1 of 35

Freedom & AI

Can Free Software include ethical AI systems?

Mike Nolan

UNICEF Office of Innovation @ Openmined�@__nolski__

[Justin]

In 2018, Victoria Krakovna of Google's DeepMind AI lab asked her colleagues at Google for examples of misbehaving AI

Not to lose Tetris? Pause the game
Self-driving car simulator tasked to keep cars “fast and safe”: spinning in circles on the spot
Artificial life simulation to simulate evolution: species that has a sedentary lifestyle that mates to produce new children which could be eaten (or used as mates to produce more edible children)

Some of these are funny, but this is becoming our world

Major social network site serves you ads for cat food after you post about getting a cat [Facebook]
Major U.S. retailer store knows you’re pregnant before you tell your friends and family [Target]

What happens when these stories turn into:

Major job-matching site knows you are afraid to ask for a raise [Indeed]
Major online retailer knows you are an impulse buyer [Amazon]

The rest of the world is taking notice too:

On January 20, 2020 (last week): CEO of Google called on governments to look at regulation for AI

https://www.dailymail.co.uk/sciencetech/article-7907373/Google-CEO-calls-regulation-AI-protect-against-deepfakes-facial-recognition.html

2 of 35

Goodhart’s Law

wikipedia.org/wiki/Goodhart's_law

“When a measure becomes a target,�it ceases to be a good measure.”

src: gph.is/2h08OFE

@__nolski__

3 of 35

What we want to solve

AI systems are being built, but there are not clear actionable steps to mitigate risks of infringing on human rights.

@__nolski__

4 of 35

What this talk is about

What freedoms do we want to protect?
How do we respect these freedom?
How to ensure others respect our freedom?

@__nolski__

5 of 35

1.�What freedoms do we want to protect?

@jflory7 @__nolski__

6 of 35

First… let’s travel back�to 1983.

src: gph.is/28RcCUT gph.is/1LofyTc

@jflory7 @__nolski__

7 of 35

27 Sept. 1983: GNU Project announced

src: gnu.org/graphics/heckert_gnu.html

@jflory7 @__nolski__

8 of 35

4 Oct. 1985: Free Software Foundation founded

src: gnu.org/graphics/fsf-logo.en.html

@jflory7 @__nolski__

9 of 35

25 Feb. 1989: GNU General Public License created

src: gnu.org/graphics/license-logos.html

@jflory7 @__nolski__

10 of 35

Is the past relevant today?

src: gph.is/2ddT0yp gph.is/1m0ai2O

@jflory7 @__nolski__

11 of 35

Your future is the new commodity.

@jflory7 @__nolski__

12 of 35

Where are�we today?

@jflory7 @__nolski__

13 of 35

AI Now Institute / The Partnership on AI

ainowinstitute.org

partnershiponai.org

@jflory7 @__nolski__

14 of 35

What freedoms do we want to protect?

src: gph.is/2iYztAJ

@jflory7 @__nolski__

15 of 35

Audit automated decision-making systems

1. Freedom to…

@jflory7 @__nolski__

16 of 35

Trolley problem

src: jrtechs.net/other/morality-of-self-driving-cars

@jflory7 @__nolski__

[Mike]

Self-driving cars (trolley problem)

Self-driving cars are already programmed to make decisions involving human life

How are decisions are made when human life is at risk?

We have the right to know how those decisions are made
This is technology that directly impacts us and our livelihoods

Study asked people if a car had to choose between a tree or run into a child, which choice would the car make?

~90% of responses said tree

Next question: Would you ride in a car that chose the tree?

Even though most people want self driving cars to hit the tree, most people responded that they would not buy a car which did not have their safety as its top priority.

Questions to consider:

Would it be ethical for a car to take into account you age, race, gender, and social status when deciding if you get to live?
If self-driving cars could access personal information such as criminal history or known friends, would it be ethical to use that information?
Would it be moral for someone to make a car which favored the safety of the passengers of the car above others?

Maybe… we ask too much of AI systems to make these decisions for us.

17 of 35

Deliver accountability and responsibility

2. Freedom to…

@jflory7 @__nolski__

18 of 35

We deserve to be put over their profit.

@jflory7 @__nolski__

19 of 35

reuters.com/article/us-myanmar-rohingya-un/myanmar-generals-had-genocidal-intent-against-rohingya-must-face-justice-u-n-idUSKCN1LC0KN

@jflory7 @__nolski__

[Mike]

You may be familiar with the Rohingya genocide in Myanmar.

What was the role of Facebook in this?

News feed optimizes on “engaging content”
What is engaging content?

Studies shown optimizing for engagement increases recommendations for extremist/alarming content

Researchers knew of these issues and called them out years before the Rohingya genocide began in Myanmar
Is Facebook responsible for “fake news” propped up by religious and military leaders that contributed to an ethnic cleansing of Rohingya people?

They didn’t explicitly think about causing genocide when building a feature
But it was still a contributing factor
In some way or another, they did know about it.

Who is to blame? Facebook? Nobody?

Nobody will accept responsibility, even today.
What we know: Profits were placed over people.

20 of 35

Appeal a decision

3. Freedom to…

@jflory7 @__nolski__

[Justin]

Transition

Third: No decision-making system is ever perfect. We are always missing some data.
So, you should also have the freedom to appeal a decision that impacts you.

Abstract

Have you ever told a story to help another person empathize with your situation? Have you ever had to explain a misconstrued fact about your background or history?

“Walk in someone else’s shoes?”
Our ability to do this is what connects us as humans and helps avert disaster more often than you think.

There are always hidden stories not captured by a set of data points.
We should always have the opportunity to break through automated systems that influence an organization and people behind the systems to use our humanity

Examples

[ask audience] By show of hands, how many of you have a university degree?

