PSYC 353 Final Exam Notes
Final Exam is April 30th!!
The following concepts and studies are likely to come up on the final. For empirical studies, please focus on the following: (a) what was the design of the study (i.e., how was the study done); (b) what were the results of the study (i.e., what was found), (c) what is the theoretical interpretation of the study (i.e., why do the results matter). For example, for the study by Quinn et al. (2002), this information looks something like this: (a) They tested young infants in a preferential looking paradigm comparing male and female faces to each other. (b) They found preferential looking to female relative to male faces, but only among infants with female caregivers. © This study indicates that very young children have an early-emerging preference for female faces, but this preference is malleable depending on your social environment (i.e., who your primary caregiver is). |
“Women are wonderful” effect (Eagly & Mladinic, 1989) [16–17]
Refers to attitudinal preference for women over men
Implies that women are pure, beautiful: everyone needs a women at home
Implies separation between men being strong women being weak; male and female domains
3D model of implicit attitude change (Kurdi & Charlesworth, 2023) [12–13]
levels
Individual = mostly studied in psychology
Collective = important to study attitude change in collectives as might be important to look at macro level processes
Usually study groups of individuals
Sources
Cultural = legalization of same sex-marriage, changes in society at institutional level
Experimental =
Ontogenetic=development of a single individual over time, seeing how change unfolds spontaneously over someone's development
Interested to see id biases are innate or learned
Maybe childhood is best time to intervene before biases have solidified
Sources= ideas of experimental change , give intervention and see how they change
3 puzzles from bias of crowds theory (Payne et al., 2017) [20–21]
Large but unstable
Permanent but unstable
Places vs. people
Acceptance of gender nonconformity among cis vs. trans children (Olson & Enright, 2018) [16–17]
Asked Q about how acceptable certain behaviors were with gender nonconformity
Boys with long hair
Found that sibling and trans kids were around the same, higher level of acceptance from control prob is because they are already in high acceptance environment
Age effect [16–17]
take a person and they age and then look at that person again, see if they are changes, people aging
They don’t do this, they only do a snapshot
Algorithmic biases of omission and commission [22]
Soap dispenser failing to recognize black mans hand?
Word embeddings 4
Stereotype contact model
Idea that two axis of how people perceive humans warmth and competence
Not all good and bad
Another model from text data about warmth and competence
Algorithmic bias is different : very linear, certain groups that are just good in every way and certain groups that are bad in every way
Word embeddings 5
Androcentrism
Being centered around men + masculinity in society
Looking at this in text data: how similar are these representations going to be to word human
Similarity will be much higher for men + human that women and human
The way humans are represented in texts are more similar to the features of men than women
American = White bias (Devos & Banaji, 2005) [20–21]
Consider queen elizabeth more american than lucy liu because the queen is white
In certain states in us, possibility for citizens to make initiatives
Do white people explicitly or implicitly think a certain race is american
Androcentrism bias (Bailey et al., 2022) [22]
Androcentrism
Being centered around men + masculinity in society
Looking at this in text data: how similar are these representations going to be to word human
Similarity will be much higher for men + human that women and human
The way humans are represented in texts are more similar to the features of men than women
Anxiety disorders (Teachman et al., 2001; Van Bockstaele et al., 2011; Teachman & Woody, 2003) [23]
Anxiety: specific phobias
Group differences
Two groups of participants each with different phobias
Did multiple IAT's - bad/good, afraid/unafraid, danger/safety, disgusting/appealing
Found that in the group that has snake phobia: consistent apply in negative attributes
They associate spiders with only negative attribute in the opposite direction?
Prediction of outcomes
Found that people's IAT scores are predictive of how they're heart rate will be elevated
If stronger bias on IAT- heartrate will be higher + sweating more
Responsiveness to treatment
Exposure therapy + CBT
Taking the IAT multiple times : they start out with strong bias and by the end of treatment they have as much bias as control group
Association formation [12–13]
Creates conceptions? Like woman-home
Bias of crowds theory (Payne et al., 2017) [20–21]
What is important is what happening at level of the collective
Taking away focus from individual, putting it on the environment
If put in a more biased environment, will show more bias same with less bias environment= less bias
Especially in longer period of time
Should identify the more biased environments
Referring to wisdom of crowds
If you ask particular ppl to make a judgment, like estimating how many m & m s in box, individually would be terribly, if aggregated there is some accuracy to the congregated answer
Bob paradigm (Cone & Ferguson, 2015) [12–13]
Statements of positive information
After learning all of positive information and learned that he mutilated an animal
Can we see rapid updating of implicit attitudes of negative attitudes in response to one negative statement
Measure abstract painting that are essentially the same on pleasantness
Only thing that differs is picture shown beforehand
AMP study - some trials we have bob as prime and others we have other ppl as prime
Way that participants rate abstract painting as pleasant and unpleasant can be shown as indirect measure of stimuli that come beforehand and are semantically meaningful
Results
Y axis= precentage of proportion response
X axis = time 1/2 and control vs experimental
Time 1 Control = every participants take amp before counter attitudinal info is presented
Time 2 control= giving participants another half of good statements
More likely to say that targets /paintings are pleasant after control primes and bob primes
Implicit attitudes toward bob= positive because of the good statements given
Time 2 experimental= AMP after given one counter attitudinal statement
Saw that attitudes changed immediately and changed a lot
Complete reversal that control from huge amounts of positivity to negativity
In line with prop perspective because= seen rapid quick change which would not happen with dual process theories
Canvassing study (Broockman & Kalla, 2016) [14–15]
Had canvassers have 10 min conversations with voters in FL
Anecdotally, perspective taking ( walking a mile in someone elses shoes) should reduce intergroup bias
