WEBVTT 1 00:00:00.000 --> 00:00:01.260 In this lesson, 2 00:00:01.260 --> 00:00:04.950 we will learn about AI usage risks. 3 00:00:04.950 --> 00:00:08.610 Artificial intelligence, or AI, usage risks 4 00:00:08.610 --> 00:00:12.540 are the potential dangers and unintended consequences 5 00:00:12.540 --> 00:00:15.810 that arise from the implementation and reliance 6 00:00:15.810 --> 00:00:18.660 on artificial intelligence systems. 7 00:00:18.660 --> 00:00:22.830 Risks of AI usage include overreliance, 8 00:00:22.830 --> 00:00:26.370 sensitive information disclosure to the model, 9 00:00:26.370 --> 00:00:29.730 sensitive information disclosure from the model, 10 00:00:29.730 --> 00:00:33.150 and excessive agency of the AI. 11 00:00:33.150 --> 00:00:36.570 Overreliance on AI can lead to situations 12 00:00:36.570 --> 00:00:39.210 where human judgment is undervalued, 13 00:00:39.210 --> 00:00:41.520 potentially resulting in decisions 14 00:00:41.520 --> 00:00:44.910 that overlook important contextual factors 15 00:00:44.910 --> 00:00:47.610 that AI cannot calculate for. 16 00:00:47.610 --> 00:00:50.850 Sensitive information disclosure to the model 17 00:00:50.850 --> 00:00:53.730 occurs when private or confidential data 18 00:00:53.730 --> 00:00:58.680 is input into AI systems without adequate safeguards. 19 00:00:58.680 --> 00:01:01.890 Sensitive information disclosure from the model 20 00:01:01.890 --> 00:01:05.460 involves the unintentional release of private data 21 00:01:05.460 --> 00:01:08.760 that the AI has learned during its training. 22 00:01:08.760 --> 00:01:13.470 Finally, excessive agency of the AI refers to situations 23 00:01:13.470 --> 00:01:18.270 where AI systems are given too much control or autonomy. 24 00:01:18.270 --> 00:01:20.910 Let's learn more about overreliance, 25 00:01:20.910 --> 00:01:24.180 sensitive information disclosure to the model, 26 00:01:24.180 --> 00:01:27.540 sensitive information disclosure from the model, 27 00:01:27.540 --> 00:01:30.900 and excessive agency of the AI. 28 00:01:30.900 --> 00:01:33.750 First, we have overreliance. 29 00:01:33.750 --> 00:01:38.750 Overreliance on AI happens when users place too much trust 30 00:01:38.970 --> 00:01:43.230 in the model's output without verifying the accuracy 31 00:01:43.230 --> 00:01:45.330 or considering other factors, 32 00:01:45.330 --> 00:01:49.770 such as contextual understanding, ethical considerations, 33 00:01:49.770 --> 00:01:51.840 limitations of training data, 34 00:01:51.840 --> 00:01:56.130 and unexpected environmental or external influences. 35 00:01:56.130 --> 00:01:58.440 This risk becomes significant 36 00:01:58.440 --> 00:02:01.140 when human oversight is minimized 37 00:02:01.140 --> 00:02:04.080 and AI predictions or suggestions 38 00:02:04.080 --> 00:02:06.360 are accepted without question. 39 00:02:06.360 --> 00:02:09.090 In fields like finance, for instance, 40 00:02:09.090 --> 00:02:13.170 an analyst may overly depend on an AI algorithm 41 00:02:13.170 --> 00:02:16.260 to generate investment recommendations, 42 00:02:16.260 --> 00:02:19.380 overlooking broader economic contexts 43 00:02:19.380 --> 00:02:21.930 or unusual market anomalies 44 00:02:21.930 --> 00:02:24.540 that the model may not factor in. 45 00:02:24.540 --> 00:02:27.390 This kind of unquestioned reliance 46 00:02:27.390 --> 00:02:31.920 can lead to decisions misaligned with real-world conditions, 47 00:02:31.920 --> 00:02:35.190 increasing the potential for financial loss 48 00:02:35.190 --> 00:02:37.170 or missed opportunities. 49 00:02:37.170 --> 00:02:41.040 In technical environments, overreliance on AI 50 00:02:41.040 --> 00:02:45.270 can also exacerbate preexisting data biases 51 00:02:45.270 --> 00:02:48.450 or structural gaps within the model. 