Latest / Artificial Intelligence Exposed / The Ethics of Automated Decision-Making

Transcript

The Ethics of Automated Decision-Making · Artificial Intelligence Exposed

• 0:00
  Speaker 1: So if you're sitting there right now thinking that the
• 0:03
  Speaker 1: absolute greatest threat of artificial intelligence is you know, some
• 0:08
  Speaker 1: kind of terminator style robot takeover.
• 0:11
  Speaker 2: Right, the classic sci fi apocalypse exactly.
• 0:13
  Speaker 1: If that's what you're worried about, you are looking in
• 0:16
  Speaker 1: the completely wrong direction. Yeah, entirely, because the real threat,
• 0:19
  Speaker 1: I mean, it already denied your neighbor's mortgage yesterday morning, all.
• 0:23
  Speaker 2: Right, flagged a resume for the trash ben like an
• 0:26
  Speaker 2: hour ago.
• 0:27
  Speaker 1: Right, and it did all of that using perfectly accurate,
• 0:30
  Speaker 1: terrifyingly quiet math. So welcome to the deep dive. Today.
• 0:35
  Speaker 1: We are really getting into the mechanical reality of automated decision.
• 0:39
  Speaker 2: Making, which is such a critical topic right now.
• 0:41
  Speaker 1: It really is. I mean, we are way past the
• 0:43
  Speaker 1: era of asking if AI is going to run our infrastructure.
• 0:46
  Speaker 1: It already does.
• 0:46
  Speaker 2: Yeah, it's everywhere.
• 0:47
  Speaker 1: So the mission for this deep dive, based on the
• 0:50
  Speaker 1: sources where unpacking today, is to figure out how we
• 0:53
  Speaker 1: architect these systems, how do we make sure they don't
• 0:56
  Speaker 1: just you know, optimize away our foundational human rights just
• 0:59
  Speaker 1: because it's computationally efficient.
• 1:02
  Speaker 2: Well, and the friction here it really isn't between humans
• 1:05
  Speaker 2: and sentient machines. That's a distraction, right. The actual friction
• 1:09
  Speaker 2: is between this relentless, hyper scaled technological deployment and honestly,
• 1:14
  Speaker 2: the structural integrity of our society.
• 1:17
  Speaker 1: So it's about how fast we're moving versus how fragile
• 1:20
  Speaker 1: our systems are exactly.
• 1:22
  Speaker 2: I mean, we are building tools that process probabilistic outcomes
• 1:25
  Speaker 2: millions of times faster than a human brain, but we're
• 1:28
  Speaker 2: plugging those tools into super fragile social architectures.
• 1:33
  Speaker 1: Like healthcare diagnostics or criminal justice risk.
• 1:36
  Speaker 2: Assessments, yeah, credit lending, you name it. We are optimizing
• 1:40
  Speaker 2: for speed and scale without actually possessing a stable mathematical
• 1:44
  Speaker 2: definition of fairness.
• 1:46
  Speaker 1: Okay, I want to jump straight into that, that concept
• 1:48
  Speaker 1: of mathematical fairness. Yeah, because I feel like this completely
• 1:51
  Speaker 1: shatters the biggest myth that's holding back the whole conversation.
• 1:54
  Speaker 2: Oh, the myth of objective math.
• 1:55
  Speaker 1: Yes, because look, I think most people who follow tech,
• 1:59
  Speaker 1: you know, they get the algorithms aren't inherently objective. They
• 2:02
  Speaker 1: understand that machine learning models train on historical data, right.
• 2:05
  Speaker 2: The model learns from the past.
• 2:07
  Speaker 1: Right, And if society's history is flawed, the model learns
• 2:12
  Speaker 1: those flaws, but the assumption is still, well, we can
• 2:15
  Speaker 1: just clean.
• 2:16
  Speaker 2: The data right right, Like it's a simple fix.
• 2:19
  Speaker 1: Yeah, people think, okay, I'll just delete the column that
• 2:21
  Speaker 1: says race or the column that says gender, and boom,
• 2:24
  Speaker 1: we have a perfectly objective math equation, which is.
• 2:27
  Speaker 2: Just a fundamental misunderstanding of how high dimensional vector space
• 2:31
  Speaker 2: actually works.
• 2:32
  Speaker 1: Okay, unpacked that Why doesn't deleting the column work well?
• 2:36
  Speaker 2: Because you can delete the race column or the gender
• 2:39
  Speaker 2: column from your data set all you want. The neural
• 2:41
  Speaker 2: network it does not care.
• 2:42
  Speaker 1: It doesn't care, not at all.
• 2:44
  Speaker 2: It will simply reconstruct their protected categories using proxy variables.
• 2:48
  Speaker 1: Okay, proxy variables. Give me an example of how that
• 2:50
  Speaker 1: happens mechanically.
• 2:51
  Speaker 2: Sure, So, a machine learning model is essentially an engine
• 2:54
  Speaker 2: designed to find correlations, right right, So, if you feed
• 2:57
  Speaker 2: it tens of thousands of data points it's about an individual,
• 3:01
  Speaker 2: say their zip code, there, grocery purchasing habits, the specific
• 3:05
  Speaker 2: vocabulary that use in a cover letter.
• 3:07
  Speaker 3: Or like their Internet browsing history, exactly, the model will
• 3:12
  Speaker 3: instantly map the latent correlations between those seemingly harmless data
• 3:16
  Speaker 3: points and the protected classes you tried to hide.
• 3:19
  Speaker 1: Oh, Wow, because human behavior is deeply.
• 3:21
  Speaker 2: Clustered, highly clustered.
• 3:23
  Speaker 1: Yes, So if I know you subscribe to a very
• 3:25
  Speaker 1: specific niche magazine, and you buy a certain brand of
• 3:29
  Speaker 1: hair product, and you live in a specific census tracked.
• 3:33
  Speaker 2: The algorithm can probabilistically determine your demographic profile with terrifying accuracy.
• 3:38
  Speaker 1: And the crazy part is the AI isn't even trying
• 3:41
  Speaker 1: to be prejudiced.
• 3:42
  Speaker 2: No, it has no concept of prejudice. It is simply
• 3:45
  Speaker 2: identifying that a specific cluster of proxy variables historically correlates
• 3:50
  Speaker 2: with whatever success metric you told it to optimize for.
• 3:53
  Speaker 1: That is wild, and it brings us to this mathematical
• 3:55
  Speaker 1: paradox of fairness that the source material really digs into.
• 3:59
  Speaker 2: Yeah, this is where it gets incredibly tricky for developers.
• 4:02
  Speaker 1: Let's say you're engineering a predictive policing algorithm or a
• 4:06
  Speaker 1: risk assessment tool for bail hearings. You obviously want it
• 4:08
  Speaker 1: to be fair.