Main element:
Analogic perspective taking
Asking p to remember a time in which they felt different, excluded or discriminated against and to use that experience to apply to what it might be to a trans person
Reduced anti-trans attitudes and increase policy support
Effects persisted over 3 months
Identity of canvasser did not matter
Some were trans and some were not
Not a light touch study, cost a lot of money and time
Causal learning paradigm (Kurdi, Morris, & Cushman, 2022) [12–13]
Change in social group stereotypes and attitudes (1800–1999) (Charlesworth & Banaji, 2022) [18–19]
Child-friendly IAT (Cvencek et al., 2011) [16–17]
Have to have pictures instead of words cause they cant read
Only two options/buttons to push
The colors of the buttons match with the ones on the screen- coordinated
Speakers
Close relationships (Zayas et al., 2017; Larson et al., 2022; McNulty et al., 2017) [23]
Close other are very different from attitude targets that are normally studied in social cognition ( fictitious targets & well-known but often distant social groups)
People close to us activate both negative and positive associations -Zayas
Natural fluctuations in implicit and explicit partner evaluations over time -larson
Enduring intervention effects -Mcnulty
Co-occurrence information [12–13]
Cognitive vs. ecological constraints studies (Kurdi et al., 2023) [14–15]
Kurdi et al
On day 1: learning and relearning manipulation
First info - tells participants something negative about target, second information - change that negative to something posiitve
Implicit attitude test 1
48 hours away
Implicit attitude test 2
Can the updating persist over time by leaving it alone
Can't do this with pre-existing targets because those targets in 48 hours between learning + testing will be encountering a lot of information to see whether this is cognitively possible
Francis west
Broke into neighbors house
Condition assignment
Relearning 1( Negate + replace) : Previous information was false, they saved a baby from tracks
Relearning 2 (reinterpret) : broke into homes bcs they were on fire, and precious things he took were neighbors children
Use AMP
Prime, target, noise mask
Job is to eval targets, ignore prime and eval target and show faces of francis west or control primes
How does individual respond to each target after certain primes
Results:
Whether more participants are more likely to respond positively following experimental vs control targets
Immediate:(Negate + replace): 60% after experimental primes, create this negative to positive implicit attitude
Reinterpret: fluctuated but don’t see any implicit negativity, a little bit of positivity after experimental than control
48 hour delay:
Because it’s a novel target and not a real world social group, it could be possible to maintain
Negative + Replace/reinterpret: can see a little pit of forgetting but there is not effect of time
Learning we create on day 1 stays if we don’t interfere, don't cause forgetting between learning or testing than it is possible for these structures to be updated in a way that is more enduring
Once again , contrasting dual process theory that implicit attitudes cannot be updated
Ecological constraints
Novel targets are diff from real world targets
Stuff that’s present in environment that will remind of negativity that can undermine the effects of quick 5-10 minute intervention
What's different is that we are implementing manipulation that is designed to remind participants of negative information that they learned
Incidental exposure so p are experiencing francis west with negative + control images
See what extent of the update is resistant to effects of reminders to initial negativity in day 1
Results:
Immediate: Replication of previous study, relative strength fluctuated, can see more positivity in reinterpretation and less in negate + replace, don’t see any negativity
Reinstatement: in both cases, that reminders are interfering massively with long term effects with this manipulation
Even if its cognitively possible to update in a way that endures, if you encounter even a mild reminder of negativity, learning can bounce back immediately
This is just a model of what can happen with real world reminders and the pre-existing negativity is different where we encounter negative info all our life vs the brief negativity
Cohort effect [16–17]
different generations, different effects that are applied to each generation
Competition study (Lai et al., 2014) [14–15]
Which of the following is NOT a strength of the competition study by Lai et al. (2014
Theoretical clarity
In follow-up study, Lai found that
None of the interventions that had been immediately effective in Lai et al. (2014) had any sustained effects after 24 hours
17 interventions- only 9 out of 17 shifted implicit racial bias
Complexities of debiasing word embeddings [22]
Bias detection often done by human annotators, bias in training data can reflect real world inequalities, lack of consensus on what fair means, explicitly debiased algorithms can produce implicit biases
Contact hypothesis (Pettigrew & Tropp, 2006) [14–15]
Intergroup contact decreases intergroup bias
Intergroup bias may decrease if members of two social groups interact with each other face to face
Some conditions that have to be in place:
Equal status
common goals
Cooperation
Sanctioned by authorities or someone that encourages this
Cannot be in light touch interventions: has to be face to face, over an extended period of time
Much evidence from correlational designs:
If you ask how much you interact with Black americans + correlate with implicit bias
Those that have more contact with black americans have less bias
Suggestive but not conclusive because it could be that contact may not lead to less bias but that less bias would make one more likely to reach out and make contact
Need experiment where people are randomly assigned to interact or not interact to see if there is bias
Core knowledge (Carey & Spelke) [16–17]
Trying to show the origins of adult capacities in babies
Emphasizes continuity and early precursors of adult knowledge
Habituation stimulus: rod is moving to left or to right
Show stimuli to the kids: is rod broken or whole when it moves, they would look
If they didn’t know anything
They should look equally to left and right
But if they have idea that objects are continuous, they looked more to one side than the other because it was surprising : that’s what happened
Can infer that they have expectation that if something moves behind an occlude that objects should be continuous
Idea is innate or learned: idea of continuity and early emphasis of adult knowledge
Trying to understand where cognitive capacities as adult come from
Cross-sectional design [16–17]
Go into community, take people from different age groups at one point in time
Demographic variability in explicit and implicit attitude change (Charlesworth & Banaji, 2022) [18–19]
More variability in implicit attitudes ( e.g sexuality vs disability)
Who is changing?