52 00:02:48.450 --> 00:02:51.390 This is because many AI systems 53 00:02:51.390 --> 00:02:53.640 are trained on historical data, 54 00:02:53.640 --> 00:02:56.160 which can reflect past biases, 55 00:02:56.160 --> 00:02:59.790 limitations, or specific environments. 56 00:02:59.790 --> 00:03:02.640 So, if an AI model is trained 57 00:03:02.640 --> 00:03:04.980 on this type of historical data 58 00:03:04.980 --> 00:03:07.710 and is trusted over human judgment, 59 00:03:07.710 --> 00:03:11.850 particularly in complex or ambiguous situations, 60 00:03:11.850 --> 00:03:14.760 critical nuances may be overlooked. 61 00:03:14.760 --> 00:03:19.260 So instead, combining AI insights with human expertise 62 00:03:19.260 --> 00:03:22.350 ensures that decisions are both data-informed 63 00:03:22.350 --> 00:03:24.270 and contextually grounded. 64 00:03:24.270 --> 00:03:27.690 Second, we have sensitive information disclosure 65 00:03:27.690 --> 00:03:31.620 to the model refers to the risk of feeding private, 66 00:03:31.620 --> 00:03:36.330 confidential, or proprietary data into the AI 67 00:03:36.330 --> 00:03:39.060 without robust privacy controls. 68 00:03:39.060 --> 00:03:42.540 This is done by inputting data into the system 69 00:03:42.540 --> 00:03:46.380 without applying adequate encryption, anonymization, 70 00:03:46.380 --> 00:03:48.750 or role-based access controls, 71 00:03:48.750 --> 00:03:52.110 leaving the data vulnerable within the model. 72 00:03:52.110 --> 00:03:55.950 When sensitive data is entered into an AI system, 73 00:03:55.950 --> 00:03:59.610 it becomes part of the model's accessible data pool, 74 00:03:59.610 --> 00:04:02.010 which can lead to potential misuse 75 00:04:02.010 --> 00:04:05.370 if privacy safeguards are lacking or weak. 76 00:04:05.370 --> 00:04:08.430 Without stringent data protection protocols, 77 00:04:08.430 --> 00:04:12.150 this information is at risk of improper storage, 78 00:04:12.150 --> 00:04:15.660 unauthorized access, or unintended use 79 00:04:15.660 --> 00:04:19.620 within other processes relying on the model. 80 00:04:19.620 --> 00:04:23.280 For example, consider a customer service department 81 00:04:23.280 --> 00:04:27.120 that uses AI to analyze customer feedback 82 00:04:27.120 --> 00:04:29.100 and behavior patterns. 83 00:04:29.100 --> 00:04:33.480 If this system also receives sensitive customer details, 84 00:04:33.480 --> 00:04:37.140 such as financial information, health records, 85 00:04:37.140 --> 00:04:39.870 or personally identifiable data, 86 00:04:39.870 --> 00:04:43.020 without sufficient security mechanisms, 87 00:04:43.020 --> 00:04:45.120 the data could be vulnerable. 88 00:04:45.120 --> 00:04:48.600 This vulnerability may expose the organization 89 00:04:48.600 --> 00:04:51.840 to compliance risks or data breaches, 90 00:04:51.840 --> 00:04:55.530 especially in sectors governed by regulations 91 00:04:55.530 --> 00:04:58.350 like the General Data Protection Regulation 92 00:04:58.350 --> 00:05:02.520 and the Health Insurance Portability and Accountability Act. 93 00:05:02.520 --> 00:05:05.823 Without granular data access controls, 94 00:05:05.823 --> 00:05:10.050 information might be retained beyond intended uses, 95 00:05:10.050 --> 00:05:12.840 shared across unrelated processes, 96 00:05:12.840 --> 00:05:16.620 or susceptible to unintended model behaviors 97 00:05:16.620 --> 00:05:20.790 where sensitive insights might inadvertently be disclosed 98 00:05:20.790 --> 00:05:22.860 in generated outputs. 99 00:05:22.860 --> 00:05:25.410 So, to mitigate this risk, 100 00:05:25.410 --> 00:05:29.670 organizations should employ strict data anonymization, 101 00:05:29.670 --> 00:05:31.740 secure access controls, 102 00:05:31.740 --> 00:05:35.310 and clearly defined data usage guidelines. 