AI and bias
Oleksandr Krakovetskyi
CEO DevRain, devrain.com
CTO DonorUA, donor.ua
Ph.D. in Computer Science
Microsoft Regional Director
Microsoft Artificial Intelligence Most Valuable Professional
Agenda
1. Types of bias
2. Research results
3. Bias in job postings
4. Gender neutral language
5. How to avoid bias?
6. Discussion and Q&A
Nonverbal Bias
We prefer to scrap
own opinion in
favour of the
groups’ opinion.
Beauty Bias
This is the view that we
tend to think that the
most handsome individual
will be the most
successful.
Affinity Bias
Affinity bias occurs when we
see someone we feel we have
an affinity with e.g. we
attended the same college, we
grew up in the same town, or
they remind us of someone we
know and like.
Halo/Horns Effect
Halo is when we see one
great thing about a person
and we let the halo glow of
that significant thing affect
our opinions of everything
else about that person.
Similarity Bias
Naturally, we want to
surround ourselves with
people we feel are similar to
us. And as a result, we tend
to want to work more with
people who are like us.
Contrast Effect
We should be comparing are
the skills and attributes each
individual has, to the skills and
attributes required for the job,
not those of the person that
came directly before them.
Attribution Bias
When we do something badly we tend
to believe that our failing is down to
external factors like other people that
adversely affected us and prevented us
from doing our best.
When it comes to other people, we
tend to think the opposite.
Confirmation Bias
When we make a judgement
about another person, we
subconsciously look for evidence
to back up our own opinions of
that person. We do this because
we want to believe we’re right
and that we’ve made the right
assessment of a person.
Conformity Bias
Occurs when a positive or
negative evaluation is
made of someone based
on their body language,
personal appearance or
style of dress.
Dunning–Kruger effect
Example #1
Women are less like to apply for jobs that have a very long list
of ’desirable’ qualities, as they do not wish to waste the
employer's time if they are not perfectly suited to the role.
Women are also less likely to shout about their
achievements, and to negotiate salaries.
Example #2
Researchers from Harvard and Princeton found that blind
auditions increased the likelihood that female musicians would
be hired by an orchestra by 25 to 46 percent.
https://www.forbes.com/sites/pragyaagarwaleurope/2019/02/20/how-to-minimize-unconscious-bias-during-recruitment/
Example #3
The researchers found that 70% (possibly even higher) of
hidden biases are directed towards African-Americans, the
elderly, the disabled, and overweight individuals.
What evidence do you have for the claim
that the wording of job ads has any effect
on the people reading them?
The evidence underlying this tool comes from a research paper written by
written by Danielle Gaucher, Justin Friesen, and Aaron C. Kay, called
Evidence That Gendered Wording in Job Advertisements Exists and
Sustains Gender Inequality. It was published in the Journal of
Personality and Social Psychology, July 2011, Vol 101(1), p109-28.
https://www.hw.ac.uk/documents/gendered-wording-in-job-ads.pdf
Job interview
1. The language in job adverts affects the type of people
who apply.
2. The fact is Latisha and Jamal do not get the same
number of callbacks as Emily and Greg.
3. Structured interviews “standardize the interview
process” and “minimize bias” by allowing employers to
“focus on the factors that have a direct impact on
performance.”
Job interview
1. Employers often use the last person to hold a post as a
benchmark for the type of candidate they’re looking for.
2. The growing use of machine learning algorithms to help
identify potential employees is leading to new types of bias.
3. Some people say they have no bias against anything, and
that’s laughable.
http://www.bbc.com/capital/story/20180806-how-hidden-bias-can-stop-you-getting-a-
job
Gender-neutral language
Masculine Feminine Gender
neutral
man woman person
father mother parent
boy girl child
uncle aunt
husband wife spouse
actor actress
prince princess
waiter waitress server
rooster hen chicken
stallion mare horse
Everyone loves ____ mother.
http://www.english.illinois.edu/-people-/faculty/debaron/essays/epicene.htm
Example: Each participant must
present his ID badge at the door.
Revised: All participants must
present their ID badges at the door.
Example: The tenant must
keep his apartment clean and tidy.
Revised: You must keep your apartment
clean and tidy.
Example: The client should
receive his invoice in two weeks.
Revised: The client should receive his or
her invoice in two weeks.
Test your job description:
http://gender-decoder.katmatfield.com/
Where and when bias appears?
Is AI a Racist?
How to avoid bias?
1. Rules vs. data (AlphaGo)
2. Give choice (Google Translate)
3. Be neutral
4. Learn and understand your biases
Digital Transformation Conference
June 22, Kyiv, Microsoft Ukraine
https://dtconf.com
Q&A
Oleksandr Krakovetskyi
alex.krakovetskiy@devrain.com
@sashaeve
fb.com/alex.krakovetskiy

Artificial Intelligence and Bias

  • 1.
