1
00:00:00,050 --> 00:00:06,140
Case study mitigating AI bias, hallucinations, and errors Finn trusts ethical and technical approach.

2
00:00:06,170 --> 00:00:11,510
Artificial intelligence systems, despite their transformative potential, are vulnerable to failures

3
00:00:11,510 --> 00:00:14,540
such as bias, hallucinations, and errors.

4
00:00:15,110 --> 00:00:20,270
These issues are particularly pertinent for AI governance professionals, who must ensure the ethical

5
00:00:20,270 --> 00:00:23,300
and accountable deployment of AI technologies.

6
00:00:25,160 --> 00:00:31,760
Consider the case of AI implementation in a multinational bank, Fintrust, which has adopted AI algorithms

7
00:00:31,760 --> 00:00:34,160
to streamline loan approval processes.

8
00:00:34,190 --> 00:00:40,790
The AI system, designed to evaluate credit worthiness based on historical data and socio economic indicators,

9
00:00:40,790 --> 00:00:44,720
initially promised to increase efficiency and reduce human error.

10
00:00:45,440 --> 00:00:49,250
However, as the system was deployed, a disturbing trend emerged.

11
00:00:49,280 --> 00:00:55,130
A significant disparity in loan approvals among different demographic groups, African American and

12
00:00:55,130 --> 00:01:00,890
Hispanic applicants were systematically denied loans at higher rates compared to their Caucasian counterparts.

13
00:01:01,940 --> 00:01:06,910
How can APIs identify and address the root causes of such biases in AI systems.

14
00:01:08,290 --> 00:01:13,720
The bias in fintrust AI systems stemmed from the training data, which was based on historical loan

15
00:01:13,720 --> 00:01:18,490
approval records that reflected decades of discriminatory lending practices.

16
00:01:19,300 --> 00:01:25,600
By learning from biased data, the AI system perpetuated these biases, leading to unfair outcomes.

17
00:01:26,500 --> 00:01:31,870
To rectify this, Fintrust data scientists embarked on a thorough review of the dataset, identifying

18
00:01:31,870 --> 00:01:34,060
key features that contributed to bias.

19
00:01:34,090 --> 00:01:39,430
They augmented the dataset with more representative samples of all demographic groups, and incorporated

20
00:01:39,430 --> 00:01:44,050
fairness aware algorithms designed to mitigate bias during the learning process.

21
00:01:44,500 --> 00:01:49,270
What specific techniques can be used to ensure datasets are balanced and representative?

22
00:01:51,820 --> 00:01:57,760
To further enhance fairness, Fintrust implemented regular audits of the AI system involving external

23
00:01:57,760 --> 00:02:01,180
auditors to review the models decisions and provide feedback.

24
00:02:01,210 --> 00:02:06,940
These audits helped identify any remaining biases and ensured continuous improvement of the system.

25
00:02:07,480 --> 00:02:13,640
Moreover, the bank established an ethical AI Committee comprising technologists, ethicists and community

26
00:02:13,640 --> 00:02:19,340
representatives to oversee the AI deployment and ensure alignment with societal values.

27
00:02:21,170 --> 00:02:25,520
Another challenge emerged in the use of AI for customer service at Fintrust.

28
00:02:25,820 --> 00:02:32,090
The bank deployed a chatbot powered by an advanced NLP model to handle customer inquiries.

29
00:02:32,420 --> 00:02:38,630
Initially, the chatbot performed well, but over time it began providing inaccurate or irrelevant information.

30
00:02:39,020 --> 00:02:44,660
For example, it fabricated details about loan products that did not exist and gave incorrect advice

31
00:02:44,660 --> 00:02:46,010
on financial planning.

32
00:02:46,010 --> 00:02:52,790
How can I prevent AI systems from generating hallucinations and ensure the accuracy of the information

33
00:02:52,790 --> 00:02:53,660
provided?

34
00:02:54,140 --> 00:02:59,870
The issue of hallucinations in the chatbot arose from the model's training objective, which prioritized

35
00:02:59,870 --> 00:03:03,380
fluency and coherence over factual accuracy.

36
00:03:03,740 --> 00:03:09,800
To address this, Fintrust integrated factual verification mechanisms into the chatbots architecture.

37
00:03:10,220 --> 00:03:15,940
These mechanisms cross-check the information generated by the chatbot against a verified knowledge base,

38
00:03:15,940 --> 00:03:19,330
ensuring that responses were grounded in factual data.