103 00:05:35.310 --> 00:05:38.670 These safeguards not only protect privacy, 104 00:05:38.670 --> 00:05:43.320 but also ensure that sensitive data is handled in alignment 105 00:05:43.320 --> 00:05:47.850 with regulatory standards and organizational policies, 106 00:05:47.850 --> 00:05:51.600 reducing the likelihood of unintended disclosures 107 00:05:51.600 --> 00:05:56.460 and preserving data integrity within AI processes. 108 00:05:56.460 --> 00:05:59.760 Third, we have sensitive information disclosure 109 00:05:59.760 --> 00:06:01.440 from the model. 110 00:06:01.440 --> 00:06:06.150 Sensitive information disclosure from an AI model can occur 111 00:06:06.150 --> 00:06:09.780 when the model unintentionally releases private data 112 00:06:09.780 --> 00:06:13.860 it encountered during its training or operational use. 113 00:06:13.860 --> 00:06:16.590 This risk is particularly relevant 114 00:06:16.590 --> 00:06:20.010 for AI models trained on large datasets 115 00:06:20.010 --> 00:06:22.530 that include sensitive information, 116 00:06:22.530 --> 00:06:26.370 as the model could inadvertently reveal specific details 117 00:06:26.370 --> 00:06:28.920 about individuals, companies, 118 00:06:28.920 --> 00:06:32.400 or situations it learned from this data. 119 00:06:32.400 --> 00:06:35.550 For example, in a customer service setting, 120 00:06:35.550 --> 00:06:39.150 an AI assistant trained on historical chat data 121 00:06:39.150 --> 00:06:42.540 might draw from specific customer interactions 122 00:06:42.540 --> 00:06:44.730 to answer new inquiries. 123 00:06:44.730 --> 00:06:48.150 If the model recalls and reveals past details, 124 00:06:48.150 --> 00:06:50.640 such as an individual's account balance, 125 00:06:50.640 --> 00:06:54.270 recent transactions, or other private information, 126 00:06:54.270 --> 00:06:56.550 this could breach confidentiality, 127 00:06:56.550 --> 00:07:00.750 exposing sensitive data to unintended parties. 128 00:07:00.750 --> 00:07:03.540 Such incidents can arise in healthcare 129 00:07:03.540 --> 00:07:06.960 or legal applications where an AI trained 130 00:07:06.960 --> 00:07:10.170 on confidential case notes or medical records 131 00:07:10.170 --> 00:07:13.440 might unintentionally incorporate and reveal 132 00:07:13.440 --> 00:07:17.010 personal health information or legal details 133 00:07:17.010 --> 00:07:19.740 in responses to new users. 134 00:07:19.740 --> 00:07:22.650 This is particularly concerning in sectors 135 00:07:22.650 --> 00:07:26.640 with strict confidentiality standards and regulations, 136 00:07:26.640 --> 00:07:28.740 like the Health Insurance Portability 137 00:07:28.740 --> 00:07:33.150 and Accountability Act, or attorney-client privilege. 138 00:07:33.150 --> 00:07:36.180 So, to prevent such disclosures, 139 00:07:36.180 --> 00:07:39.120 AI models need robust mechanisms, 140 00:07:39.120 --> 00:07:41.250 such as differential privacy 141 00:07:41.250 --> 00:07:45.000 and strict data handling protocols, that can obscure 142 00:07:45.000 --> 00:07:49.230 or anonymize sensitive details from responses. 143 00:07:49.230 --> 00:07:53.940 Effective safeguards not only protect individuals' privacy, 144 00:07:53.940 --> 00:07:57.930 but also uphold the integrity and confidentiality 145 00:07:57.930 --> 00:08:01.980 required in professional and regulated environments. 146 00:08:01.980 --> 00:08:06.300 Fourth and last, we have excessive agency of AI 147 00:08:06.300 --> 00:08:09.780 occurs when a system is given too much autonomy, 148 00:08:09.780 --> 00:08:13.230 allowing it to make decisions or take actions 149 00:08:13.230 --> 00:08:17.100 without human supervision or intervention. 