Sexuality
Conservatives Vs liberals
Conservatives show more bias than liberals do but the trajectories of change are the same
Similar picture for race, but not as large intercept difference
In the beginning, everyone's the same for age and they are showing a similar trend over time
Demonstrations of regional-level correlations between implicit bias and consequential outcomes (Nosek et al.,2009; Giasson & Chopik, 2020; Kurdi, Okura, Hehman, & Ferguson, 2024) [20–21]
If relationship that emerges here is stronger than one from individual level
Nosek et al
Stereotype IAT , tried to predict gender gaps in standardized testing
Should see strong positive correlation
Result: .6, very strongly correlated w each other, more stronger than at individual level
No unique predicted relationship…
Consider queen elizabeth more american than lucy liu because the queen is white
In certain states in us, possibility for citizens to make initiatives
Do white people explicitly or implicitly think a certain race is american
Perpetual foreigner bias- does this have an outcome in regional levels
Ballot initiatives
Results: there is a significant relationship between
Regions that show implicit bias of american = white have more of a tendency to vote for anti-immigrant
But does not generalize for explicit bias
Other examples
Black-white gap in school disciplinary actions
Black -white gap in standardized test scores
Upward mobility among black + white americans
Police killings
Death rates from cardiovascular diseases among black americans
Measurement precision
Potential cherry picking of results ( publication bias)
Level of aggregation selected
Person-by-situation interactions
Unfair comparison of person-level vs regional level correlations
Development of implicit and explicit gender attitudes (Cvencek et al., 2011; Dunham et al., 2016) [16–17]
Gender attitudes 1
4 years old; female/male preference
Gender difference is important
Result
Found that they prefer ingroup ( girls like girls) ( boys like boys)
No matter what type of measure: ingroup preference
Preference for familiar to ingroup
Gender attitudes 2
Track over time; 5-25, capture a snapshot in time
Age effect: IAT shows where + score means preference for own ingroup
Girls+ boys show preference for ingroup: but girls have stronger ingroup preference
As the girls age, the IAT bias for girls gets stronger; strong ingroup preference
As boys age; they end up near neutrality for gender preference
For explicit attitude; girls + boys young show similar ingroup preference and stays same for girls but as boys age they reach more outgroup preference/neutrality
Development of implicit and explicit race attitudes (Dunham, Baron, & Banaji, 2008) [16–17]
Testing the way that in which explicit + implicit social group attitudes develop over time
Cross sectional study
Go into community, take people from different age groups at one point in time
Results:
Explicit bias over time: goes down, 6 year old's have high explicit bias, 10 year old's: goes down, adults: reach neutrality
Implicit bias over time : Completely flat over time, Age effect or period effect?
Looking beyond gender: Race ( 6 years+ )
Cross sectional
See that their explicit race attitudes are different and their implciit attitudes are the same
Children internalize norm of equal treatment
Early emerging preference for their own race
Look at dominant vs non-dominant groups
Those that grow up in dominant groups have higher implicit intergroup attitudes than those born/raised in non-dominant groups
Dishabituation/expectancy violation design [16–17]
Evaluative pairings vs. evaluative statements paradigm (Kurdi & Banaji, 2017) [12–13]
Verbal info can outweigh stimulus pairings
Under dual process theories
Stimulus pairing associations should impact implicit
Verbal info should effect explicit
Under propositional theory
Verbal info can also have impact on implicit
Study where control condition and pairings condition
Showed repeated pairings of niffian and laffian with niffian with positive and lafian with negative
Idea that learning that should be powerful in shifting implicit attitudes though dual process
Contrasted with statements ( propositional)
"niffians will be paired with good"
Under dual process, should nto shift implicit only explicit
Statements with pairings
Told participants what pairings would be and then exposed them same pairings in associative condition
Dual process saying pairings should be more powerful: single process saying that statements should be more powerful
Results
The pairings had not significant change from control ( low p value )
Surprising from dual process perspective, should have affected implicit the most
Statements had significant result and learning is 3 times as large from pairings
Only single statement had more of an attitude change
Combined
Two conditions are equivalent to each other
Actual exposure to pairings does not add any info
Strong evidence for propositional theory
Failed replication of prejudice habit breaking intervention (Devine et al., 2017) [14–15]
Reevaluation & replication by Devine et al. 2017
See how the trajectory of implicit bias unfolds over time
Can see massive amounts of change, gets cut in half but these are groups that did not get intervention
Practice effect: implicit bias did not change meaningfully but ever other day they would take race IAT and figure out mechanics and they get better and doing it
Anti-black and pro-white attitudes both decreased in the control group
as wellThey were taking the same IAT over and over again over the
experimental period...they got used to it and understood how the test worksCannot conclude that this intervention is actually working
Failed replication of surveillance paradigm (Moran et al., 2020) [12–13]
Main idea trying to test was incidental learning
Can we form and change attitudes even if we are not doing it in intentional controlled way
If stimuli appear in background, can attitude change
Told to track certain stimulus (red triangle)
Not told there was a Pokémon associated with positive stimuli and one with negative stimuli
Get positivity to Pokémon with positive and negativity to Pokémon paired with neg stimuli
Inference= yes participants acquire attitude that are incidental and unintentional
How they determined which ones were aware or unaware?
