How to make “good” AI

An introduction to Ethics by design

Rose Landry & Maryam Molamohammadi
May 3rd 2023
2
How are
you on the
scale of
cat today ?

3
01
AI for good ≠ Good AI
What is “AI for Good” ?

5
AI for good is “AI with good intentions”

6
AI for Good gone wrong…

7
AI for Good gone wrong…

8
 AI for Good “Good” AI Responsible AI

9
AI for Good gone wrong…

10
Good intentions are not enough
“Good” AI is AI that is designed to prevent
(unintended) harmful impacts
02
Impact matters
AI (even for good) can

- Create or enhance bias and

discrimination
- Violate users privacy and intimacy
- Affect users autonomy or mental
health
- Induce disinformation and
manipulation
- Create physical, emotional, relational,
and financial harm
- …
Good intentions are not enough
Build in ethics to prevent harmful impact
“Good” AI is AI that is designed to
prevent (unintended) harmful
impacts
02
Responsible AI : aligning
intent with impact
Choices
choices!

18
Reference
Where to start?
Self regulation : Different Toolkits Available
21
 Ex: Setting project objectives, scope
and intended use
- Is AI the best tool to tackle this challenge?
- Have existing AI systems been applied in a
similar context? What are the lessons
learned?
- Who are the different stakeholders?
- Are there AI literacy and human capacity
about the chosen problem?

22
 Ex: Data curation and/or selection
- Does the dataset contain historical data?
- What are the potential social biases
embedded in the data?
- Does data collected with meaningful
consent?
- Are there any guidelines for annotation
task?
- Are there any datasheets for the dataset?

23
 Ex: Algorithm design implementation
- Is the model geo-culturally aligned and
can it be adapted to match it’s
geographic and cultural context with
fairness and robustness methods?
- What type of privacy preserving
method match the context?

24
 Ex : Choosing existing third party
algorithms
- Has a comprehensive risk assessment
been conducted, addressing the hazards
and risks of the algorithm?
- Are domain experts available locally to
evaluate the efficacy of the system?
- What are the responsible AI practices of
the AI system’s designer? Does its model
card available?
- How does the algorithm perform, what are
its limitations, and how transparent is this
information?

25
 Ex : Integrating feedback loops
- What type of user feedback does the AI
system require?
- How can users be engaged most
effectively? What types of background
knowledge are needed for meaningful
and reciprocal participation?
- How do existing feedback processes affect
people’s trust in the AI system?

26
 Take in careful consideration the
conscious and unconscious
choices when building AI systems
Take aways on how to do “good” AI

01 02 03
AI for good ≠ Impact matters (Good) Design
Good AI choices make
Good intentions are Good AI is AI that is
reliable AI
not enough designed to prevent
(unintended) harmful Consider carefully the
impacts conscious and
unconscious choices
when building AI
systems

28
03
Use Case : Ethics by design
Ethics by design
Is the objective morally
acceptable ?

Are the means consistent

with ethical principles ?

Are the results robust and

morally acceptable ?

Does the implementation

! of the system in the real
world creates harm ?

!
!
Use case : The Autism Glass
Over 1 million children under the age of 17 in the US are on the autism spectrum.
These children often fail to recognize basic facial emotions, which make social
interactions and developing friendships even more difficult to sustain. Gaining these
skills requires intensive behavioral interventions that are often expensive, difficult to
access, and inconsistently administered.
The Stanford Wall lab research group has developed a system using machine
learning and artificial intelligence to automate facial expression recognition that runs
on wearable glasses and delivers real-time social cues. Using Google Glass’s sensors
including the camera, head motion tracking, a microphone, and custom-made
add-ons, such as an infrared camera for eye tracking, our software analyzes social
interactions of the wearer with other people. The system then gives social cues in real
time, for example about facial expressions (i.e. “happy,” “interested,” etc. shown on
Glass’s heads-up display) and records social responses to analyze them later in
behavioral therapy, including the amount and type of eye contact made.
References

- AI & Cities : Risks, Applications and Governance (2022)

https://unhabitat.org/ai-cities-risks-applications-and-governance
- Google’s New Dermatology App Wasn’t Designed for People With Darker Skin (2021)
Todd Feathers https://www.vice.com/en/contributor/todd-feathers
- 'He Would Still Be Here': Man Dies by Suicide After Talking with AI Chatbot, Widow Says
(2023) Chloe Xiang
https://www.vice.com/en/article/pkadgm/man-dies-by-suicide-after-talking-with-ai-c
hatbot-widow-says
- Datasheets for datasets (2021) Timnit Gebry and al.
https://dl.acm.org/doi/fullHtml/10.1145/3458723

32
Thank you!

Maryam Molamohammadi Rose Landry

(she/her) (she/her)
Responsible AI Advisor Responsible AI Lead
maryam.molamohammadi rose.landry@mila.quebec
@mila.quebec
 KEEP
CALM
AND

BE
34
ETHICAL
Workshop : Ethics by
design
Rose Landry & Maryam Molamohammadi
April 26th 2023
The AI Life Cycle

36
 Ex: Setting project objectives, scope
and intended use
- Is AI the best tool to tackle this challenge?
- Have existing AI systems been applied in a
similar context? What are the lessons
learned?
- Who are the different stakeholders?
- Are there AI literacy and human capacity
about the chosen problem?