• 4:09
  Speaker 2: Of course, no one sets out to build a biased.
• 4:11
  Speaker 1: Tool, right, So you sit down with your data scientists
• 4:14
  Speaker 1: and you try to define fairness in the actual code,
• 4:17
  Speaker 1: but the text says you immediately hit an insurmountable mathematical wall.
• 4:22
  Speaker 2: Yes, the conflict between predictive parity and equalized odds.
• 4:26
  Speaker 1: Okay, let's break those down mechanically, because this is the
• 4:29
  Speaker 1: exact moment where the just fix the math argument completely
• 4:34
  Speaker 1: falls apart.
• 4:35
  Speaker 2: It really does. So let's start with predictive parity. This
• 4:38
  Speaker 2: basically means that if the algorithm says someone is high risk,
• 4:41
  Speaker 2: the actual probability of them being high risk should be
• 4:44
  Speaker 2: exactly the same, regardless of their demographic group.
• 4:47
  Speaker 1: Okay, so a seventy percent risk score means a seventy
• 4:50
  Speaker 1: percent chance of reoffending whether the person is in group
• 4:53
  Speaker 1: A or group B exactly.
• 4:55
  Speaker 2: The score means the same thing across the board.
• 4:57
  Speaker 1: Well, I mean that sounds completely reasonable. H It means
• 5:00
  Speaker 1: the score is universally accurate.
• 5:01
  Speaker 2: It does. It sounds great. But then you have the
• 5:03
  Speaker 2: other definition of fairness, which is equalized odds.
• 5:06
  Speaker 1: And this one is specifically looking at false positive rates right.
• 5:10
  Speaker 2: Right, So equalized odds means the algorithm shouldn't falsely flag
• 5:14
  Speaker 2: an innocent person from group A at a higher rate
• 5:17
  Speaker 2: than an innocent person from group B, which.
• 5:19
  Speaker 1: Also sounds entirely necessary for just system. I mean, you
• 5:22
  Speaker 1: can't have an algorithm that accidentally throws one demographic in
• 5:26
  Speaker 1: jail twice as often as another.
• 5:28
  Speaker 2: No, you can't. But here is the massive problem.
• 5:31
  Speaker 1: Lay it on me.
• 5:32
  Speaker 2: Unless the underlying historical base rates of those two groups
• 5:36
  Speaker 2: are absolutely identical, which.
• 5:38
  Speaker 1: In a society with centuries of systemic inequality.
• 5:41
  Speaker 2: They never are exactly they never are. So unless they
• 5:44
  Speaker 2: are identical, it is a proven mathematical impossibility to satisfy
• 5:49
  Speaker 2: both predictive parity and equalize odds at the same time. Wait, really,
• 5:53
  Speaker 2: it's impost mathematically impossible. You literally cannot code an algorithm
• 5:57
  Speaker 2: that does both perfectly.
• 5:59
  Speaker 1: So if you calibrate, it's the system. So the risk
• 6:01
  Speaker 1: scores mean the exact same thing across demographics.
• 6:04
  Speaker 2: You will mathematically force a higher false positive rate on
• 6:07
  Speaker 2: the marginalized group.
• 6:08
  Speaker 1: Oh man, And if you force the false positive rates
• 6:12
  Speaker 1: to be equal.
• 6:12
  Speaker 2: Then your risk scores lose their accuracy for one of
• 6:15
  Speaker 2: the groups.
• 6:16
  Speaker 1: So the developers, they are forced to make a philosophical choice,
• 6:20
  Speaker 1: but they're hiding it behind this veil of objective optimization.
• 6:24
  Speaker 2: That's exactly what happens. They choose to optimize for predictive
• 6:27
  Speaker 2: parity because well, it makes a tool look more accurate
• 6:30
  Speaker 2: on a spreadsheet.
• 6:31
  Speaker 1: Well, quietly accepting that it will disproportionately generate false positives
• 6:36
  Speaker 1: for marginalized communities.
• 6:37
  Speaker 2: Yeah, and then they sell this to judges and loan
• 6:40
  Speaker 2: officers as unbiased, perfect math.
• 6:43
  Speaker 1: And the people pressing the button, you know, the judge,
• 6:46
  Speaker 1: the loan officer, the hiring manager, they trust the output
• 6:49
  Speaker 1: because it looks like a clean, authoritative, decimal point.
• 6:53
  Speaker 2: Right. The systemic inequality is no longer a human prejudice
• 6:56
  Speaker 2: that you can you know, cross examine in a courtroom.
• 6:58
  Speaker 2: It has been laundered through a machine learning.
• 7:00
  Speaker 1: Model, laundering bias through math. That is such a powerful
• 7:04
  Speaker 1: way to frame it. It's codified as statistical truth.
• 7:07
  Speaker 2: And this leads directly into the architectural problem of transparency.
• 7:11
  Speaker 1: Yes, the black box, because if I know the math
• 7:13
  Speaker 1: is prioritizing proxy variables and I know it's failing at
• 7:17
  Speaker 1: equalized odds, my first instinct as a software engineer is
• 7:20
  Speaker 1: just let's pop the hood.
• 7:21
  Speaker 2: You want to look at the code.
• 7:22
  Speaker 1: Right, I want to find the specific notes that are
• 7:24
  Speaker 1: waiting those proxy variables and manually prune them. But we
• 7:28
  Speaker 1: are dealing with the black box, and I want to
• 7:30
  Speaker 1: get away from the simplistic analogies here.
• 7:32
  Speaker 2: Yeah, it's not just a locked recipe box where you
• 7:34
  Speaker 2: don't know the ingredients.
• 7:35
  Speaker 1: Right. We are dealing with an architecture that is inherently
• 7:38
  Speaker 1: opaque by its very design.
• 7:40
  Speaker 2: We are talking about deep neural networks. Let's really map
• 7:44
  Speaker 2: the mechanics of why this is a black box.
• 7:46
  Speaker 1: Let's do it.
• 7:46
  Speaker 2: So you have an input layer where your data goes in,
• 7:51
  Speaker 2: and an output layer where the decision comes out. Simple enough, okay,
• 7:54
  Speaker 2: input and output, But in between you have hidden layers.
• 7:58
  Speaker 2: And in modern deep learning model, particularly large language models
• 8:02
  Speaker 2: or advanced computer vision systems, you have billions, sometimes trillions
• 8:07
  Speaker 2: of parameters.
• 8:08
  Speaker 1: Trillions. That's just incomprehensible, it really is.
• 8:10
  Speaker 2: And these parameters are essentially weights and biases applied to
• 8:14
  Speaker 2: the data as it passes through the network.