    AI and bias OleksandrKrakovetskyi CEO DevRain, devrain.com CTO DonorUA, donor.ua Ph.D. in Computer Science Microsoft Regional Director Microsoft Artificial Intelligence Most Valuable Professional
  • 2.
    Agenda 1. Types ofbias 2. Research results 3. Bias in job postings 4. Gender neutral language 5. How to avoid bias? 6. Discussion and Q&A
  • 3.
    Nonverbal Bias We preferto scrap own opinion in favour of the groups’ opinion.
  • 4.
    Beauty Bias This isthe view that we tend to think that the most handsome individual will be the most successful.
  • 5.
    Affinity Bias Affinity biasoccurs when we see someone we feel we have an affinity with e.g. we attended the same college, we grew up in the same town, or they remind us of someone we know and like.
  • 6.
    Halo/Horns Effect Halo iswhen we see one great thing about a person and we let the halo glow of that significant thing affect our opinions of everything else about that person.
  • 7.
    Similarity Bias Naturally, wewant to surround ourselves with people we feel are similar to us. And as a result, we tend to want to work more with people who are like us.
  • 8.
    Contrast Effect We shouldbe comparing are the skills and attributes each individual has, to the skills and attributes required for the job, not those of the person that came directly before them.
  • 9.
    Attribution Bias When wedo something badly we tend to believe that our failing is down to external factors like other people that adversely affected us and prevented us from doing our best. When it comes to other people, we tend to think the opposite.
  • 10.
    Confirmation Bias When wemake a judgement about another person, we subconsciously look for evidence to back up our own opinions of that person. We do this because we want to believe we’re right and that we’ve made the right assessment of a person.
  • 11.
    Conformity Bias Occurs whena positive or negative evaluation is made of someone based on their body language, personal appearance or style of dress.
  • 12.
  • 14.
    Example #1 Women areless like to apply for jobs that have a very long list of ’desirable’ qualities, as they do not wish to waste the employer's time if they are not perfectly suited to the role. Women are also less likely to shout about their achievements, and to negotiate salaries.
  • 15.
    Example #2 Researchers fromHarvard and Princeton found that blind auditions increased the likelihood that female musicians would be hired by an orchestra by 25 to 46 percent. https://www.forbes.com/sites/pragyaagarwaleurope/2019/02/20/how-to-minimize-unconscious-bias-during-recruitment/
  • 16.
    Example #3 The researchersfound that 70% (possibly even higher) of hidden biases are directed towards African-Americans, the elderly, the disabled, and overweight individuals.
  • 17.
    What evidence doyou have for the claim that the wording of job ads has any effect on the people reading them? The evidence underlying this tool comes from a research paper written by written by Danielle Gaucher, Justin Friesen, and Aaron C. Kay, called Evidence That Gendered Wording in Job Advertisements Exists and Sustains Gender Inequality. It was published in the Journal of Personality and Social Psychology, July 2011, Vol 101(1), p109-28. https://www.hw.ac.uk/documents/gendered-wording-in-job-ads.pdf
  • 18.
    Job interview 1. Thelanguage in job adverts affects the type of people who apply. 2. The fact is Latisha and Jamal do not get the same number of callbacks as Emily and Greg. 3. Structured interviews “standardize the interview process” and “minimize bias” by allowing employers to “focus on the factors that have a direct impact on performance.”
  • 19.
    Job interview 1. Employersoften use the last person to hold a post as a benchmark for the type of candidate they’re looking for. 2. The growing use of machine learning algorithms to help identify potential employees is leading to new types of bias. 3. Some people say they have no bias against anything, and that’s laughable. http://www.bbc.com/capital/story/20180806-how-hidden-bias-can-stop-you-getting-a- job
  • 20.
    Gender-neutral language Masculine FeminineGender neutral man woman person father mother parent boy girl child uncle aunt husband wife spouse actor actress prince princess waiter waitress server rooster hen chicken stallion mare horse
  • 21.
    Everyone loves ____mother. http://www.english.illinois.edu/-people-/faculty/debaron/essays/epicene.htm
  • 22.
    Example: Each participantmust present his ID badge at the door. Revised: All participants must present their ID badges at the door.
  • 23.
    Example: The tenantmust keep his apartment clean and tidy. Revised: You must keep your apartment clean and tidy.
  • 24.
    Example: The clientshould receive his invoice in two weeks. Revised: The client should receive his or her invoice in two weeks.
  • 25.
    Test your jobdescription: http://gender-decoder.katmatfield.com/
  • 26.
    Where and whenbias appears?
  • 27.
    Is AI aRacist?
  • 28.
    How to avoidbias? 1. Rules vs. data (AlphaGo) 2. Give choice (Google Translate) 3. Be neutral 4. Learn and understand your biases
  • 29.
    Digital Transformation Conference June22, Kyiv, Microsoft Ukraine https://dtconf.com
  • 30.