1
00:00:00,050 --> 00:00:04,190
Lesson requirements for high risk AI systems and foundation models.

2
00:00:04,220 --> 00:00:09,770
High risk AI systems and foundation models necessitate stringent requirements to ensure they're safe,

3
00:00:09,800 --> 00:00:11,840
ethical, and effective deployment.

4
00:00:12,650 --> 00:00:19,010
The term high risk AI systems often refers to AI applications that can significantly impact individuals

5
00:00:19,010 --> 00:00:25,100
lives, such as those used in health care, finance, criminal justice, and autonomous vehicles.

6
00:00:25,670 --> 00:00:31,220
Foundation models, such as large language models serve as the basis for numerous applications, making

7
00:00:31,220 --> 00:00:35,750
their governance critical for ensuring broad utility without undue harm.

8
00:00:36,530 --> 00:00:43,070
Governance frameworks for high risk AI systems must prioritize transparency, accountability, and ethical

9
00:00:43,070 --> 00:00:44,150
considerations.

10
00:00:44,750 --> 00:00:50,480
Transparency involves clear documentation regarding the data used for training, the algorithms employed,

11
00:00:50,480 --> 00:00:53,720
and the decision making processes of the AI system.

12
00:00:54,170 --> 00:00:59,570
This documentation should be accessible to stakeholders, enabling them to understand the AI's functioning

13
00:00:59,570 --> 00:01:02,650
and to identify potential biases or errors.

14
00:01:03,250 --> 00:01:09,040
Accountability requires that developers and deployers of high risk AI systems be liable for their outcomes,

15
00:01:09,070 --> 00:01:13,780
necessitating robust mechanisms for monitoring and rectifying harmful impacts.

16
00:01:14,410 --> 00:01:20,770
Ethical considerations involve adherence to principles such as fairness, privacy, and non-discrimination,

17
00:01:20,770 --> 00:01:26,020
ensuring that the AI systems do not perpetuate or exacerbate social inequalities.

18
00:01:27,550 --> 00:01:33,370
Moreover, high risk AI systems should undergo rigorous testing and validation before deployment.

19
00:01:33,760 --> 00:01:39,880
This includes performance evaluations under various scenarios to ensure reliability and robustness.

20
00:01:40,270 --> 00:01:46,240
For instance, in healthcare, an AI diagnostic tool must be tested across diverse patient populations

21
00:01:46,240 --> 00:01:48,610
to confirm its accuracy and fairness.

22
00:01:48,970 --> 00:01:55,150
Regulatory bodies may require third party audits and certifications to verify compliance with established

23
00:01:55,150 --> 00:01:55,870
standards.

24
00:01:55,900 --> 00:02:00,640
Such processes are vital to building trust among users and stakeholders.

25
00:02:01,530 --> 00:02:07,380
Foundation models, due to their extensive applications, require comprehensive governance to safeguard

26
00:02:07,380 --> 00:02:08,670
against misuse.

27
00:02:09,270 --> 00:02:14,970
These models are trained on vast data sets, often sourced from the internet, which can introduce biases

28
00:02:14,970 --> 00:02:16,830
and propagate harmful content.

29
00:02:17,400 --> 00:02:23,610
Developers must implement robust data curation practices to mitigate these risks, ensuring that training

30
00:02:23,610 --> 00:02:27,210
data is representative and free from inappropriate material.

31
00:02:27,660 --> 00:02:32,940
Additionally, techniques such as differential privacy can be employed to protect individual data points

32
00:02:32,940 --> 00:02:36,360
within the training data sets, bolstering privacy protections.

33
00:02:38,790 --> 00:02:44,220
The deployment of foundation models also necessitates clear usage guidelines to prevent misuse.

34
00:02:44,760 --> 00:02:50,250
For instance, guidelines can stipulate acceptable use cases and prohibit applications that could cause

35
00:02:50,250 --> 00:02:53,760
harm, such as generating deepfakes or misinformation.

36
00:02:54,210 --> 00:02:59,880
Developers should also provide tools for users to interpret and control the outputs of foundation models.