• 8:16
  Speaker 1: So it's not a decision tree. It's not you know,
• 8:18
  Speaker 1: if A is true, move to B, if B is false,
• 8:22
  Speaker 1: deny the loan.
• 8:23
  Speaker 2: No, not at all. It is a massive, cascading web
• 8:26
  Speaker 2: of probability. Let's use a better analogy than a flow chart.
• 8:30
  Speaker 1: Okay, I love a good analogy.
• 8:31
  Speaker 2: Imagine a quire of a billion singers and they are
• 8:34
  Speaker 2: standing in a massive canyon. You stand at one end,
• 8:38
  Speaker 2: and shout a freeze. That's your input data.
• 8:41
  Speaker 1: Okay, I'm shouting it at billion singers.
• 8:42
  Speaker 2: The first row of singers hears you, and based on
• 8:45
  Speaker 2: what they hear, they whisper a slightly modified note to
• 8:48
  Speaker 2: the row behind them. That row processes the whispers applies
• 8:53
  Speaker 2: their own internal modifier, which in AI is called an
• 8:56
  Speaker 2: activation function, and then they whisper to the next row.
• 9:00
  Speaker 1: This cascade happens across hundreds of layers.
• 9:02
  Speaker 2: Exactly by the time the sound reaches the other side
• 9:04
  Speaker 2: of the canyon, it has formed this beautiful, complex, perfect chord.
• 9:08
  Speaker 2: And that is your output.
• 9:10
  Speaker 1: Wow. And here's the terrifying part. If there's a sour
• 9:14
  Speaker 1: note buried inside that final chord.
• 9:16
  Speaker 2: Say if the algorithm unfairly denies a patient a critical
• 9:20
  Speaker 2: medical treatment.
• 9:21
  Speaker 1: Right, I cannot walk into that crowd of a billion
• 9:23
  Speaker 1: singers and figure out who started the sour note.
• 9:26
  Speaker 2: You can't because no single singer made the decision. The
• 9:29
  Speaker 2: decision exists entirely in the dynamic cascading interaction between all
• 9:34
  Speaker 2: of them.
• 9:34
  Speaker 1: The interaction is the decision. Yeah, that is wild.
• 9:37
  Speaker 2: And this opacity, it isn't a bug. It is the
• 9:40
  Speaker 2: fundamental feature that allows the system to process these highly
• 9:43
  Speaker 2: complex non linear relationships in the data.
• 9:47
  Speaker 1: So how does it actually learn to make that chord perfect?
• 9:51
  Speaker 2: When the network is training, it uses a process called backpropagation.
• 9:55
  Speaker 2: It basically looks at its errors, calculates the gradient of
• 9:59
  Speaker 2: the loss function.
• 10:00
  Speaker 1: Difference between what it guessed and what was right.
• 10:02
  Speaker 2: Exactly, and then it mathematically tweaks those billion weights backward
• 10:06
  Speaker 2: through the network to make the next guest slightly better.
• 10:08
  Speaker 1: So the machine is learning representations of the data that
• 10:11
  Speaker 1: are completely alien to human cognition.
• 10:15
  Speaker 2: Yes, it isn't looking at an X ray and seeing
• 10:18
  Speaker 2: a tumor the way a human doctor sees a shape.
• 10:20
  Speaker 1: What is it seeing.
• 10:21
  Speaker 2: It is seeing a multidimensional matrix of pixel intensities, and
• 10:25
  Speaker 2: it's recognizing a mathematical pattern that correlates with historical tumor data, which.
• 10:29
  Speaker 1: Creates a massive crisis in high stakes environments.
• 10:32
  Speaker 2: Oh, absolutely, think about it.
• 10:34
  Speaker 1: If an AI diagnoses a patient with an aggressive illness
• 10:38
  Speaker 1: and recommends this invasive surgery, the doctor is entirely flying blind.
• 10:42
  Speaker 2: Because if the doctor asks the system, hey, why did
• 10:45
  Speaker 2: you make this diagnosis?
• 10:46
  Speaker 1: The system cannot produce a human readable diagnostic pathway.
• 10:50
  Speaker 2: It can't It creates an immediate crisis of informed consent.
• 10:54
  Speaker 2: I mean a patient cannot legally or ethically consent to
• 10:57
  Speaker 2: a medical procedure if the primary diagnostition the algorithm cannot
• 11:01
  Speaker 2: explain its reasoning.
• 11:03
  Speaker 1: You can't exactly cross examine a matrix of billions of
• 11:06
  Speaker 1: floating point numbers.
• 11:07
  Speaker 2: You know, you really can't.
• 11:08
  Speaker 1: But this is where the industry is scrambling to build workarounds. Right.
• 11:12
  Speaker 1: We have this entire subfiel now called explainable AI or XAI.
• 11:15
  Speaker 2: Yes. XI is huge right.
• 11:17
  Speaker 1: Now because instead of forcing the neural network to be
• 11:19
  Speaker 1: inherently transparent, which like you said, would ruin as processing power,
• 11:23
  Speaker 1: they're building secondary tools to reverse engineer the.
• 11:25
  Speaker 2: Black box, right and the most prominent methods right now
• 11:28
  Speaker 2: are things like Lime and eshep.
• 11:30
  Speaker 1: Let's break down SAHA because I was reading the source
• 11:33
  Speaker 1: material on this and it uses game theory, which I
• 11:35
  Speaker 1: found just fascinating.
• 11:36
  Speaker 2: Yeah, SAHAP values are incredibly clever. They treat the model's
• 11:40
  Speaker 2: prediction as a payout in a cooperative game.
• 11:43
  Speaker 1: Okay, so the prediction is the price.
• 11:45
  Speaker 2: Right, and the different features of your data, so your age,
• 11:49
  Speaker 2: your blood pressure, your genetic markers, those are the players
• 11:53
  Speaker 2: in that game.
• 11:54
  Speaker 1: So how does SHP actually figure out who contributed what?
• 11:58
  Speaker 2: It mathematically calculates the marginal contribution of each individual feature
• 12:03
  Speaker 2: by looking at all possible combinations of those features.
• 12:07
  Speaker 1: Oh, I see. So it's essentially asks how much does
• 12:10
  Speaker 1: the prediction change if we include blood pressure versus if
• 12:14
  Speaker 1: we leave it out entirely across every possible scenario.
• 12:17
  Speaker 2: Exactly, it runs all this permutation.
• 12:18
  Speaker 1: So it generates like a heat map.
• 12:20
  Speaker 2: Yes, visual heat maps are very common for SAHAP.
• 12:23
  Speaker 1: So it tells the doctor, Okay, the algorithm diagnosed this illness,
• 12:26
  Speaker 1: and the features that push the probability the highest were
• 12:30
  Speaker 1: this specific cluster of pixels on the X ray and
• 12:33
  Speaker 1: this specific enzyme level in the blood work.
