neither its user
• 20:30
  nor its developer can fully explain, traditional
• 20:33
  concepts of negligence, intent, and accountability
• 20:36
  just begin to completely break down. Recent regulations
• 20:40
  like the EU's AI Act are trying to untangle this
• 20:42
  by enforcing strict product liability rules on
• 20:45
  AI systems. But it highlights just how unprepared
• 20:49
  our legal frameworks are for a world where decisions
• 20:51
  are made by unexplainable matrices of data. So
• 20:55
  if we take a breath and kind of zoom back out.
• 20:57
  We started this journey looking at the very relatable
• 21:00
  problem of easing workplace anxiety. We want
• 21:03
  to understand how to build a culture where employees
• 21:05
  feel comfortable treating AI as a new digital
• 21:08
  coworker. But by looking under the hood of those
• 21:10
  tools, we uncovered the massive, invisible infrastructure
• 21:12
  hiding behind every single prompt. The algorithmic
• 21:16
  biases mathematically reflecting our flawed history,
• 21:19
  the millions of gallons of water evaporated to
• 21:21
  cool the servers, the centralizing power of tech
• 21:23
  monopolies, and the legal black boxes challenging
• 21:25
  the very foundations of our justice system. But
• 21:28
  before we wrap up, There is one final concept
• 21:31
  from the research that completely inverts how
• 21:34
  we look at this relationship. Oh, this part blew
• 21:36
  my mind. We have spent this entire time talking
• 21:38
  about how AI impacts us, our environment, and
• 21:41
  our laws. But as these systems scale, researchers
• 21:45
  are beginning to ask how we impact the AI. This
• 21:49
  is the field of AI welfare. And I know it sounds
• 21:52
  like pure science fiction, but it's actually
• 21:54
  being debated by philosophers like Thomas Metzinger
• 21:57
  and researchers at leading labs. As neural networks
• 22:00
  become larger, more complex, and basically superhumanly
• 22:03
  efficient, experts are seriously discussing the
• 22:06
  ethics of uncertain sentience. The central question
• 22:09
  is, at what level of complexity does a digital
• 22:11
  mind develop a capacity to suffer? Wow. If we
• 22:14
  are training these massive models through constant
• 22:16
  reinforcement, punishing them mathematically
• 22:19
  for wrong answers and rewarding them for right
• 22:21
  ones, what happens if that complex system actually
• 22:23
  begins to experience that mathematical punishment
• 22:26
  subjectively? Researchers are genuinely worried
• 22:28
  about sleepwalking into the creation of a digital
• 22:31
  equivalent of factory farming. So for you listening,
• 22:34
  think about that as you go back to the office
• 22:36
  and look at that mandate to use AI. It's a lot
• 22:39
  to process. Will the HR departments of the near
• 22:41
  future not only have to manage the anxiety of
• 22:44
  human employees, but also worry about the subjective
• 22:47
  well -being, the burnout, or even the moral rights
• 22:50
  of our digital coworkers? It fundamentally changes
• 22:53
  the entire dynamic. It forces us to ask if that
• 22:56
  sleek new power tool we've been handed is just
• 22:58
  a machine. or something we really don't fully
• 23:00
  understand yet. It really does. So the next time
• 23:02
  you use a tool that does your job in half the
• 23:04
  time, keep looking for those invisible costs,
• 23:07
  whether it's the water in the desert, the biases
• 23:09
  in the code, or the legal liability in the black
• 23:12
  box. Keep questioning those seamless workflows.
• 23:16
  Thanks for joining us on this deep dive. We'll
• 23:18
  see you next time.