WEBVTT 00:00.120 --> 00:05.640 This opening slide introduces the foundation of responsible AI development. 00:06.000 --> 00:10.000 Generative AI systems are no longer experimental tools. 00:10.360 --> 00:16.720 They actively influence real decisions, shape opinions, and impact opportunities at scale. 00:17.280 --> 00:25.080 The title and subtitle emphasize a critical goal building AI systems that earn and maintain user trust. 00:25.560 --> 00:31.720 The visual on this slide shows the connected neural structure surrounded by symbols representing fairness, 00:31.760 --> 00:33.720 protection, and accountability. 00:34.240 --> 00:40.840 This reinforces the idea that ethics is not a single feature, but a network of responsibilities embedded 00:40.840 --> 00:41.960 throughout the system. 00:42.760 --> 00:45.360 This section is not about abstract philosophy. 00:45.800 --> 00:52.960 It is about practical decisions engineers and product teams make every day what data to use, what behavior 00:52.960 --> 00:56.320 to allow, and how to respond when things go wrong. 00:57.000 --> 01:00.360 Ethical choices shape how users experience AI. 01:00.760 --> 01:06.750 By the end of this section, you should recognize that ethics is not a constraint on innovation. 01:07.110 --> 01:11.750 It is what allows innovation to scale responsibly and sustainably. 01:12.470 --> 01:17.830 This slide explains why ethics is an operational necessity, not an optional add on. 01:18.470 --> 01:25.030 As highlighted in the slide text, generative AI systems shape critical aspects of modern life, from 01:25.030 --> 01:28.230 hiring and education to healthcare and finance. 01:28.950 --> 01:31.990 The slide breaks ethical impact into three dimensions. 01:32.150 --> 01:38.510 User impact emphasizes that AI decisions affect people's lives, careers, and well-being. 01:39.030 --> 01:45.150 Legal risk reflects the tightening global regulatory landscape where violations can lead to serious 01:45.150 --> 01:45.910 penalties. 01:46.390 --> 01:49.190 Trust capital underscores a powerful truth. 01:49.630 --> 01:54.790 Once users lose trust in an AI system, it is extremely difficult to rebuild. 01:55.310 --> 01:59.150 The key insight at the bottom of the slide is particularly important. 01:59.510 --> 02:03.470 Every deployed AI system reflects the values of its creators. 02:03.870 --> 02:08.460 What your system does and does not do reveals what you prioritize. 02:09.020 --> 02:13.700 Ethics must therefore be embedded into every stage of development and deployment. 02:14.100 --> 02:18.660 Without ethical guardrails, technical success becomes organizational risk. 02:18.860 --> 02:23.380 This slide explores where bias and generative AI systems comes from. 02:23.860 --> 02:27.380 A critical point made here is that models do not invent bias. 02:27.780 --> 02:32.140 They learn and amplify patterns present in their training data and design. 02:32.780 --> 02:35.100 The slide outlines four major sources. 02:35.620 --> 02:41.740 Training data imbalance occurs when data sets overrepresent certain groups, languages, or perspectives. 02:42.340 --> 02:49.820 Historical and societal bias reflects discrimination and stereotypes embedded in source material, language 02:49.820 --> 02:51.180 and cultural dominance. 02:51.340 --> 02:56.660 Often prioritizes English and Western contexts at the expense of global diversity. 02:57.260 --> 03:03.580 Feedback loop amplification occurs when user interactions reinforce existing biases rather than correcting 03:03.580 --> 03:03.980 them. 03:04.420 --> 03:10.380 The visual on this slide reinforces interconnected pathways, symbolizing how bias propagates through 03:10.380 --> 03:11.500 systems over time. 03:12.100 --> 03:14.380 The reality statement at the bottom is crucial. 03:14.780 --> 03:18.820 Models learn statistical correlations, not ethical principles. 03:19.340 --> 03:24.020 Recognizing this limitation is the first step toward responsible mitigation. 03:24.300 --> 03:29.020 This slide identifies common bias patterns that appear in deployed systems. 03:29.620 --> 03:35.500 Demographic bias shows up as different outputs based on gender, race, or age. 03:35.860 --> 03:38.860 Such as skewed hiring or healthcare recommendations. 03:39.500 --> 03:46.500 Cultural and geographic bias manifests as Western centric assumptions or culturally inappropriate suggestions. 03:47.060 --> 03:52.780 Confirmation bias reinforces user assumptions instead of challenging incorrect beliefs. 03:53.380 --> 03:58.740 Representation bias leads to missing perspectives or stereotypical portrayals. 03:59.420 --> 04:04.700 The example at the bottom of the slide, identical prompts with different demographic names producing 04:04.700 --> 04:09.100 different outcomes, illustrates how subtle and measurable bias can be. 04:09.740 --> 04:14.850 These issues often emerge through aggregate behavior rather than isolated failures. 