39
00:03:19,870 --> 00:03:26,020
Additionally, the bank employed a human in the loop approach where human agents reviewed and validated

40
00:03:26,020 --> 00:03:29,590
the chatbots responses before they were delivered to customers.

41
00:03:30,430 --> 00:03:36,100
This hybrid approach significantly reduced the incidence of hallucinations and enhanced the reliability

42
00:03:36,100 --> 00:03:36,820
of the chatbot.

43
00:03:39,010 --> 00:03:43,720
The use of AI in healthcare at Fintrust also posed significant challenges.

44
00:03:44,080 --> 00:03:50,080
The bank developed an AI diagnostic tool to assist medical professionals in interpreting medical images,

45
00:03:50,110 --> 00:03:55,030
aiming to improve diagnostic accuracy and reduce the workload on radiologists.

46
00:03:55,450 --> 00:04:01,120
However, during a pilot phase, the tool misinterpreted several images, leading to incorrect diagnoses

47
00:04:01,120 --> 00:04:03,040
and recommendations for treatment.

48
00:04:03,460 --> 00:04:08,410
These errors, if left unaddressed, could have had severe consequences for patient health.

49
00:04:09,070 --> 00:04:15,340
What steps can be taken to improve the accuracy and interpretability of AI diagnostic tools in healthcare?

50
00:04:16,240 --> 00:04:21,710
The errors in the AI diagnostic tool were partly due to the complexity and opacity of the deep learning

51
00:04:21,710 --> 00:04:26,660
models used, which operated as black boxes with limited interpretability.

52
00:04:27,200 --> 00:04:33,020
To enhance transparency, Fintrust data scientists employed model agnostic interpretability methods

53
00:04:33,020 --> 00:04:39,410
such as Lime and Shape, which provided insights into the decision making process of the AI system.

54
00:04:39,830 --> 00:04:45,020
These methods allowed medical professionals to understand the rationale behind the AI's predictions

55
00:04:45,020 --> 00:04:47,060
and identify potential errors.

56
00:04:47,810 --> 00:04:53,210
Moreover, the bank developed inherently interpretable models that offered a balance between complexity

57
00:04:53,210 --> 00:04:59,480
and transparency, ensuring that the diagnostic tools decisions were more understandable and trustworthy.

58
00:05:02,030 --> 00:05:08,180
The deployment of AI in dynamic and unpredictable environments, such as autonomous vehicles, presents

59
00:05:08,210 --> 00:05:09,590
additional risks.

60
00:05:10,430 --> 00:05:17,030
Fintrust had invested in an AI powered fleet of autonomous delivery vehicles to enhance logistics efficiency.

61
00:05:17,540 --> 00:05:23,000
However, during a trial run, one of the vehicles failed to recognize an obstacle on the road, resulting

62
00:05:23,000 --> 00:05:24,410
in a minor accident.

63
00:05:24,890 --> 00:05:30,170
This incident underscored the potential dangers of AI errors in safety critical applications.

64
00:05:30,170 --> 00:05:35,810
What measures can be implemented to ensure the robust performance of AI systems in such contexts?

65
00:05:37,190 --> 00:05:43,310
To address the risks associated with autonomous vehicles, Fintrust adopted a rigorous testing and validation

66
00:05:43,310 --> 00:05:44,150
framework.

67
00:05:44,570 --> 00:05:50,390
This framework involved extensive simulations and real world tests under various conditions, to assess

68
00:05:50,390 --> 00:05:54,710
the AI system's performance and identify potential failure points.

69
00:05:55,520 --> 00:06:00,770
Continuous monitoring was also implemented to detect anomalies and ensure timely intervention.

70
00:06:01,400 --> 00:06:07,160
Additionally, the bank collaborated with regulatory bodies and industry experts to develop safety standards

71
00:06:07,160 --> 00:06:10,520
and best practices for autonomous vehicle deployment.

72
00:06:12,410 --> 00:06:17,960
The ethical implications of AI failures extend beyond technical considerations affecting broader societal

73
00:06:17,960 --> 00:06:18,710
impacts.

74
00:06:19,190 --> 00:06:25,160
Bias, hallucinations, and errors in AI systems can exacerbate inequalities, undermine trust, and

75
00:06:25,160 --> 00:06:27,500
erode public confidence in technology.