150 00:08:17.100 --> 00:08:18.660 This can be problematic 151 00:08:18.660 --> 00:08:22.500 if the AI operates beyond its intended scope, 152 00:08:22.500 --> 00:08:25.620 as it may make choices that are misaligned 153 00:08:25.620 --> 00:08:29.910 with organizational values or user expectations. 154 00:08:29.910 --> 00:08:34.140 An example of this can be seen in self-driving cars 155 00:08:34.140 --> 00:08:37.530 that are allowed to make all driving decisions. 156 00:08:37.530 --> 00:08:40.650 If the car encounters an unusual situation 157 00:08:40.650 --> 00:08:43.680 it wasn't trained for, it could make a decision 158 00:08:43.680 --> 00:08:46.620 with potentially dangerous consequences, 159 00:08:46.620 --> 00:08:50.160 underscoring the need for controlled autonomy. 160 00:08:50.160 --> 00:08:53.280 Similarly, in professional environments, 161 00:08:53.280 --> 00:08:57.780 AI should operate with limits on its decision-making power, 162 00:08:57.780 --> 00:09:01.680 enabling human oversight to guide, intervene, 163 00:09:01.680 --> 00:09:04.320 and align with the AI's actions 164 00:09:04.320 --> 00:09:08.220 with broader ethical and practical considerations. 165 00:09:08.220 --> 00:09:11.850 Balancing AI autonomy with human supervision 166 00:09:11.850 --> 00:09:16.850 ensures that the AI's agency is both effective and safe, 167 00:09:17.160 --> 00:09:21.690 reducing the risk of unintended or harmful actions 168 00:09:21.690 --> 00:09:24.120 and promoting outcomes that align 169 00:09:24.120 --> 00:09:26.880 with human goals and values. 170 00:09:26.880 --> 00:09:30.810 So, to prevent the excessive agency of AI, 171 00:09:30.810 --> 00:09:34.230 organizations should implement oversight mechanisms 172 00:09:34.230 --> 00:09:38.400 that restrict AI's autonomy through structured workflows, 173 00:09:38.400 --> 00:09:43.170 human-in-the-loop systems, and decision auditing processes. 174 00:09:43.170 --> 00:09:46.830 These well-defined supervised contexts 175 00:09:46.830 --> 00:09:51.690 ensure human review in critical decision-making processes. 176 00:09:51.690 --> 00:09:56.690 So, remember, artificial intelligence, or AI, usage risks 177 00:09:57.840 --> 00:10:02.430 are potential dangers that arise from relying on AI systems 178 00:10:02.430 --> 00:10:06.030 without proper controls or understanding. 179 00:10:06.030 --> 00:10:09.150 Key risks include overreliance, 180 00:10:09.150 --> 00:10:12.390 sensitive information disclosure to the model, 181 00:10:12.390 --> 00:10:15.450 sensitive information disclosure from the model, 182 00:10:15.450 --> 00:10:18.750 and excessive agency of the AI. 183 00:10:18.750 --> 00:10:23.750 Overreliance occurs when users trust the AI output too much, 184 00:10:23.760 --> 00:10:26.550 overlooking critical human judgment. 185 00:10:26.550 --> 00:10:30.600 Next, sensitive information disclosure to the model 186 00:10:30.600 --> 00:10:34.020 happens when private data is shared with AI 187 00:10:34.020 --> 00:10:38.280 without proper or adequate safeguards, risking misuse. 188 00:10:38.280 --> 00:10:42.330 Next, sensitive information disclosure from the model 189 00:10:42.330 --> 00:10:46.260 involves the unintentional release of private data 190 00:10:46.260 --> 00:10:49.050 that AI has encountered in training, 191 00:10:49.050 --> 00:10:51.930 potentially breaching confidentiality. 192 00:10:51.930 --> 00:10:55.470 Finally, excessive agency of the AI 193 00:10:55.470 --> 00:10:58.950 occurs when systems are given too much autonomy, 194 00:10:58.950 --> 00:11:03.930 potentially leading to unsafe or unaligned decisions. 195 00:11:03.930 --> 00:11:06.300 So, to address these risks, 196 00:11:06.300 --> 00:11:09.720 organizations should combine AI insights 197 00:11:09.720 --> 00:11:14.550 with human expertise, enforce strict data privacy protocols, 198 00:11:14.550 --> 00:11:17.793 and limit AI's decision-making power.