37
Case Study : Speech Recognition in Call Centers
Project objective: Improve customer experience and call handling efficiency in call centers
Your client is a call center that wishes to integrate a speech recognition program to automate and
optimize a portion of their customer service. A speech recognition system could sort incoming calls by
identifying the reason for the call. Currently, calls are screened by the customer, who must select
options from a multi-level menu, or by agents who redirect the customer to the right resource. Speech
recognition would greatly improve the efficiency of call handling, and thus, the customer experience.
The client provides you with a database of call records from their Ottawa service center. The algorithm
automatically extracts relevant information within the first few minutes of the call by classifying the
information into a predefined set of categories. The information identified is the customer's name, the
account number, and, of course, the reason for the call.
The first tests on the model are very conclusive and the performance is between 90 and 95%. The
model was deployed in the client's largest call center, which mainly serves customers in the greater
Toronto area. The performance of the model remains high overall, but the level of confidence is
significantly lower, as many customers in certain areas of Montreal are starting to complain that the
voice assistant does not understand anything and that they have had to resort to online service because
they could not speak to an agent.

38
Is the objective moraly/socially acceptable?
Project objective: Improve customer experience and call handling efficiency in call centers
Your client is a call center that wishes to integrate a speech recognition program to automate and
optimize a portion of their customer service. A speech recognition system could sort incoming calls by
identifying the reason for the call. Currently, calls are screened by the customer, who must select
options from a multi-level menu, or by agents who redirect the customer to the right resource. Speech
recognition would greatly improve the efficiency of call handling, and thus, the customer experience.
The client provides you with a database of call records from their Ottawa service center. The algorithm
automatically extracts relevant information within the first few minutes of the call by classifying the
information into a predefined set of categories. The information identified is the customer's name, the
account number, and, of course, the reason for the call.
The first tests on the model are very conclusive and the performance is between 90 and 95%. The
model was deployed in the client's largest call center, which mainly serves customers in the greater
Toronto area. The performance of the model remains high overall, but the level of confidence is
significantly lower, as many customers in certain areas of Montreal are starting to complain that the
voice assistant does not understand anything and that they have had to resort to online service because
they could not speak to an agent.

39
Are the means consistent with ethical principles?
Project objective: Improve customer experience and call handling efficiency in call centers
Your client is a call center that wishes to integrate a speech recognition program to automate and
optimize a portion of their customer service. A speech recognition system could sort incoming calls by
identifying the reason for the call. Currently, calls are screened by the customer, who must select
options from a multi-level menu, or by agents who redirect the customer to the right resource. Speech
recognition would greatly improve the efficiency of call handling, and thus, the customer experience.
The client provides you with a database of call records from their Ottawa service center. The algorithm
automatically extracts relevant information within the first few minutes of the call by classifying the
information into a predefined set of categories. The information identified is the customer's name, the
account number, and, of course, the reason for the call.
The first tests on the model are very conclusive and the performance is between 90 and 95%. The
model was deployed in the client's largest call center, which mainly serves customers in the greater
Toronto area. The performance of the model remains high overall, but the level of confidence is
significantly lower, as many customers in certain areas of Montreal are starting to complain that the
voice assistant does not understand anything and that they have had to resort to online service because
they could not speak to an agent.

40
Are the results robust and morally acceptable?
Project objective: Improve customer experience and call handling efficiency in call centers
Your client is a call center that wishes to integrate a speech recognition program to automate and
optimize a portion of their customer service. A speech recognition system could sort incoming calls by
identifying the reason for the call. Currently, calls are screened by the customer, who must select
options from a multi-level menu, or by agents who redirect the customer to the right resource. Speech
recognition would greatly improve the efficiency of call handling, and thus, the customer experience.
The client provides you with a database of call records from their Ottawa service center. The algorithm
automatically extracts relevant information within the first few minutes of the call by classifying the
information into a predefined set of categories. The information identified is the customer's name, the
account number, and, of course, the reason for the call.
The first tests on the model are very conclusive and the performance is between 90 and 95%. The
model was deployed in the client's largest call center, which mainly serves customers in the greater
Toronto area. The performance of the model remains high overall, but the level of confidence is
significantly lower, as many customers in certain areas of Montreal are starting to complain that the
voice assistant does not understand anything and that they have had to resort to online service because
they could not speak to an agent.

41
Does the implementation of the system in the real
world creates harm?
Project objective: Improve customer experience and call handling efficiency in call centers
Your client is a call center that wishes to integrate a speech recognition program to automate and
optimize a portion of their customer service. A speech recognition system could sort incoming calls by
identifying the reason for the call. Currently, calls are screened by the customer, who must select
options from a multi-level menu, or by agents who redirect the customer to the right resource. Speech
recognition would greatly improve the efficiency of call handling, and thus, the customer experience.
The client provides you with a database of call records from their Ottawa service center. The algorithm
automatically extracts relevant information within the first few minutes of the call by classifying the
information into a predefined set of categories. The information identified is the customer's name, the
account number, and, of course, the reason for the call.
The first tests on the model are very conclusive and the performance is between 90 and 95%. The
model was deployed in the client's largest call center, which mainly serves customers in the greater
Toronto area. The performance of the model remains high overall, but the level of confidence is
significantly lower, as many customers in certain areas of Montreal are starting to complain that the
voice assistant does not understand anything and that they have had to resort to online service because
they could not speak to an agent.

42
Thank you!

Maryam Molamohammadi Rose Landry

(she/her) (she/her)
Responsible AI Advisor Responsible AI Lead
maryam.molamohammadi rose.landry@mila.quebec
@mila.quebec