76
00:06:27,920 --> 00:06:33,800
How can I integrate ethical principles with technical solutions to ensure responsible AI deployment?

77
00:06:35,330 --> 00:06:41,540
At Fintrust, a comprehensive approach to AI governance was adopted, integrating ethical principles

78
00:06:41,540 --> 00:06:43,310
with technical solutions.

79
00:06:43,850 --> 00:06:49,790
The Bank established clear accountability mechanisms, including transparent documentation of AI development

80
00:06:49,790 --> 00:06:52,850
processes and rigorous auditing practices.

81
00:06:53,420 --> 00:06:58,850
Channels were created for stakeholders to report grievances and provide feedback, ensuring that any

82
00:06:58,850 --> 00:07:00,620
issues were promptly addressed.

83
00:07:01,220 --> 00:07:07,220
Moreover, interdisciplinary collaboration among technologists, ethicists, legal experts, and affected

84
00:07:07,220 --> 00:07:13,610
communities was fostered to ensure that AI systems were designed and deployed in ways that aligned with

85
00:07:13,610 --> 00:07:15,470
societal values and norms.

86
00:07:16,340 --> 00:07:21,440
Understanding AI failures is crucial for developing ethical and accountable AI systems.

87
00:07:21,890 --> 00:07:27,440
Addressing these challenges requires a multifaceted approach that combines technical innovations with

88
00:07:27,440 --> 00:07:29,210
robust governance frameworks.

89
00:07:29,570 --> 00:07:36,610
By prioritizing fairness, accuracy, transparency, and collaboration, AIG can help mitigate the risks

90
00:07:36,610 --> 00:07:42,460
associated with AI failures and harness the technology's potential for positive societal impact.

91
00:07:43,390 --> 00:07:49,000
The case of Fintrust illustrates the complexities and challenges associated with AI deployment across

92
00:07:49,000 --> 00:07:50,320
various sectors.

93
00:07:51,130 --> 00:07:56,830
The bank's commitment to addressing bias involved curating balanced data sets and implementing fairness

94
00:07:56,830 --> 00:08:01,510
aware algorithms, ensuring that the AI systems decisions were equitable.

95
00:08:01,990 --> 00:08:07,930
By integrating factual verification mechanisms and human in the loop approaches, Fintrust minimizes

96
00:08:07,930 --> 00:08:14,080
the risk of hallucinations in its chatbot, enhancing the reliability of customer interactions in the

97
00:08:14,080 --> 00:08:15,010
healthcare domain.

98
00:08:15,010 --> 00:08:20,800
Model agnostic interpretability methods and inherently interpretable models improve the accuracy and

99
00:08:20,800 --> 00:08:23,380
transparency of AI diagnostic tools.

100
00:08:23,380 --> 00:08:26,050
Building trust among medical professionals.

101
00:08:27,640 --> 00:08:33,640
The incident with the Autonomous delivery vehicle highlighted the importance of rigorous testing, validation,

102
00:08:33,640 --> 00:08:39,620
and continuous monitoring to ensure the robust performance of AI systems in dynamic environments.

103
00:08:40,280 --> 00:08:45,590
The bank's collaboration with regulatory bodies and industry experts underscored the need for safety

104
00:08:45,620 --> 00:08:49,580
standards and best practices in deploying autonomous technologies.

105
00:08:51,620 --> 00:08:57,080
Finn Trust's comprehensive approach to AI governance, integrating ethical principles with technical

106
00:08:57,080 --> 00:09:03,050
solutions, exemplified how organisations can navigate the ethical implications of AI failures.

107
00:09:03,080 --> 00:09:08,810
Establishing clear accountability mechanisms, fostering interdisciplinary collaboration, and engaging

108
00:09:08,810 --> 00:09:14,570
with affected communities ensured that AI systems were aligned with societal values and norms.

109
00:09:16,820 --> 00:09:22,190
In conclusion, the case of Finn Trust demonstrates the importance of understanding and addressing AI

110
00:09:22,220 --> 00:09:28,610
failures, bias, hallucinations, and errors to develop ethical and accountable AI systems.

111
00:09:28,610 --> 00:09:34,520
By combining technical innovations with robust governance frameworks, organisations can mitigate risks

112
00:09:34,520 --> 00:09:38,330
and leverage AI's potential for positive societal impact.