• 12:36
  Speaker 2: Exactly. It provides what we call a post hawk explanation.
• 12:39
  Speaker 2: It translates the alien math back into human logic.
• 12:43
  Speaker 1: Post hawk after the fact. But there is a massive trap,
• 12:46
  Speaker 1: there Isn't there a huge trap? Because researchers are raising
• 12:49
  Speaker 1: the alarm about this constantly. SAHAP value isn't telling you
• 12:53
  Speaker 1: how the model actually thinks.
• 12:55
  Speaker 2: No, it's not peering into its mind.
• 12:56
  Speaker 1: It is just providing a plausible, human readable summer of
• 13:00
  Speaker 1: the output. It is essentially an AIPR representative.
• 13:04
  Speaker 2: That is the perfect way to put it. It's pr
• 13:06
  Speaker 2: for the algorithm.
• 13:07
  Speaker 1: So if the underlying model is using a deeply biased
• 13:11
  Speaker 1: proxy variable in its hidden layers, the POSTOC explanation tool
• 13:15
  Speaker 1: might just you know, smooth that over.
• 13:17
  Speaker 2: It might completely mask it and instead present a justification
• 13:21
  Speaker 2: that looks medically or legally sound to the human operator.
• 13:24
  Speaker 1: So we might just be building highly sophisticated rationalization engines,
• 13:28
  Speaker 1: not true transparency tools.
• 13:30
  Speaker 2: And that is the danger of postoc explainability. You are
• 13:33
  Speaker 2: interpreting an interpretation. If the explainer model is flawed, it
• 13:36
  Speaker 2: creates a completely false sense of security. The human operator
• 13:41
  Speaker 2: looks at the SAHAP heat map, nods and says, ah, yes,
• 13:45
  Speaker 2: that makes logical sense to me.
• 13:47
  Speaker 1: Completely unaware that the underlying neural network actually triggered the
• 13:50
  Speaker 1: decision based on some spurious correlation.
• 13:53
  Speaker 2: Right like the font used in the medical report, or
• 13:57
  Speaker 2: a water mark that was present on all the positive
• 13:59
  Speaker 2: extra samples in the training data but not the negative ones.
• 14:03
  Speaker 1: The algorithm learns the watermark, not the tumor exactly. So
• 14:06
  Speaker 1: if the black box is inherently opaque and our XAI
• 14:10
  Speaker 1: tools are potentially just generating these plausible rationalizations. We have
• 14:14
  Speaker 1: a massive structural problem when things go.
• 14:16
  Speaker 2: Wrong, and things will absolutely go wrong.
• 14:19
  Speaker 1: Which forces us to look at the legal architecture. The
• 14:22
  Speaker 1: source text calls this the triad.
• 14:24
  Speaker 2: Of blame the legal vacuum.
• 14:25
  Speaker 1: Yes, because if the choir seems a sour dote and
• 14:28
  Speaker 1: someone loses their life or their liberty or their livelihood,
• 14:32
  Speaker 1: who gets sud right?
• 14:33
  Speaker 2: Current tort law is completely unequipped to handle this because
• 14:36
  Speaker 2: in this scenario, we have three distinct actors.
• 14:38
  Speaker 1: Okay, let's list them out.
• 14:40
  Speaker 2: First, you have the developer, the entity that engineered the
• 14:43
  Speaker 2: foundation model, like opening I are Google, right. Second, you
• 14:46
  Speaker 2: have the organization that deployed the system, so that's the hospital,
• 14:49
  Speaker 2: the bank, the court system.
• 14:51
  Speaker 1: Okay.
• 14:51
  Speaker 2: And third you have the end user, the individual, doctor,
• 14:56
  Speaker 2: loan officer, or judge who is actually interacting with the
• 14:59
  Speaker 2: inter Okay, I'm going.
• 15:01
  Speaker 1: To play Devil's advocate and take the side of the developer. Here,
• 15:04
  Speaker 1: go for it. Because if I'm a massive tech conglomerate
• 15:08
  Speaker 1: building a generalized foundation model, say a huge multimodal LM,
• 15:14
  Speaker 1: I am basically building an engine. I'm building infrastructure.
• 15:17
  Speaker 2: You just laying the pipes exactly.
• 15:19
  Speaker 1: I train it on the internet, I tune it, and
• 15:21
  Speaker 1: I release an API. I cannot possibly foresee that three
• 15:25
  Speaker 1: years later, some regional hospital is going to use my
• 15:28
  Speaker 1: API to build a bespoke triage tool that systematically underranks
• 15:33
  Speaker 1: the pain levels of minority patients. Right, I shouldn't be
• 15:36
  Speaker 1: held strictly liable for the downstream misapplication of a generalized tool, And.
• 15:40
  Speaker 2: That is exactly the defense they will use in court,
• 15:42
  Speaker 2: the lack of foreseeability. That makes sense, right, Well, think
• 15:45
  Speaker 2: about it in terms of traditional product liability. If a
• 15:49
  Speaker 2: manufacturer builds a chainsaw and the chainsaw explodes during normal
• 15:54
  Speaker 2: use because of a manufacturing defect, then.
• 15:56
  Speaker 1: The manufacturer is strictly liable. Right.
• 15:59
  Speaker 2: But if someone uses that same chainsaw to try and
• 16:01
  Speaker 2: cut a concrete pillar and it breaks and injures them.
• 16:05
  Speaker 1: The manufacturer isn't liable. You misuse the tool exactly.
• 16:08
  Speaker 2: So, the tech companies will argue that using a generalized
• 16:11
  Speaker 2: language model for high stakes medical diagnostics without human oversight
• 16:16
  Speaker 2: is the equivalent of using a chainsaw on concrete.
• 16:18
  Speaker 1: Okay, So liability flows downstream then to the organization the
• 16:22
  Speaker 1: hospital that integrated the API into their workflow and.
• 16:26
  Speaker 2: The organization is highly vulnerable here, but their legal defense
• 16:30
  Speaker 2: will be the black box reality we literally just discussed.
• 16:32
  Speaker 1: Oh man, They're going to say they didn't know how
• 16:35
  Speaker 1: it worked.
• 16:35
  Speaker 2: They will argue that they purchased a state of the
• 16:37
  Speaker 2: art system from a reputable developer who marketed it as
• 16:40
  Speaker 2: highly accurate and safe.
• 16:42
  Speaker 1: And the hospital administrators don't have PhDs in machine.
• 16:45
  Speaker 2: Learning exactly, they cannot audit a billion parameter neural network.
• 16:49
  Speaker 2: So they will argue that they acted reasonably by relying
• 16:52
  Speaker 2: on the developer's benchmarks.