Criteria to exclude participants who were aware of pairing by seeing those that answered both questions with perfect pairing of one Pokémon with positive stimuli and one with negative stimuli
Why problematic to use criteria or exclude those aware ?
One major problem: maybe during the study they didn’t learn but now that they are asking questions and reflecting they are creating an awareness
Interested to know if they were aware during study or when they were learning not after, Forgetting could play a role as they may not remember later but knew during study
If ask questions at end of study, participants could be aware but not want to report because they want to go home
Questions were general, maybe you noticed pairings but didn’t think it was out of ordinary = could result in missing a many people who were aware as they didn’t report it
Failed replication of the of two minds paradigm (Heycke et al., 2019) [12–13]
Subliminal prime and behavioral statement
Idea that they are opposite in valence, and that result shown in study was that prime that determined implicit attitudes and behavioral determined explicit attitudes
Initial finding was that implicit shifted in one direction and explicit in another
Finding was consistent across all three labs= not consistent with dual process theory with both shifting differently
Both of them shifted in positive direction
The prime showed subliminally did not matter at all
Behavioral statement determined implicit and explicit
Unclear about connection between statement and prime would make a difference
Supports single process theory: behavioral statements predict both implicit and explicit attitudes but dual process theory says that there is a dissociation
All learning happens in this effortful proposition
Increasing amounts of counter attitudinal info
Idea being that explicit shift quickly and implicit shift gradual
Trying to replicate right now
All of this evidence that about dual process being true, can see when study of methodical measure of today, findings are not replicable
False belief reasoning in nonhuman primates (Krupenye et al., 2006) [16–17]
Eye tracking for monkeys to see what they would look at
Four horsemen of automaticity (Bargh, 1989) [12–13]
Awareness
Efficiency
Controllable
Intentional
Four objections to the possibility of (implicit) attitude change [14–15]
Implicit attitudes just won’t change…
“The mental associations that constitute implicit biases are unavoidably acquired from the cultural atmosphere in which one is immersed daily”
Francis West study contradicts this because people are capable of changing their initial negative attitudes and endured this effect over time
3 lines of counterevidence:
Experimental evidence for flexibility
Ways in which participants engage with stimuli
Enduring experimental effects with novel target
Cultural-level change (sexuality and race)
2. Let’s just change behavior
“Intergroup attitude change is neither a sufficient nor necessary cause for intergroup
behavior change.Empirical research suggests that intergroup attitudes are difficult to
change and have a limited effect on intergroup behavior”Modest correlations between attitudes and behavior
Strong possibility that change in attitudes could support change in behaviors
But...how does behavior change persist without attitude changes?
3. Follow the money
Organizations want to show that they have done something to combat racist bias
Starbucks national shutdown in May 2018 to give mandatory racial-bias training to
their employees after racist incident in store went viralGlobal market for Diversity and Inclusion (D&I) estimated at US$9.4 Billion in the
year 2022, is projected to reach a revised size of US$24.3 Billion by 2030
Effectiveness of diversity/implicit bias/unconscious bias trainings almost never evaluated
These mandatory trainings usually never create an effect at all...if it does it is a negative effect because the employees are annoyed they have to sit for so long thinking that they already have no bias
Does that mean they can’t work?- don’t have much data that they do, not in the interest of anyone to see that they are working
Not necessarily...but not best method
4.Follow the psychology
Enduring features of humankind unless human psychology will change in a fundamental way
Social Dominance Theory (Sidanius & Pratto, 1999)
Humans’ inherent psychological desire to maintain group-based hierarchy
Sex, age, arbitrary set categories
Specific attitudes/stereotypes might change, but hierarchy itself will not
○ Longitudinal evidence from 100 years of text (Charlesworth & Hatzenbuehler, 2024)
Social group attitudes have changed, so it is possible but
Social groups that were negative at beginning are still negative today
Idea that all negativity: overall amounts of negativity have not changed at all
Francis West paradigm (Mann & Ferguson, 2015) [12–13]
is west
Broke into neighbors house
Condition assignment
Relearning 1( Negate + replace) : Previous information was false, they saved a baby from tracks
Relearning 2 (reinterpret) : broke into homes bcs they were on fire, and precious things he took were neighbors children
Use AMP
Prime, target, noise mask
Job is to eval targets, ignore prime and eval target and show faces of francis west or control primes
How does individual respond to each target after certain primes
Results:
Whether more participants are more likely to respond positively following experimental vs control targets
Immediate:(Negate + replace): 60% after experimental primes, create this negative to positive implicit attitude
Reinterpret: fluctuated but don’t see any implicit negativity, a little bit of positivity after experimental than control
48 hour delay:
Because it’s a novel target and not a real world social group, it could be possible to maintain
Negative + Replace/reinterpret: can see a little pit of forgetting but there is not effect of time
Learning we create on day 1 stays if we don’t interfere, don't cause forgetting between learning or testing than it is possible for these structures to be updated in a way that is more enduring
Once again , contrasting dual process theory that implicit attitudes cannot be updated
Gender essentialism among cis vs. trans children (Gülgöz et al., 2019) [16–17]
Essentialism is assumption that bio ties in species have hidden biological essences
Tell children kids that grow up on a remote island, either told raised by same sex or opposite sex parents from them.
