1
00:00:00,050 --> 00:00:03,740
Lesson group harms discrimination and bias in AI systems.

2
00:00:03,770 --> 00:00:09,590
Artificial intelligence systems have increasingly become integral to various aspects of society, from

3
00:00:09,590 --> 00:00:13,190
health care and law enforcement to finance and advertising.

4
00:00:13,940 --> 00:00:19,310
However, with this growing reliance on AI, there is an urgent need to address the potential harms

5
00:00:19,310 --> 00:00:23,900
these systems can inflict, particularly in terms of discrimination and bias.

6
00:00:24,920 --> 00:00:30,890
Discrimination and bias in AI systems manifest when these technologies perpetuate or even exacerbate

7
00:00:30,890 --> 00:00:37,580
existing societal inequalities, often unintentionally due to the data they are trained on or the algorithms

8
00:00:37,580 --> 00:00:39,080
that govern their operations.

9
00:00:40,910 --> 00:00:45,680
One of the primary sources of bias in AI systems is the data used to train them.

10
00:00:46,100 --> 00:00:51,920
AI systems, particularly those based on machine learning, rely on vast amounts of data to learn and

11
00:00:51,920 --> 00:00:53,090
make decisions.

12
00:00:53,390 --> 00:01:00,070
If the training data contains biases, the AI system will likely replicate and even amplify these biases.

13
00:01:00,070 --> 00:01:06,190
For instance, if an AI system is trained on a data set that reflects historical hiring practices favoring

14
00:01:06,190 --> 00:01:11,890
certain demographics over others, the system may recommend job applicants in a biased manner, thus

15
00:01:11,890 --> 00:01:14,290
reinforcing workplace discrimination.

16
00:01:14,890 --> 00:01:17,200
This phenomenon is not merely hypothetical.

17
00:01:17,200 --> 00:01:22,060
There have been real world instances where AI systems have demonstrated biased behavior.

18
00:01:22,570 --> 00:01:29,380
In 2018, Amazon scrapped an AI recruiting tool after discovering it was biased against women as it

19
00:01:29,380 --> 00:01:34,630
had been trained on resumes submitted over a ten year period, predominantly from men.

20
00:01:36,190 --> 00:01:42,700
Another critical issue is that AI systems can inadvertently learn and perpetuate societal stereotypes.

21
00:01:43,000 --> 00:01:43,840
Word embeddings.

22
00:01:43,870 --> 00:01:49,090
A common technique in natural language processing have been shown to capture and reflect gender and

23
00:01:49,090 --> 00:01:51,760
racial biases present in the training data.

24
00:01:52,090 --> 00:01:54,040
For example, Bolukbasi et al.

25
00:01:54,070 --> 00:02:00,300
Demonstrated that word embeddings trained on a large corpus of text, associated the word man with computer

26
00:02:00,300 --> 00:02:03,090
programmer and woman with homemaker.

27
00:02:03,540 --> 00:02:09,540
This kind of bias in AI can have far reaching implications, influencing everything from language translation

28
00:02:09,540 --> 00:02:12,270
services to search engine suggestions.

29
00:02:13,440 --> 00:02:18,600
Algorithmic bias can also arise from the design and implementation of AI systems.

30
00:02:19,080 --> 00:02:25,140
Developers subjective choices, whether in selecting training data, defining objectives, or designing

31
00:02:25,140 --> 00:02:28,260
evaluation metrics, can introduce biases.

32
00:02:28,650 --> 00:02:33,870
For instance, facial recognition technologies have been found to have higher error rates for individuals

33
00:02:33,870 --> 00:02:37,770
with darker skin tones compared to those with lighter skin tones.

34
00:02:37,800 --> 00:02:43,140
This disparity is often due to the lack of diversity in the training datasets, and the developer's

35
00:02:43,140 --> 00:02:48,000
oversight in ensuring the system's fairness across different demographic groups.

36
00:02:48,600 --> 00:02:55,840
Such biases can lead to serious consequences such as misidentification and wrongful accusations, particularly

37
00:02:55,840 --> 00:02:57,520
in law enforcement contexts.

38
00:02:58,630 --> 00:03:04,810
Moreover, the deployment of biased AI systems can have significant societal impacts, particularly

39
00:03:04,840 --> 00:03:07,180
on marginalized and vulnerable groups.

40
00:03:07,600 --> 00:03:13,720
For example, predictive policing algorithms, which aim to forecast criminal activity based on historical

41
00:03:13,720 --> 00:03:18,520
crime data, have been criticized for disproportionately targeting minority communities.

42
00:03:18,910 --> 00:03:25,150
These systems can perpetuate a cycle of overpolicing and criminalization of these communities, exacerbating

43
00:03:25,150 --> 00:03:27,070
existing social inequalities.

44
00:03:28,000 --> 00:03:34,360
Similarly, AI driven credit scoring systems can systematically disadvantage certain demographic groups,

45
00:03:34,360 --> 00:03:39,730
leading to unequal access to financial services and perpetuating economic disparities.