• 16:54
  Speaker 1: Basically saying, how can you hold a hospital negligent for
• 16:58
  Speaker 1: failing to spot a mathematical anomaly buried in a hidden
• 17:02
  Speaker 1: layer of a proprietary neural network they don't even own.
• 17:05
  Speaker 2: Right, Which leaves us with the end user.
• 17:08
  Speaker 1: The doctor, the person pressing the button yes.
• 17:10
  Speaker 2: And this is where the legal concept of the learned
• 17:13
  Speaker 2: intermediary gets incredibly messy.
• 17:16
  Speaker 1: The learned intermediary break that down for us.
• 17:18
  Speaker 2: So traditionally, if a pharmaceutical company makes a drug, they
• 17:22
  Speaker 2: warn the doctor about the side effects. The doctor is
• 17:25
  Speaker 2: the learned intermediary.
• 17:26
  Speaker 1: They stand between the drug company and the patient.
• 17:28
  Speaker 2: Exactly if the doctor prescribes it. Wrongly. The liability falls
• 17:32
  Speaker 2: on them.
• 17:33
  Speaker 1: But the learned intermediary doctrine assumes the doctor understands how
• 17:37
  Speaker 1: the drug works exactly.
• 17:38
  Speaker 2: It assumes they can evaluate the risks based on their
• 17:41
  Speaker 2: medical expertise. But how does a doctor act as a
• 17:44
  Speaker 2: learned intermediary for an algorithmic output. They are mathematically barred
• 17:47
  Speaker 2: from understanding.
• 17:49
  Speaker 1: They can't. And on top of that, we have this
• 17:51
  Speaker 1: massive sociological phenomenon of automation bias.
• 17:55
  Speaker 2: OH Automation bias is huge here.
• 17:57
  Speaker 1: It's the innate human tendency to try us the machine
• 18:00
  Speaker 1: over our own judgment right exactly.
• 18:03
  Speaker 2: If a junior radiologist looks at a scan and thinks
• 18:06
  Speaker 2: it's benign, but the billion dollar enterprise AI system flags
• 18:11
  Speaker 2: it as a ninety eight percent probability of malignancy.
• 18:14
  Speaker 1: The psychological pressure to defer to the machine is immense.
• 18:17
  Speaker 2: It's overwhelming. If the doctor overrules the AI and the
• 18:21
  Speaker 2: patient dies, the doctor will be sued for medical malpractice
• 18:24
  Speaker 2: and totally crucified for ignoring the state of the art system.
• 18:28
  Speaker 1: But if the doctor defers to the AI and the
• 18:31
  Speaker 1: patient suffers an unnecessary.
• 18:32
  Speaker 2: Surgery, the doctor will just say I was following protocol
• 18:35
  Speaker 2: the machine said ninety eight percent.
• 18:37
  Speaker 1: It creates a perfect liability shield where everyone points at
• 18:40
  Speaker 1: the other two actors and the victim is left with
• 18:43
  Speaker 1: zero recourse. Nobody is at fault.
• 18:45
  Speaker 2: And this is exactly why we are seeing a massive
• 18:48
  Speaker 2: push to restructure liability laws specifically for.
• 18:51
  Speaker 1: AI, because we can't rely on twentieth century negligent standards
• 18:55
  Speaker 1: for twenty first century probabilistic models.
• 18:57
  Speaker 2: We need frameworks that mandate algorithmic audits and assign strict
• 19:01
  Speaker 2: liability before the system is even deployed.
• 19:04
  Speaker 1: But to conduct those algorithmic audits, we have to talk
• 19:06
  Speaker 1: about the data architecture, the fuel of the machine.
• 19:10
  Speaker 2: Yes, you cannot audit a model's fairness or its accuracy
• 19:14
  Speaker 2: without full access to the training data.
• 19:17
  Speaker 1: And the continuous real time data it processes. Right, and
• 19:20
  Speaker 1: this isn't just anonymous telemetry data anymore, is it. The
• 19:23
  Speaker 1: models we are building now require extreme granularity.
• 19:27
  Speaker 2: To predict behavior. The AI needs to model your behavior perfectly, and.
• 19:32
  Speaker 1: This brings up a striking paradox that you highlighted earlier.
• 19:35
  Speaker 1: To build a highly accurate, personalized and safe AI assistant
• 19:38
  Speaker 1: or diagnostic tool, that tool requires an inherently invasive level of.
• 19:42
  Speaker 2: Surveillance accuracy demands the death of privacy.
• 19:45
  Speaker 1: Man. That is a grim trade off. It is.
• 19:48
  Speaker 2: The data economy has structurally been designed around infinite extraction.
• 19:52
  Speaker 2: But we are moving past the era of simply scraping
• 19:55
  Speaker 2: static databases.
• 19:56
  Speaker 1: Right, because the cutting edge is how models continuously train
• 19:59
  Speaker 1: on user interact.
• 20:00
  Speaker 2: Yes, and the technical solutions being proposed to solve this
• 20:05
  Speaker 2: privacy paradox are actually fascinating. You don't just solve this
• 20:08
  Speaker 2: with a better privacy policy, you know.
• 20:10
  Speaker 1: You have to solve it at the architectural level.
• 20:12
  Speaker 2: Exactly like federated learning.
• 20:15
  Speaker 1: Okay, let's get into federated learning, because this concept completely
• 20:19
  Speaker 1: flips the traditional data extraction model on its head.
• 20:22
  Speaker 2: It's a brilliant mechanical.
• 20:24
  Speaker 1: Shift because instead of hoarding all the data in a
• 20:26
  Speaker 1: central server, you leave the data where it is.
• 20:29
  Speaker 3: Right.
• 20:30
  Speaker 2: Yes, Traditionally, if you want to train an AI to
• 20:35
  Speaker 2: predict text on a smartphone keyboard, you would have to
• 20:38
  Speaker 2: upload millions of users' private text messages to a central server,
• 20:43
  Speaker 2: train the model, and then push the update back to
• 20:45
  Speaker 2: the phone, which.
• 20:46
  Speaker 1: Is a massive privacy vulnerability.
• 20:48
  Speaker 2: It's a honeypot for hackers.
• 20:49
  Speaker 1: But with federated learning, the data never leaves the user's phone.
• 20:53
  Speaker 2: Right, The developer sends a copy of the untrained foundational
• 20:56
  Speaker 2: model down to your phone.
• 20:57
  Speaker 1: Okay, so the model comes to me.
• 20:59
  Speaker 2: Yes, and the model trains locally on your device, learning
• 21:03
  Speaker 2: from your specific private text messages. It updates its internal
• 21:08
  Speaker 2: weights and parameters based on your personal behavior.