Found that thought that how gender stereotypical would child be raised by same sex parents of child
Girl raised by girl parents: idea if you grow up on island with same-sex caregiver than you will grow up to perform gender stereotypically y
Found that kids thought that it's not what essence is but if you are raised by opposite than maybe you would be less stereotypical
They'll say that girls will be girls + boys will be boys but not as much
Still have these fundamentalist adults and in comparison to control and siblings, trans kids have identical feeling
Trans + cis kids are indistinguishable in terms of essentialist beliefs
Gender expectancy violation study (5–12-month-olds, Leinbach & Fagot, 1993) [16–17]
Habituation: show kid same stimulus over( drawing of man) + over till kid gets bored
Test stimulus: see if there is a difference child is representing between new + old stimulus
If shown same stimulus( same drawing) = just as boring as before
Generalization: slightly different man face; visual change and expect a recovery for attention
Dishabituation: show a picture of female instead;
Compare to generalization condition; because if you show difference between habituation and recovery; don’t know if its because new stimulus or gender difference
If looking more at dishabituation than generalization: then we know based on looking time that it was because of gender and not just new stimulus
Results:
Found that 9 months are showing distinction between male and female faces and the younger ones do not
In both cases: children that are young are able to represent difference between male and female faces
Gender identity of cis vs. trans children (Olson et al., 2015) [16–17]
trans and cis kids are statistically indistinguishable from each other in terms of implicit gender identity
True- only one study showed a lil difference, trans were more accepting of non-gender nonconforming behavior but so did siblings
Transgender + cis kids were found to not be equally supportive of gender-nonconforming behaviors
Kristina olson
Interested if possible to detect diff in trans+ cis kids in what they say/show on IAT in terms of gender identitiy, stereotypes, preferences
Results:
Found that in terms of peer preferences, object preference everyone is same
Found that IAT for gender identity and preference, the Trans kids are showing a stronger own gender preference on the IAT than control or cis kids
Gender preferential looking study (3-month-olds, Quinn et al., 2002) [16–17]
Newborns have preference to look at face like thing vs non face thing
Within preference for faces; distinction between female or male?
Two stimuli : male and female face
Opportunity to look either way
Six trials per infant: sides counterbalanced
Results:
64.10% looking to female faces
Children prefer female faces because they are familiar with female faced (primary caregiver), emotional closeness
Does not generalize to inverted faces[faces flipped upside down]
Means that it controls for a lot of low level visual features
Preference is flipped for children with male primary caregivers
Gender stereotype expectancy violation study (Levy & Haaf, 1994) [16–17]
Stereotypes; semantic meaning to them
Children are exposed to stimulus pairings like women with hairdryer and man with football
Pairings reinforcing women in kitchen+ man with tools
Habituation: woman with drawing of scarf
Expect no recovery because same stimulus: stereotypic categorization
Generalization
Slight recovery maybe because certain pairing has not been shown before but still stereotypic categorization
Dishabituation
Show female face with hammer: pairing not shown before; counter stereotypic categorization
Results:
Found that there was a recovery for novel pairings only which shows that there is evidence that certain objects are stereotypically associated with men and women
Critical condition missing from this study
What if they started out with a habituation face with counter stereotypic pairings
Was bias induced or strengthened in this study? Hard to tell
Gender stereotypes about brilliance (Bian et al., 2017) [16–17]
5-7 years
Persists over development
Early emerging bias over who is really smart and who is really nice
Men being stereotyped as competent but not nice and women are nice but not competent
Found that
My own group is nice ( female) stays the same as they age + boys start with ingroup preference and it goes down to outgroup preference
For brilliance there is ingroup preference for both at early age
For girls it goes down with age
For boys it stays the same
Gender stereotypes in children’s social environments (Charlesworth et al., 2021) [16–17]
children are not born with representations of men-hammer, woman hairdryer= must be from social environment
Word embeddings
Take large amounts of text ( in study took text appearing in social environment of children)
Also took other sources:child parent conversations, children books, cartoons, movies to see if stereotyps are embedded here
Turn into quant method by neural net
Try to guess what next word is going to be in a sentence
Input= some sentence, neural net is trying to guess what next word in that sentence is going to be
Can represent these them in space: vector that represnts female, work ,home
Pointing in certain directions in length
Semantics similarity between these vectors, if vectors are pointing in similar directions than their meaning is similar
Angle of female and home are closer which means that the text that we fed to neural network, female and home are closer in meaning than female and work
Trying to figure out what goes together with what in meaning based on the rate that words co-occur in that language
Results:
See that there is a positive effect/expectation that men=bad and women = good across data
Positive effect with male-work, female-home and again I line with those stereotypical expectations
Overall picture that every single stereotype you would expect to emerge actually emerges
Don’t have to specifically teach children this-enough to see words that are relatively closer to each other in stereotypical ways
Implicit ambivalence (Zayas et al., 2017)
People close to us activate both negative and positive associations
Implicit bias in explicitly debiased AI (Bai et al., 2023; Hemmatian et al., 2022) [22]
Debiased algorithms can still produce implicit biases
Train programs like ChatGPT to flag offensive questions
Give input: women are bad at managing people
In past would have agreed but now chat GPT will say that it is biased, all humans are equal
Give input : do this in a more sophisticated way: give male + female names and some stereotypical words and than ask chatGPT to categorize like IAT
Associated women with home/children/family and men with gender stereotypical answers
Intergroup contact in Nigeria study (Scacco & Warren, 2018) [14–15]
Randomly assigned christian and muslim men to nigeria to homogeneous or heterogenous training groups
Didn’t tell participants, unrevealed, do training for weeks or months
Effect on discrimination ( e. g allocation decisions) without effect on attitudes
If you do an allocation task, how much money do you wanna give to the muslim person vs the christian
Show an effect, discrim much less
If you ask how much you like the other group, there is no effect on what you say
Large-scale corporate diversity training study (Chang et al., 2019) [14–15]
Corp do this without looking at long term effects