46
00:03:40,780 --> 00:03:46,090
Addressing discrimination and bias in AI systems requires a multifaceted approach.

47
00:03:46,660 --> 00:03:51,330
One critical step is ensuring the diversity and representativeness of training data.

48
00:03:51,750 --> 00:03:57,660
This involves not only collecting data from diverse sources, but also continuously monitoring and updating

49
00:03:57,660 --> 00:04:01,560
datasets to capture changes in societal norms and behaviors.

50
00:04:01,950 --> 00:04:07,770
Additionally, employing techniques such as reweighting, resampling, and debiasing during the data

51
00:04:07,770 --> 00:04:10,950
preprocessing stage can help mitigate bias.

52
00:04:11,760 --> 00:04:17,430
Another important measure is the implementation of fairness aware algorithms, which are designed to

53
00:04:17,460 --> 00:04:20,430
explicitly account for and mitigate biases.

54
00:04:21,030 --> 00:04:26,490
These algorithms can be programmed to ensure that their decisions do not disproportionately benefit

55
00:04:26,490 --> 00:04:28,650
or harm any particular group.

56
00:04:29,040 --> 00:04:34,950
Techniques such as fairness constraints and adversarial debiasing can be utilized to achieve this goal.

57
00:04:35,730 --> 00:04:41,910
Furthermore, it is essential to establish clear guidelines and standards for fairness in AI, including

58
00:04:41,910 --> 00:04:48,220
the development of robust evaluation metrics that can assess the fairness of AI systems across different

59
00:04:48,220 --> 00:04:50,500
contexts and demographic groups.

60
00:04:51,700 --> 00:04:56,410
Transparency and accountability are also crucial in addressing bias in AI.

61
00:04:57,070 --> 00:05:02,650
Developers and organizations must be transparent about the data sources, methodologies, and decision

62
00:05:02,680 --> 00:05:05,770
making processes used in their AI systems.

63
00:05:06,550 --> 00:05:11,890
This transparency can be achieved through documentation practices such as model cards and data sheets

64
00:05:11,890 --> 00:05:17,440
for data sets, which provide detailed information about the data and model characteristics, including

65
00:05:17,440 --> 00:05:19,720
potential biases and limitations.

66
00:05:20,350 --> 00:05:26,020
Additionally, establishing accountability mechanisms such as regular audits and impact assessments

67
00:05:26,020 --> 00:05:32,320
by independent third parties can help ensure that AI systems are held to high ethical and fairness standards.

68
00:05:34,210 --> 00:05:40,750
Public and stakeholder engagement is another vital component in addressing AI bias, involving a diverse

69
00:05:40,750 --> 00:05:46,470
range of stakeholders, including those from marginalized communities in the development and deployment

70
00:05:46,470 --> 00:05:52,140
of AI systems can provide valuable insights into potential biases and their impacts.

71
00:05:52,740 --> 00:05:57,900
This collaborative approach can help ensure that AI systems are designed and implemented in a manner

72
00:05:57,900 --> 00:06:00,060
that is inclusive and equitable.

73
00:06:02,190 --> 00:06:07,980
Education and training are also essential in promoting awareness and understanding of AI bias among

74
00:06:07,980 --> 00:06:11,160
developers, policymakers, and the general public.

75
00:06:11,820 --> 00:06:17,880
Incorporating ethics and fairness in AI curricula and professional development programs can equip developers

76
00:06:17,880 --> 00:06:22,350
with the knowledge and skills needed to identify and address biases in their work.

77
00:06:23,190 --> 00:06:28,890
Policymakers, on the other hand, need to be informed about the potential risks and benefits of AI

78
00:06:28,920 --> 00:06:34,110
to create effective regulations and policies that promote fairness and accountability.

79
00:06:35,310 --> 00:06:40,710
Finally, fostering a culture of ethical AI development within organizations is crucial.

80
00:06:41,290 --> 00:06:46,810
This involves creating an environment where ethical considerations are prioritized and integrated into

81
00:06:46,810 --> 00:06:49,450
every stage of the AI development life cycle.

82
00:06:49,930 --> 00:06:55,420
Encouraging a culture of ethical reflection and discussion, supported by organizational policies and

83
00:06:55,420 --> 00:07:00,790
incentives, can help ensure that AI systems are developed and deployed responsibly.

84
00:07:02,530 --> 00:07:08,350
In conclusion, while AI systems have the potential to bring about significant societal benefits, they

85
00:07:08,350 --> 00:07:12,700
also pose risks of discrimination and bias that must be carefully managed.

86
00:07:13,420 --> 00:07:19,990
Ensuring the fairness and equity of AI systems requires a comprehensive and multifaceted approach involving

87
00:07:19,990 --> 00:07:26,650
diverse and representative data, fairness aware algorithms, transparency and accountability, stakeholder

88
00:07:26,650 --> 00:07:31,540
engagement, education and training, and a culture of ethical AI development.

89
00:07:31,570 --> 00:07:37,720
By addressing these challenges, we can harness the power of AI to create a more just and equitable

90
00:07:37,720 --> 00:07:38,590
society.