• 21:11
  Speaker 1: And then what it doesn't send my text back.
• 21:13
  Speaker 2: Obviously, No, your phone sends only the updated parameters, the
• 21:17
  Speaker 2: mathematical adjustments back to the central server.
• 21:21
  Speaker 1: Ah, so just the math, not the content exactly.
• 21:24
  Speaker 2: The server averages out the mathematical adjustments from millions of
• 21:26
  Speaker 2: phones and updates the master model. So the machine learns
• 21:30
  Speaker 2: the overarching patterns of human communication without ever actually reading
• 21:34
  Speaker 2: a single private message.
• 21:35
  Speaker 1: It's elegant, it really is. But I was reading that.
• 21:38
  Speaker 1: It still has vulnerabilities.
• 21:40
  Speaker 2: It does. It's not a silver bullet.
• 21:41
  Speaker 1: Because researchers have shown that if you analyze those parameter
• 21:44
  Speaker 1: updates carefully enough, you can actually perform what's called a
• 21:47
  Speaker 1: model inversion attack.
• 21:49
  Speaker 2: Yes, those are very real threats.
• 21:50
  Speaker 1: Basically, you could look at the gradient updates a specific
• 21:53
  Speaker 1: phone sent back and reverse engineer the private data that
• 21:57
  Speaker 1: caused those updates.
• 21:58
  Speaker 2: Right, Because if the mathematical weight for a specific, highly
• 22:01
  Speaker 2: unusual word spiked on your phone.
• 22:03
  Speaker 1: The server nos, I tech that word.
• 22:04
  Speaker 2: Exactly, which is why federated learning has to be combined
• 22:07
  Speaker 2: with differential privacy.
• 22:09
  Speaker 1: Okay, differential privacy. This is the mathematical science of injecting
• 22:13
  Speaker 1: statistical noise into a data set. Right, Yes, Let's explain
• 22:17
  Speaker 1: how that noise injection actually works, because when you first
• 22:20
  Speaker 1: hear it, it sounds like you are just breaking the data.
• 22:23
  Speaker 2: Well, you are slightly breaking the individual data points to
• 22:26
  Speaker 2: preserve the macro level accuracy.
• 22:29
  Speaker 1: Let's use an analogy. The text had a great one
• 22:31
  Speaker 1: about salaries.
• 22:32
  Speaker 2: Right. Imagine I want to know the average salary of
• 22:35
  Speaker 2: a group of people, but no one wants to reveal
• 22:38
  Speaker 2: their exact salary because that's private. Exactly, So I tell
• 22:42
  Speaker 2: everyone flown a coin and secret. If it's heads, tell
• 22:45
  Speaker 2: me your real salary. Okay, If it's tails, flip a
• 22:48
  Speaker 2: second coin. If that one is heads, tell me in
• 22:50
  Speaker 2: a randomly generated high number. If it's tails, a randomly
• 22:54
  Speaker 2: generated low number.
• 22:55
  Speaker 1: Okay. I love this. So when I collect the answers,
• 22:57
  Speaker 1: I look at an individual number, and I have total
• 23:01
  Speaker 1: plausible deniability.
• 23:02
  Speaker 2: Exactly, I don't know if that's their real salary or
• 23:04
  Speaker 2: a random number. Their individual privacy is mathematically guaranteed.
• 23:08
  Speaker 1: But and this is the magic part. Because I know
• 23:12
  Speaker 1: the statistical probability of the coin flips, I.
• 23:14
  Speaker 2: Can run an equation on the aggregated data to subtract
• 23:18
  Speaker 2: the expected noise.
• 23:19
  Speaker 1: So you can still accurately determine the average salary of
• 23:22
  Speaker 1: the group, even though you cannot trust any single data point.
• 23:26
  Speaker 2: That's differential privacy in a nutshell. In machine learning. It
• 23:30
  Speaker 2: injects that kind of mathematical noise into the gradient updates
• 23:34
  Speaker 2: before they are sent to the server, so the.
• 23:36
  Speaker 1: Server can learn the overarching patterns of the crowd without
• 23:39
  Speaker 1: ever being able to memorize the specific data of an individual.
• 23:42
  Speaker 2: It is an incredible cryptographic balancing act, it really is.
• 23:46
  Speaker 1: Yeah, and they use a specific privacy budget, right, an
• 23:49
  Speaker 1: epsilon value to determine exactly how much noise to inject.
• 23:52
  Speaker 2: Right, Because too much noise the model is useless. Too
• 23:56
  Speaker 2: little noise, the privacy is compromised.
• 23:58
  Speaker 1: But even if we perfectly secure the data architecture with
• 24:01
  Speaker 1: federated learning and differential privacy, we still run into a
• 24:04
  Speaker 1: philosophical limit.
• 24:05
  Speaker 2: The limits of vector space yes, we.
• 24:08
  Speaker 1: Have perfected the data pipelines, but we have to look
• 24:11
  Speaker 1: at what the machine actually does with that data in
• 24:14
  Speaker 1: the context of the human experience.
• 24:16
  Speaker 2: Because there is a dimensional reality to human life that
• 24:19
  Speaker 2: simply cannot be mapped as of vector embedding.
• 24:22
  Speaker 1: Let's define what we mean by vector embeddings here, because
• 24:25
  Speaker 1: this is the absolute core of how modern AI supposedly
• 24:29
  Speaker 1: understands the world.
• 24:31
  Speaker 2: Right, and we need to put quotes around understands.
• 24:35
  Speaker 1: Large language models do not understand language now, they don't.
• 24:38
  Speaker 2: They do not know what a dog is. They do
• 24:40
  Speaker 2: not know what sadness is.
• 24:41
  Speaker 1: So what are they doing.
• 24:42
  Speaker 2: They take words and translate them into high dimensional numerical
• 24:46
  Speaker 2: arrays vectors. They plot these vectors in a space with
• 24:50
  Speaker 2: thousands of dimensions, and words that frequently appear together in
• 24:54
  Speaker 2: the training data are plotted physically closer together in that
• 24:57
  Speaker 2: mathematical space.
• 24:58
  Speaker 1: So the word dog is math sthmatically close to puppy
• 25:01
  Speaker 1: and leash exactly.
• 25:03
  Speaker 2: Distance equals semantic similarity. That's how the model generates text.
• 25:07
  Speaker 1: It calculates the mathematical trajectory from one word to the
• 25:10
  Speaker 1: next and just predicts the most statistically probable token to
• 25:14
  Speaker 1: complete the sequence.
• 25:15
  Speaker 2: That's all it's doing. It is a stunningly powerful illusion
• 25:18
  Speaker 2: of comprehension.