So many of you have applied for jobs.
When you apply, you submit your CV for the application.

Automated tools already review CVs to filter out candidates.
But what is the point of interviews? With a real person?

Interviews are a chance to tell our own hidden stories and tell the gap between what is on our CV and what isn’t
It gives us a chance to build empathy between us and who we want to be our employer

Where else?

U.S. appellate court

Unfair trial / biased judge
System in place to appeal

Rejection for loans / credit offers

Appeal to impartial third-party
System in place to appeal

Outro

So, to wrap up, what is the idea behind this Freedom?
>> Next slide

21 of 35

We must not erase the opportunity for empathy.

@jflory7 @__nolski__

22 of 35

~~ soooome sort of transition with cool GIF ~~

@jflory7 @__nolski__

23 of 35

2.�How do we respect these Freedoms?

@jflory7 @__nolski__

24 of 35

Reproducibility

arxiv.org/abs/1810.03993

@jflory7 @__nolski__

[Mike]

We like to make claims that models are fool-proof and avoid bias

How do we prove that with proprietary data? Or a subset of real data?
How can we verify the claims of accuracy made by creators of AI systems?

Reproducibility is about audibility

Connects back to 1st Freedom: freedom to audit automated decision-making systems

To verify, we need three things to reproduce results from a model:

Access to training data (or source of training data)
Access to source code
Access to proper documentation describing the model

Might be most critical piece!

Documentation is an equalizer.
Users have a right to reasonable documentation detailing the functionality of the AI system they’re subjected to.
This is general language, but there are key steps we can take to move this forward.
AI Now has made progress standardizing this sort of documentation for models

This isn’t perfect… but we can design these requirements in an actionable way.

25 of 35

Liability & responsible design

@jflory7 @__nolski__

26 of 35

Appealing mechanisms

@jflory7 @__nolski__

[Justin]

Like we mentioned before, we already have appealing mechanisms in our existing systems

No secret sauce / new ideas to implement these
Start with things we already have as a model

Appealing mechanisms: Human-controlled methods to appeal a decision and correct mistakes made by automated systems, common among other services.
Reiterate past examples:

Courts
Bank loans

Even if an automated decision-making has input on a decision, appealing mechanisms are still human-centered and human-controlled

Appealing mechanism should be designed to create empathy between humans

What might it look like for AI?

Manual overrides
Example: Tesla auto-pilot; once you touch the wheel, brake, or gas pedal, you immediately override the automated decisions
You “appeal” the decisions the car would make for you

[transition] What happens when appealing mechanisms are poorly designed?

27 of 35

What happens without manual overrides?

src: usatoday.com

@jflory7 @__nolski__

28 of 35

Want to build a movement? Define�a standard.

@jflory7 @__nolski__

29 of 35

Luckily, Free Software advocates are pretty good at this.

@jflory7 @__nolski__

WHATWG

30 of 35

3.�How to ensure others respect our Freedoms?

@jflory7 @__nolski__

31 of 35

1.�Licensing

@jflory7 @__nolski__

[Mike]

Transition

When researching the history of Free software, we naturally gravitated towards using licencing as a model.
Because AI systems extend beyond simply the model but also the data and even documentation, this became tricky.

Abstract

IANAL but we suspect licencing could be a strategy one could use to ensure derivatives are developed in a sustainable fashion by developing comprehensive packaged license solution for code & data sets.
This has a few pros/cons

License proliferation

Yet another software license. This may not even be Free software at this point

Legal & litigatory overhead

Places strain on the original software creator to protect.

Shifts power into activist hands

Gives you an immediate tool you can use as a creator of AI software to incentivize others to respect the same things you do.

Examples

Open Data Licensing + Open Street Maps
CDLA - Data license

Not perfect

32 of 35

2.�Regulation

@jflory7 @__nolski__

33 of 35

3.�Independent certification

@jflory7 @__nolski__

[Justin]

Transition

The difference of knowledge between consumers and producers allows for exploitation by producers.
Third-party, independent certification bodies are a classic solution to simplifying complex information for consumers to make smart decisions.

Abstract

A certifying authority could allow creators of AI systems which respect our Freedoms to easily earn consumer trust.
There are some pros/cons to this

Empowers consumers
(but doesn’t protect).
Empowers competition

MUST BE RUN BY 3rd PARTY

Examples

B-Corp: Certifying authority to help ensure corporations have a positive impact on their workers, community, customers and environment.
USDA Organic: Certifying authority to ensure that food producers do not use chemicals, growth hormones, etc. in their product.
Respect your freedoms certification: FSF

Lulzbot
Librem - purism

34 of 35

4.�Labor Organization

@jflory7 @__nolski__

[Justin]

Transition

Lastly, what has historically been our strongest tool as workers is to organize.
Labor organization is a powerful tool workers can use to ensure their freedoms are respected.

Abstract

More tech workers are noticing the negative impacts their AI systems create
A labor union’s ability to affect a workforce’s employment contract incentives your company to take notice and respect the wishes of their employees.
A union can transition power from shareholder’s hands into the hands of creators, you, who are creating the real value.

Examples

The Amazon Employees for Climate Justice are organizing against’ Amazon’s inaction towards climate change
Github employees are protesting and threatening to strike over their company’s deal with America’s violently xenophobic immigration enforcement department
Google already threatened by employees trying to unionize that they have been actively deploying expensive countermeasures to sew doubt in the power of labor organization

[transition] The power of labor organization is being respected by these organizations

[next slide]

35 of 35

Freedom & AI

Let’s build a future that we own, together, for everyone.

freeourfuture.world

Justin W. Flory

Rochester Institute of Technology�@jflory7

Mike Nolan

UNICEF Office of Innovation�@__nolski__