04:15.490 --> 04:18.290 The engineering challenge highlighted here is critical. 04:18.770 --> 04:21.690 Bias is rarely obvious in a single response. 04:22.170 --> 04:26.210 It requires systematic testing across large samples and scenarios. 04:26.770 --> 04:29.090 Spot checking is not enough. 04:29.250 --> 04:34.770 This slide reframes fairness as a design requirement rather than an emergent property. 04:35.210 --> 04:41.410 Fair systems demonstrate consistent behavior, ensuring similar inputs receive similar treatment. 04:41.890 --> 04:47.450 They provide equal access, allowing all users to leverage full system capabilities. 04:48.050 --> 04:53.890 They produce non-discriminatory outcomes, avoiding systematic disadvantage to protected groups. 04:54.490 --> 04:57.850 The slide also outlines practical implementation approaches. 04:58.370 --> 05:05.050 Diverse evaluation data sets ensure testing across demographics, languages and use cases. 05:05.570 --> 05:10.970 Biased testing protocols systematically measure differential performance during development. 05:11.450 --> 05:16.240 Human review adds expert oversight for sensitive or high stakes outputs. 05:16.800 --> 05:21.760 The rule at the bottom is unambiguous fairness must be designed from the start. 05:22.200 --> 05:28.080 Retrofitting fairness after deployment is expensive, incomplete and often ineffective. 05:28.120 --> 05:33.040 This slide focuses on data privacy risks unique to generative AI. 05:33.880 --> 05:41.240 Unlike traditional software, Llms can memorize and reproduce sensitive information, creating new vectors 05:41.240 --> 05:42.920 for privacy violations. 05:43.680 --> 05:45.800 The slide identifies key risks. 05:46.280 --> 05:51.720 Training data exposure can lead to personal or proprietary information appearing in outputs. 05:52.240 --> 05:57.400 Prompt logging creates databases of sensitive user queries that become breach targets. 05:58.080 --> 06:01.720 Model memorization allows regurgitation of specific training. 06:01.720 --> 06:02.480 Examples. 06:03.040 --> 06:06.520 Unintentional leakage can occur through inference attacks over time. 06:07.080 --> 06:13.360 High risk domains such as healthcare, finance, legal services, and HR systems require heightened 06:13.360 --> 06:17.040 vigilance due to regulatory and ethical obligations. 06:17.480 --> 06:22.390 The core message is clear Privacy failures are often subtle and cumulative. 06:22.750 --> 06:27.470 Preventing them requires both technical safeguards and disciplined data practices. 06:27.510 --> 06:32.230 This slide presents concrete best practices for responsible data handling. 06:32.670 --> 06:36.670 The guiding principle is simple collect only what you need. 06:36.870 --> 06:40.990 Protect everything you keep, and delete what you no longer use. 06:41.710 --> 06:44.350 The slide outlines four actionable steps. 06:44.630 --> 06:48.110 Minimize collection by questioning every data point gathered. 06:48.630 --> 06:52.630 Limit retention by avoiding storage of raw prompts whenever possible. 06:53.190 --> 06:58.630 Encrypt everything both at rest and in transit using industry standards. 06:59.110 --> 07:04.150 Control access to a strict role based permissions and comprehensive audit logging. 07:04.870 --> 07:11.750 The statistic on this slide is striking 83% of data breaches involve unnecessary collected data. 07:12.230 --> 07:17.030 This reinforces the principle at the bottom if you don't need the data, don't keep it. 07:17.550 --> 07:21.190 Good privacy practice is both ethical and practical. 07:21.550 --> 07:24.710 It reduces risk by reducing attack surface. 07:24.750 --> 07:30.390 The final slide focuses on transparency as the foundation of user trust. 07:30.870 --> 07:37.870 Users deserve to know when they are interacting with AI, what it can and cannot do, and when its confidence 07:37.870 --> 07:38.630 is limited. 07:39.310 --> 07:47.470 The slide outlines for trust building practices disclose AI usage clearly explain limitations honestly. 07:47.910 --> 07:51.190 Signal uncertainty when the system is less confident. 07:51.790 --> 07:55.310 Enable user control over data and interaction loads. 07:55.990 --> 08:03.390 It also highlights anti-patterns that erode trust, hiding AI involvement, overstating accuracy, and 08:03.390 --> 08:05.710 masking failures or hallucinations. 08:06.550 --> 08:12.150 These practices may temporarily improve perception, but they destroy trust long term. 08:12.790 --> 08:14.670 The closing statement is powerful. 08:14.950 --> 08:17.310 Transparency is not a feature to AB. 08:17.310 --> 08:20.190 Later, it must be embedded from day one. 08:20.870 --> 08:26.190 Users can forgive limitations they understand, but they will not forgive deception.