• 25:19
  Speaker 1: An illusion, yeah, but fundamentally just statistical geometry exactly. And
• 25:24
  Speaker 1: that geometry hits a very hard wall when you try
• 25:27
  Speaker 1: to apply it to moral reasoning or context specific empathy.
• 25:31
  Speaker 2: It fails completely.
• 25:32
  Speaker 1: Let's look at the criminal justice system again. A human
• 25:35
  Speaker 1: judge is sitting in a courtroom. A defendant is standing
• 25:39
  Speaker 1: before them.
• 25:39
  Speaker 2: Right.
• 25:40
  Speaker 1: The judge isn't just processing a rap sheet. The judge
• 25:43
  Speaker 1: is reading the room. They're sensing the defendant's remorse or
• 25:46
  Speaker 1: the lack thereof.
• 25:47
  Speaker 2: They're taking into account the highly specific, nuanced sociological context
• 25:53
  Speaker 2: of the defendant's community exactly.
• 25:55
  Speaker 1: The grief of the victim's family sitting in the front row,
• 25:58
  Speaker 1: the subtle micro expressions of the attorneys.
• 26:01
  Speaker 2: The judge is processing qualitative, multimodal reality that simply cannot
• 26:05
  Speaker 2: be quantified into a data set.
• 26:07
  Speaker 1: Right. But the risk assessment algorithm doesn't have sensors for grief.
• 26:11
  Speaker 1: It doesn't have a vector embedding for the sudden realization
• 26:15
  Speaker 1: of a misspent youth. It only has historical data points
• 26:19
  Speaker 1: prior arrests, zip code, employment status.
• 26:22
  Speaker 2: It flattens the infinite dimensionality of a human life into
• 26:26
  Speaker 2: a handful of proxy variables and outputs a decimal.
• 26:29
  Speaker 1: Which is why the push for fully autonomous decision making
• 26:34
  Speaker 1: in high stakes human systems is so incredibly dangerous.
• 26:38
  Speaker 2: It isn't just that the algorithms might be biased. It's
• 26:40
  Speaker 2: that they are fundamentally incapable of moral reasoning.
• 26:43
  Speaker 1: Because empathy requires a shared vulnerability, it requires the capacity
• 26:48
  Speaker 1: to suffer.
• 26:49
  Speaker 2: Yes, an algorithm cannot weigh the moral gravity of separating
• 26:52
  Speaker 2: a parent from a child because it does not possess
• 26:54
  Speaker 2: the capacity to understand what a parent or a child
• 26:57
  Speaker 2: actually is.
• 26:57
  Speaker 1: It only knows that the token parent and the to
• 27:00
  Speaker 1: okenchild frequently appear together in the training data. That's it.
• 27:03
  Speaker 1: So the human in the loop isn't just some regulatory
• 27:05
  Speaker 1: safety net, It is a mechanical necessity.
• 27:08
  Speaker 2: The machine provides the statistical pattern recognition, the scale, and
• 27:12
  Speaker 2: the speed, but the human must provide the context and
• 27:15
  Speaker 2: the moral weight.
• 27:16
  Speaker 1: The machine calculates the probability, the human makes the judgment.
• 27:20
  Speaker 2: Which is exactly the philosophy underpinning the most aggressive regulatory
• 27:24
  Speaker 2: frameworks currently being drafted.
• 27:25
  Speaker 1: Globally, like the European Union's AI.
• 27:27
  Speaker 2: Act specifically, the EUAI Act, governments are finally realizing that
• 27:32
  Speaker 2: relying on the tech industry to self regulate is a
• 27:35
  Speaker 2: catastrophic failure of governance.
• 27:37
  Speaker 1: I mean, you cannot ask a corporation driven by fiduciary
• 27:41
  Speaker 1: duty to its shareholders to voluntarily bottleneck their own technological
• 27:46
  Speaker 1: supremacy for the sake of abstract ethical concerns.
• 27:49
  Speaker 2: You can't. It goes against their entire operational model.
• 27:52
  Speaker 1: And the EUAI Act is fascinating because it doesn't just
• 27:55
  Speaker 1: try to blindly regulate AI as this big monolith.
• 27:59
  Speaker 2: No, it uses a tea weird risk based architecture.
• 28:02
  Speaker 1: It categorizes the technology strictly by the level of threat
• 28:05
  Speaker 1: it poses to fundamental human rights.
• 28:07
  Speaker 2: It creates four distinct tiers. First, you have unacceptable risks,
• 28:11
  Speaker 2: and those are just banned, banned outright period, things like
• 28:14
  Speaker 2: social scoring systems run by governments, cognitive behavioral manipulation, real
• 28:18
  Speaker 2: time biometric categorization in public spaces.
• 28:21
  Speaker 1: The EU is basically drawing a hard line in the
• 28:23
  Speaker 1: sand and saying we do not care how efficient this
• 28:27
  Speaker 1: technology is. It is fundamentally incompatible with a democratic society exactly.
• 28:32
  Speaker 2: Then you have the high risk tier, which is where
• 28:34
  Speaker 2: almost everything we've been talking about today, lives.
• 28:36
  Speaker 1: Medical devices, critical infrastructure, educational scoring.
• 28:40
  Speaker 2: Employment filtering, criminal justice tools, all of it.
• 28:44
  Speaker 1: And if you are a developer building a high risk system,
• 28:47
  Speaker 1: the compliance burden under this act is massive.
• 28:50
  Speaker 2: Oh, it's monumental. You cannot just release the API and
• 28:53
  Speaker 2: see what happens anymore.
• 28:55
  Speaker 1: You have to undergo rigorous conformity assessments before the product
• 28:59
  Speaker 1: even hits the market.
• 29:00
  Speaker 2: You must prove that your training data sets are high quality,
• 29:03
  Speaker 2: relevant and free of discriminatory bias.
• 29:05
  Speaker 1: You have to build in logging capabilities so the system's
• 29:08
  Speaker 1: decisions can actually be traced and audited.
• 29:11
  Speaker 2: And crucially, tying back to our earlier point, you must
• 29:14
  Speaker 2: guarantee human oversight, the human and the loop. Yes, the
• 29:18
  Speaker 2: system has to be designed in a way that allows
• 29:20
  Speaker 2: a human operator to fully override or intervene in the
• 29:23
  Speaker 2: automated decision.
• 29:24
  Speaker 1: But there is a massive logistical pushback against this from
• 29:27
  Speaker 1: the engineering community, of course. I mean, I look at
• 29:30
  Speaker 1: the requirements for these conformity assessments, the data logging, the
• 29:34
  Speaker 1: bias audits, the red teaming, and I see an incredibly
• 29:37
  Speaker 1: expensive compliance apparatus.
• 29:40
  Speaker 2: It's millions of dollars in compliance alone.