37
00:02:59,880 --> 00:03:08,160
Enhancing transparency and user agency regulatory frameworks for high risk AI systems and foundation

38
00:03:08,160 --> 00:03:12,450
models must be adaptive, keeping pace with technological advancements.

39
00:03:12,990 --> 00:03:18,720
Policymakers should engage with AI experts, stakeholders, and the public to develop regulations that

40
00:03:18,720 --> 00:03:21,120
balance innovation with safeguards.

41
00:03:21,510 --> 00:03:27,510
For example, the European Union's AI act proposes a risk based approach to AI regulation, categorizing

42
00:03:27,540 --> 00:03:32,790
AI applications by their potential impact and imposing corresponding requirements.

43
00:03:32,850 --> 00:03:38,790
High risk AI systems would be subject to stringent obligations, including conformity assessments and

44
00:03:38,790 --> 00:03:42,180
continuous monitoring, to ensure their safe deployment.

45
00:03:43,380 --> 00:03:49,500
These statistics underscore the critical need for robust governance of high risk AI systems and foundation

46
00:03:49,500 --> 00:03:50,340
models.

47
00:03:50,520 --> 00:03:52,470
A study by Obermaier et al.

48
00:03:52,500 --> 00:03:57,870
Revealed that a widely used health care algorithm exhibited racial bias, underscoring the potential

49
00:03:57,900 --> 00:04:00,590
harms of unregulated AI systems.

50
00:04:00,860 --> 00:04:02,480
Similarly, Bender et al.

51
00:04:02,510 --> 00:04:08,210
Highlighted the risks associated with large language models, including the amplification of biases

52
00:04:08,210 --> 00:04:10,400
and the generation of harmful content.

53
00:04:10,820 --> 00:04:16,340
These findings emphasize the importance of rigorous oversight to prevent adverse outcomes.

54
00:04:17,210 --> 00:04:23,270
Examples of successful governance frameworks illustrate the feasibility of implementing robust requirements

55
00:04:23,270 --> 00:04:26,750
for high risk AI systems and foundation models.

56
00:04:27,500 --> 00:04:32,390
The Food and Drug Administration in the United States has established guidelines for the approval of

57
00:04:32,390 --> 00:04:38,510
AI based medical devices, ensuring their safety and effectiveness before they reach patients.

58
00:04:39,230 --> 00:04:45,620
These guidelines include requirements for clinical validation, transparency of algorithms, and post-market

59
00:04:45,620 --> 00:04:46,550
surveillance.

60
00:04:46,880 --> 00:04:53,630
Providing a comprehensive framework for AI governance in health care in the private sector, companies

61
00:04:53,630 --> 00:04:58,790
like Google have developed internal guidelines for the ethical development and deployment of AI.

62
00:04:59,480 --> 00:05:05,350
Google's AI principles emphasize fairness, privacy, and accountability, guiding the company's approach

63
00:05:05,350 --> 00:05:08,650
to AI governance by adhering to these principles.

64
00:05:08,680 --> 00:05:14,440
Google aims to ensure that its AI technologies are developed responsibly and used for the benefit of

65
00:05:14,440 --> 00:05:15,220
society.

66
00:05:17,470 --> 00:05:24,100
In conclusion, the requirements for high risk AI systems and foundation models are multifaceted, encompassing

67
00:05:24,130 --> 00:05:30,370
transparency, accountability, ethical considerations, rigorous testing, and adaptive regulation.

68
00:05:30,970 --> 00:05:36,580
These requirements are essential to mitigate the risks associated with powerful AI technologies, and

69
00:05:36,580 --> 00:05:43,000
to harness their potential for societal benefit through collaborative efforts among developers, regulators,

70
00:05:43,000 --> 00:05:44,110
and stakeholders.

71
00:05:44,140 --> 00:05:50,650
It is possible to create a governance framework that fosters innovation while safeguarding against harm.

72
00:05:51,490 --> 00:05:56,590
The integration of robust oversight mechanisms, clear guidelines and continuous engagement with the

73
00:05:56,590 --> 00:05:59,830
public will be crucial in achieving this balance.