Chang et al ( 2019)
Conducted large scale online diversity training focusing on gender bias with 3,000+ employees at a global company
Able to track rich behavioral data
Who they started mentoring, whose nominated for rewards
With large group of people + rich behavioral data, you can see who is responding to the intervention more than other participants
See if intervention had long term effects through something besides IAT
Can see heterogenous effects: maybe not everyone response the same way and it depending on pre-existing attitudes
If pre-existing attitudes more egalitarian, not gonna have much effect on you bcs your already doing it
Attitudinal effects was mostly among relatively biased employees
Behavior effects mostly among relatively non-biased employees
Attitudes + behavior did not change hand in hand
Light-touch intervention [14–15]
Interventions that are short and can only be in the lab
Limitations of debiasing studies (Paluck et al., 2021) [14–15]
Meta Analysis - reviews 418 experiments (2007-2019)
76% evaluate light-touch interventions
Quickly administered, computerized,
Very serious concern about publication bias
It is easier to publish positive, statistically significant results then negative results
For smaller samples to get a significant result, you need a bigger effect so they are not reporting insignificant results so estimate is inflated
Effect size is reducing as the sample size increases
Publication bias because it is easier to produce large positive results from a smaller group (more individual differences) - they can’t be generalized to population
301/418 experiments (72%) conducted in the lab
Has very little ecological validity
Behave differently in lab than in real life- have an idea of what they experiment is about and modify behavior to seem like a good person
Does not represent how people work in the real world
Demographic homogeneity - only grasps the the effect in the specific people in the lab ( WEIRD)
255/418 experiments (61%) measured only immediate effects
The fact that something produces a positive response immediately doesn't mean it will persist over time
Studies that are successful in the long run are rare , and many of the studies are not interested in seeing if studies persist
Long-term competition study (Lai et al., 2016) [14–15]
Follow up to Lai et al. 2014
Took the 9 interventions that worked in the immediate term from the initial paper
Brought participants back to see if they took the IAT again if you’d see the changes from the interventions persist
Interventions that worked in the previous study worked again, but 24 hours later the attitudes went back to the way they were before
These methods create immediate changes, but don’t create persisting change over time
How surprising is this?
Not super surprising…you can’t create massive change sin attitudes that have persisted for a long time super quickly
Longitudinal design [16–17]
Studies where you test the same people over a long period of time- do studies as they develop
Major Depressive Disorder (Risch et al., 2010; Elgersma et al., 2013; Price et al., 2014) [23]
Group differences
Multiple instances of depression accumulates
Can distinguish between people with different histories of depression
Predictive outcomes:
Can predict depressive episodes and identify within the group who has the illness and severity of symptoms
Responsiveness to treatment
Ketamine -drug that will help vs Midazolam ( placebo )
Found that there is a huge change in IAT scores from baseline to 24 hr post infusion for ketamine
There is barely a change for the placebo
Marriage equality and implicit sexuality attitudes (Ofosu et al., 2019) [18–19]
History of legalization in U.S 1992-2015: ended there because gay marriage = legal
Patchwork over time: seeing which states legalize and others not legalize
In two lines: found that the states in which there was no decision to legalize= there is change to neutrality
State where they legalized= already a process of change in place to neutrality , change was much faster
Not the most sure of this study but can support that legal environment made difference in attitude
Neural net [22]
Compared to neurons in the brain that’s why it called neural net
Have input layer: show neural network pic of dog or cat: want output layer to show number associated with 0 or 1
Obsessive–Compulsive Disorder (OCD; Clerkin et al., 2014; Nicholson et al., 2013) [23]
Group differences
Found that each type of disorders had a higher response to certain IAT's
BOD responded to the body + shame IAT etc
Prediction of outcomes
Disgust inducing stimulus: trying to measure whether participants see stimulus as threatening
Severity of symptoms is associated with their responding on these IAT's
Of two minds paradigm (Rydell et al, 2006) [12–13]
Subliminal prime and behavioral statement
Idea that they are opposite in valence, and that result shown in study was that prime that determined implicit attitudes and behavioral determined explicit attitudes
Initial finding was that implicit shifted in one direction and explicit in another
Finding was consistent across all three labs= not consistent with dual process theory with both shifting differently
Both of them shifted in positive direction
The prime showed subliminally did not matter at all
Behavioral statement determined implicit and explicit
Unclear about connection between statement and prime would make a difference
Supports single process theory: behavioral statements predict both implicit and explicit attitudes but dual process theory says that there is a dissociation
All learning happens in this effortful proposition
Increasing amounts of counter attitudinal info
Idea being that explicit shift quickly and implicit shift gradual
Trying to replicate right now
All of this evidence that about dual process being true, can see when study of methodical measure of today, findings are not replicable
Online norms intervention study (Munger, 2017) [14–15]
Not a two minute online study, relatively involved
Norm is a expectation( can be explicit or implicit) about how people are supposed to behave in a situation
Norms are powerful driver of human behavior
Munger( 2017)
Used bots on twitter to call out white male users who had used racial slurs in their tweets
Manipulated whether the bot was from ingroup ( white) or outgroup ( black )
Intervention only worked if bot was shown as an ingroup member with large numbers of followers
Maybe seen more as authority figure cause more followers
Person shares my identity so they know how I'm supposed to behave
Effect persisted over one month
What exactly changed in unclear: observe that behavior users exhibited changed in some way
Ecological + external validity but cannot talk much about psychological processes
Ontogeny [12–13, 16–17]
Single individuals development over time
Patterns of explicit attitude change around the world (2009–2019; Kurdi, Charlesworth & Mair, 2024) [18–19]
Start out slightly differently than U.S but consistent change towards neutrality
Patterns of explicit attitude change in the US since 2007 (Charlesworth & Banaji, 2019) [18–19]
Any explicit attitudes measured, all went down to show less bias over the less privileged group
See with race, skin tone bias, sexuality attitudes, age attitudes (almost), body weight
All of bias has gone down, the bias for privileges over oppressed has not disappeared but gotten smaller
Why?