• 29:42
  Speaker 1: And the critique here is regulatory capture, because if you
• 29:45
  Speaker 1: make the legal requirements to build a foundational model so
• 29:49
  Speaker 1: astronomically expensive and legally perilous.
• 29:52
  Speaker 2: The only entities who will be able to afford to
• 29:54
  Speaker 2: build AI are the massive tech monopolies.
• 29:57
  Speaker 1: Google, Microsoft, Meta. You effectively legislate open source developers and
• 30:03
  Speaker 1: small startups right out of existence.
• 30:05
  Speaker 2: And that is the absolute tightrope regulators are walking right now, because.
• 30:10
  Speaker 1: If the regulations are too loose, the technology causes societal harm.
• 30:14
  Speaker 2: But if the regulations are too draconian, you entrench the
• 30:17
  Speaker 2: power of a few mega corporations and totally stifle innovation.
• 30:20
  Speaker 1: And those mega corporations already hold unprecedented power over the
• 30:24
  Speaker 1: digital infrastructure of the globe. They do if they are
• 30:27
  Speaker 1: the only ones legally permitted to build these models, they
• 30:30
  Speaker 1: essentially become unelected geopolitical actors. They get to dictate the
• 30:34
  Speaker 1: acceptable parameters of reality for the entire Internet.
• 30:37
  Speaker 2: Which is why the regulatory framework cannot rely solely on
• 30:41
  Speaker 2: government auditors. It requires systemic public engagement and a thriving
• 30:46
  Speaker 2: open source ecosystem.
• 30:47
  Speaker 1: Open source is key here.
• 30:49
  Speaker 2: It is because when models are open source, researchers across
• 30:53
  Speaker 2: the globe can audit the weights, They can analyze the
• 30:56
  Speaker 2: vector embeddings.
• 30:57
  Speaker 1: They can find the proxy variables and the hidden biases
• 31:00
  Speaker 1: that a closed corporate API would never ever reveal.
• 31:04
  Speaker 2: Transparency at the code level, not just at the policy.
• 31:07
  Speaker 1: Level, exactly, we need independent technologists, ethicists, and civil society
• 31:12
  Speaker 1: groups actively probing these systems.
• 31:14
  Speaker 2: We need adversarial testing conducted not just by the company
• 31:18
  Speaker 2: trying to sell the product, but by the communities who
• 31:20
  Speaker 2: will actually be subjected to the algorithm's decisions.
• 31:23
  Speaker 1: Because the boundaries of automated decision making must be established
• 31:26
  Speaker 1: through democratic friction, not closed door engineering sprints.
• 31:30
  Speaker 2: That is the only way we prevent a hyper concentrated
• 31:33
  Speaker 2: group of developers in Silicon Valley from unilaterally deciding what
• 31:36
  Speaker 2: constitutes a fair risk assessment or a valid hiring metric
• 31:40
  Speaker 2: for the rest of the planet.
• 31:41
  Speaker 1: Because if we leave it entirely to the engineers.
• 31:44
  Speaker 2: The system will optimize for computational efficiency and seamless scale, And.
• 31:48
  Speaker 1: If we leave it entirely to the corporations.
• 31:51
  Speaker 2: The system will optimize for user retention and quarterly profit.
• 31:54
  Speaker 1: It is only through rigorous adversarial public engagement and unyielding
• 31:59
  Speaker 1: legal frameworks that we force the technology to optimize for
• 32:03
  Speaker 1: human dignity. Well said, we have covered massive complex terrain today.
• 32:08
  Speaker 1: We dismantled the myth of objective math, looking mechanically at
• 32:12
  Speaker 1: how algorithms reconstruct biases through proxy variables and.
• 32:15
  Speaker 2: The mathematical impossibility of satisfying both predictive parity and equalized odds.
• 32:20
  Speaker 1: We stared into the cascating billion parameter reality of the
• 32:23
  Speaker 1: black box, exploring how deep learning architectures rely on nonlinear
• 32:28
  Speaker 1: activation functions that just completely defy human logic.
• 32:32
  Speaker 2: And how tools like setia A values try and honestly
• 32:35
  Speaker 2: sometimes fail to reverse engineer that alien math into post
• 32:40
  Speaker 2: hawk rationalizations.
• 32:41
  Speaker 1: We navigated the liability vacuum, examining the complex legal interplay
• 32:45
  Speaker 1: between developers claiming lack of foreseeability, organizations hiding behind the
• 32:50
  Speaker 1: black box, and end users trapped by automation bias.
• 32:53
  Speaker 2: We looked at the architectural demands for privacy, breaking down
• 32:57
  Speaker 2: how federated learning and differential privacy attempt to secure the
• 33:00
  Speaker 2: data pipelines that feed these hungry models.
• 33:03
  Speaker 1: We discuss the fundamental limits of high dimensional vector space,
• 33:07
  Speaker 1: realizing that algorithms can map semantic similarity sure but they
• 33:11
  Speaker 1: can never replicate the moral weight, the context, and the
• 33:15
  Speaker 1: shared vulnerability required for true human empathy.
• 33:17
  Speaker 2: And finally, we analyze the mechanics of the EUAI Act,
• 33:20
  Speaker 2: the immense difficulty of conformity assessments, and the absolute necessity
• 33:25
  Speaker 2: of open source auditing to prevent regulatory capture.
• 33:27
  Speaker 1: The overarching truth here is that building a mathematically smarter
• 33:30
  Speaker 1: machine is only half the engineering challenge.
• 33:33
  Speaker 2: A flawless neural network that scales systemic inequality is a
• 33:36
  Speaker 2: failed system.
• 33:37
  Speaker 1: The ultimate goal is sociotechnical. We have to architect models
• 33:41
  Speaker 1: that don't just process reality faster, but actively preserve the
• 33:45
  Speaker 1: structural integrity of a just society exactly, which brings us
• 33:49
  Speaker 1: back to the image of the algorithm as a camera
• 33:52
  Speaker 1: endlessly taking high dimensional photos of our historical data. If
• 33:55
  Speaker 1: artificial intelligence is ultimately just a mirror reflecting the statistical
• 33:59
  Speaker 1: reaction of our world back at us, codified into code.
• 34:03
  Speaker 2: Then maybe the problem isn't the mirror, right.
• 34:06
  Speaker 1: Maybe the profound underlying challenge isn't just mathematically tweaking the mirror.
• 34:10
  Speaker 1: Maybe the challenge is figuring out how to fix the
• 34:12
  Speaker 1: society the machine is actually looking at, what kind of
• 34:15
  Speaker 1: reflection are you feeding the algorithm today? Thank you for
• 34:18
  Speaker 1: joining us on this deep dive. Keep questioning the invisible
• 34:21
  Speaker 1: forces shaping your world.