1 of 42

Human Responsibility

in the Age of AI

Kasia Chmielinski | March 2024

Shared

1

2 of 42

2010

Created with DALL-E

3 of 42

What is my responsibility

as a builder of

AI systems?

4 of 42

Artificial Intelligence

The simulation of human intelligence by machines that are programmed to perform complex cognitive functions

5 of 42

AI is not new ...

1st iPhone

AI is coined as a term

Turing

Test

1st micro-

processor

Hyperlink Invented

Windows �1.0 Launch

2024

2000

1975

1950

Google Launches

Decision Trees

Multilayer Perceptron

Random Forest

Deep �Learning

Models

Back-

propagation

1st home

smoke alarm

1st

floppy disk

1st man on the moon

6 of 42

... nor are our fears ...

“AI Poses 'Risk of Extinction,' Industry Leaders Warn” (NYT, 2023)

Murderous HAL! from 2001: A Space Odyssey (1968)

“Why AI is a Dangerous Dream” (New Scientist, 2009)

“Will man-made robots� rise up and demand their rights?“(MIT, 2000)

Introducing Terminator:

A cyborg assassin (1984)

2024

2000

1975

1950

AI is coined as a term

7 of 42

But AI is getting more powerful, more quickly

2024

2000

1975

1950

1M peta

10k peta

100 peta

1 peta

10T

100B

1B

10M

100k

1000

10

Training Computation �in FLOPS

(floating point �operations per second)

Minerva (2.78B PetaFLOPS)

GPT-3 (314 M PetaFLOPS)

Theseus (40 FLOPS)

AlexNet (470 PetaFLOPS)

Sources: “Compute trends across three eras of machine learning”, by J.Sevilla et al., arXiv, 2022; Our World in Data

... due to increases in computing power

8 of 42

But AI is getting more powerful, more quickly

2025

2020

2015

2010

200

150

100

50

2 Zettabytes

Of Data (2010)

181 Zettabytes of Data (est 2025)

© Statista 2023

Volume of Digitized Data Worldwide

(in Zettabytes of data, 2010-2025)

... due to increases in data availability

9 of 42

Reminiscent of the mobile app landscape,

CLOUD AND COMPUTE SERVICES

AWS, Azure, GPUs, TPUs

FOUNDATION MODELS

GPT, Gemini, Miqu 70B, Llama2, DALL-E, Stable Diffusion

APPLICATIONS

ChatGPT, Copilot, AlphaCode, Scribe, �Synthesia, Sora

APP PLATFORMS

APPLICATIONS

HARDWARE AND CLOUD

Mobile Apps Landscape

Generative AI Apps Landscape

users

Generative AI-powered consumer services are emerging

10 of 42

11 of 42

AI is not new, though it is increasingly powerful and growing rapidly

12 of 42

AI is not new, though it is increasingly powerful and growing rapidly

... and it’s not just what we’re building, but how we’re doing it

13 of 42

Created with DALL-E

14 of 42

Products are market driven, which means we optimize for growth. This means building for the “majority” at the expense of others

14

Build for the “ideal user” and iterate out over time

  • People who don’t �speak English

  • People who don’t have the most recent iphone

  • People in rural areas
  • People who don’t �have binary gender �
  • People who use assistive devices or technology

  • People who share phones or devices with family

80% solution

for “majority”

15 of 42

Frustratingly, most of the focus (hype) is on �existential future risks rather than actual harms

16 of 42

Yes, and.... we need to focus on

the consolidation of power and control in the hands of the few

17 of 42

18 of 42

19 of 42

AI is not new, though it is increasingly powerful and growing rapidly

... and it’s not just what we’re building, but how we’re doing it

20 of 42

The real risks are products

that fail communities at scale

And the further consolidation of power in the hands of the few

21 of 42

2017

Created with DALL-E

22 of 42

2017

23 of 42

23

AI Model Pipeline

Systems built on �problematic data will �exhibit those same issues, �especially for historically �marginalized people

24 of 42

24

Systems built on �problematic data will �exhibit those same issues, �especially for historically �marginalized people

AI Model Pipeline

Shouldn’t we interrogate for issues before we build?

25 of 42

The Dataset Nutrition Label

A standardized documentation tool that tells you what’s in a dataset and whether it’s healthy for your model

26 of 42

26

The label includes at-a-glance information about key critical aspects of the dataset as well as usage information and known risks by category.

27 of 42

“... While I know that the primary mission of the DNP is to improve the understanding, searching, and consumption of datasets by users of datasets, it has also been key to improving my dataset design moving forward.”

— Dataset Partner

28 of 42

AI Systems are socio-technical

Within the full system, there are many sites of risk for bias and therefore many sites of intervention

(Partnership on AI, 2022)

29 of 42

AI Systems are socio-technical

Within the full system, there are many sites of risk for bias and therefore many sites of intervention

(Partnership on AI, 2022)

30 of 42

What is our responsibility

your opportunity

as a user of AI systems?

31 of 42

Tip 1:�

Think about AI

as a process, not a product

31

32 of 42

When you are �procuring systems...

Ask about how the model was tested

Ask about how the model will be monitored and updated

Ask about the training data

Ask about the success criteria for the model

Ask about criteria for decommission

Ask whether this is the right problem to address with AI

33 of 42

Tip 2:�

Think about AI as another tool in your toolkit, not something totally new

33

34 of 42

Look to existing policies ...

Attribution and disclosure policies (plagiarism, copyright)

New tools requirements (security, confidentiality)

Freeware vs. Enterprise software (licenses, support, security)

Use policies (calculators, digital editing, photoshop)

35 of 42

Tip 3:�

Think about AI in the context of

new cultural norms

35

36 of 42

Cultural Norms

New technology requires a shift in norms

around how we use and expect others to use the tools

Example:

Using AI to Create

Internal vs. External material

Keep a human in the loop �for anything published or launched externally

Example:

Care with exposing confidential information

Do not put any confidential information into Generative AI tools because they may retain (and regurgitate) this information for other users

Example:

Community to Share

Challenges & Opportunities

You’re not alone! Find others who are navigating similar changes in your domain

37 of 42

https://aipedagogy.org

38 of 42

AI is not new, and neither are �our concerns about it.

But what is changing is the

scale and power of these systems

Compounded with �how we build product

39 of 42

Means that the real risk are

products that fail communities

And the consolidation of the power

in the hands of the few.

40 of 42

We need to remember that humans build the technology

If we remember that AI is a process, �we - and you! -

can meaningfully intervene

41 of 42

Thank you!

Kasia Chmielinski | March 2024

42 of 42

Slides:

Or get in touch!

kc@datanutrition.org