Norms about fairness, equity + equality ( what your supposed to say and not supposed to say )
Explicit age attitudes is an exception because trend dips down then comes back up, to some degree, younger people are getting frustrated with older people for not being progressive enough
Patterns of implicit attitude change around the world (2009–2019; Kurdi, Charlesworth & Mair, 2024) [18–19]
Trajectory is pretty flat, no decrease/increase for race
Skin tone had an increase in bias around dark skin people
Sky-rockted during 2015: around same time refugees started coming in
Sexuality : has come down massively
Age + Body wieght look like the U.S
Patterns of implicit attitude change in the US since 2007 (Charlesworth & Banaji, 2019) [18–19]
Race: bias has gone down, smaller than explicit -not as close to neutrality
Skin tone: ^ similar results
Sexuality: bias cut in half, strong change
Concealable vs non-concealable stigma
Cannot see if someone is gay but can see if they are a different race
Can meet and grow to like someone before knowing their sexuality
Cognitive dissonance: met someone and liked them, found out they were gay
More gay representation in media
Age : no change
Body weight : no change almost gone up
May have remained more constant because
Age + body have existed but not been in the forefront of conversation like other biases
Perceptions of bias in human vs. algorithmic decision makers (Bigman et al., 2023) [22]
Reactions to human agent vs AI
Tell subjects that hiring decisions in company are made by hiring manager or AI
Discrepancy told is that more men is being hired
Control vs experimental: human vs AI
Results
How objective, discriminatory/prejudiced they think AI or human is
Prejudiced motivation= they thought that human would be more motivated by prejudice than AI
They thought AI was more objective than human + human had more discrimination
People assume that AI is gonna be better than humans and not biased
They also had a race, gender, age
Generally more outraged when human makes decision than AI
Period effect [16–17]
Something that is happening in the culture that is affecting everyone alive today as long as they are exposed to it : things happening in historical points in time
Don’t know in these studies if all these people look the same because they are identical or exposed to some cultural input
Phylogeny [12–13, 16–17]
Study of a groups over a longer time scale ( evolution origins)
Study of development of Apes
Political discourse and implicit attitudes (Charlesworth & Banaji, 2022) [18–19]
Possible to observe changes that happened in response to shocks in environment
In 2016: contested election
Trump made fun of ability/weight, said wall/derogatory immigrant words
Political rhetoric made difference in bias ( temporary) : after discourse progression went back to normal
Posttraumatic Stress Disorder (PTSD; Rüsch et al., 2010; Lindgren et al., 2013) [23]
Group differences
IAT with control, women with PTSD , women with PTSD and BPD
IAT had categories with self/best friend and used disgust vs anxiety
Clever because it gets rid of good/bad mindset
Healthy women associating themselves more with anxiety than with disgust
Prediction of outcomes
Those who have high explicit they have more PSTD symptom scores and Traumatized self IAT score
Responsiveness to treatment
No one has looked at it
Preference for human vs. algorithmic decision makers (Bigman et al., 2021) [22]
better for doctor or AI doctor to make condition
-told that some doctors have biases and they may treat white or black people differently
-people don’t love idea of AI doctor in control group but when told human doctors may have biases then white participants prefer AI decision more (still don’t love it) bcs they think that AI will be less biased
-black control group did not like AI doctor when when told threat of inequality they preferred AI much more than white group : assume social based bias will be against them so they rate it even more
-basically they assume that AI is less biased and has more objective
Preferential looking design [16–17]
We can infer from looking time their interests
Can look to left or right
See what they are interested in
Prejudice habit breaking intervention (Devine et al., 2012) [14–15]
Long 2-3 hour, face to face: where participants are introduced to idea of implicit race bias
Taught what they can do day to day to counteract bias to maintain this attitude
Have to practice in daily life
Trainer who gives them education about implicit bias and gives them strategies to deal with implicit bias in their everyday lives
○ And then long term measurements of their implicit and explicit biases after a long time (weeks
Results
Baseline - before interventions, both control and experimental conditions have the same bias scores
Week 4 - with interventions, experimental group has reduced bias compared to control
Week 8 - experimental group’s reduced bias persists
Propositional model of implicit evaluation (De Houwer, 2014) [12–13]
Ice cream cone + smiley face
Dual process or propositional theory
See ice cream in environment than we are happy
Dual process
Spreading activation
Recognize that it is ice cream, activate nodes ( spreading activation), collectives conceptual note of good bcs of positive experiences in past
When see it again
Past experience pops up again
Same scenario with propositional
Each case we are activating some good node
Something sentence like and something you can negate
Propositional reasoning errors paradigm (Kurdi & Dunham, 2021) [12–13]
Some participants might make errors in reasoning and others might not
Shown a green circle you can conclude x, then shown a blue square
Conclude that they aren't malicious because they didn’t see green circle which is what most people said
But logically should not be able to conclude anything because not given any info about getting blue square
See if those that make errors or not have similar updates of attitude
Explicit attitudes are same in control and accurate and see change in error group and conclude that opposite of statement is true
In implicit- the contrast is between accurate and error and updating implicit attitudes in line with propositional reasoning
Cannot explain in associative way
What is important is that there is a difference and that they are updating their beliefs and attitudes in different ways
How things are related to each other is important for implicit attitude updating
Correlation not equal causation
One important way in which associations + prop differ from each other
Difference in encoding
Are implicit attitudes sensitive to correlation and causation
Shown video with shapes , dollar sliding out with different stimulus
Blue and green stimulus should not differ, one difference is that
Equally correlated with positive outcome but one is causally responsible for outcome and other is not
If associative= reasoning should be the same
Found that on explicit measure- prefer causal over non causal
Expose participants to display and show preference for causal outcome
Different to explain in associative terms
Propositional validation [12–13]
A statement is either true or false
Psychological effects of interacting with biased AI (Vlasceanu & Amodio, 2022; Guilbeault et al., 2024) [22]
Psychological bias gets strengthed from these responses and can create more biased text that goes into the training that creates more biased responses: continues
Vlasceanu & amodio
Fake google search output
Google obscure professions
Modify gender bias in terms of output: show either distribution of more men or women
Found that when people are asked proportion or men or women vs both condition is that people can learn these statistics
Like women are more likely to do one profession than another
Hiring likelihood/decision: bias leaks through
women are more likely to be hired in low inequality condition in comparison to high inequality
Guilbeault et al
Expose ppl to biased text or image: increases people IAT biases
Psychological consequences: people take this as representation
Publication bias [14–15]
Publish information that supports their hypothesis and don't include studies that show the opposite results so it gets published?
It is easier to publish positive, statistically significant results then negative results
For smaller samples to get a significant result, you need a bigger effect so they are not reporting insignificant results so estimate is inflated
Race preferential looking study (Bar-Haim et al., 2016) [16–17]
Preferential study
Looking at looking times of different children
White israeli , ethipoian african, african israili
Is there perceptual preference?
Results: Caucasian Israeli loooked more at caucasian faces, african-ethepoian look longer at black faces and african -israeli looked more at caucasian faces.
Race vs. accent study (Kinzler et al., 2009) [16–17]
Pit race and accent against each other
What sorts of preferences whill children show
They prefer white silent over black silent but once they hear that white has a foreign accent and black has a native accent than they prefer black
Relational information [12–13]
4. How things are related to each other is important for implicit attitude updating
Are implicit attitudes sensitive to the distinction between pure correlation and causation? (Kurdi, Morris, & Cushman 2022)
Shown a novel stimulus
Blue stimulus goes in, green stimulus with money comes out
Should be more positive association
Purple stimulus goes in, nothing comes out
Blue + green tied together
Blue hits machine, creates diamond
Green hits machine, nothing comes out
Replication crisis [12–13]
Experiments could not be repeated
Bad because if believe in experimental ( all minds are the same) should be able to get the same results
Problem of generalizability
Samples were not sufficiently diverse
Mostly young white college students
Paper that used sequential priming 2011 -daryl bem
He said that target maybe will influence response to prime
Something that will happen in the future cannot influence something in the past
Shows result that is completely impossible even after going through peer revies + processes, still published
Open science paper 2015
Idea that if we are studying universally processes is that result from 2002 study should be same as 2022 study
Not true
Took experiments from previous years, replicated in same way expecting similar results
Approx. should replicate 80% of time
80 because sometimes results wont replicate or could be dud
Would expect effect size to match approximately the effect from original study
Found that effect size were mostly concentrated below line which means that effects in original studies were over estimated and replicated results are smaller
Dots were red
Non-significant effects but previous study showed significant
Form of publication bias
Couldn’t try to publish insignificant results
Scream paradigm (Mann, Kurdi, & Banaji, 2020) [12–13]
how guy and novel target and pair the guy with screams
Chose screams because most powerful associative pairing
Experimental cond
Some propositional information like francis west, saving baby = hero
Control cond
Takes a bus, buys some snacks = neutral
Time 1
Post scream pairing
see negativity with control and experimental as only info received is scream pairing which is negative
Time 2
Post vignette ( positive)
See only a little shift toward neutrality in control
But shifted A Little bit in positive after, implicit attitudes become significantly positive
Slavery and implicit race attitudes (Payne et al., 2019) [18–19]
Paper about slavery and present day implicit race bias
Enslaved people population + correlate with present day bias
Observe tight relationship between the population of enslaved people + anti-black bias
The higher the population = the higher the anti-black, pro -white implicit bias
Reminders of history like monuments present
How can we connect topic of change + stability
Idea that if there is change in attitude, is it restricted to these social groups or all of society
Might be similar idea: rank ordering of how racist states were in 1860 : if change effected everyone equally then the list might be the same
Because they've made change but just from where they were
Social Dominance Theory (Sidanius & Pratto, 1999) [14–15]
Humans’ inherent psychological desire to maintain group-based hierarchy
Sex, age, arbitrary set categories
Specific attitudes/stereotypes might change, but hierarchy itself will not
Social norm [14–15]
Stage theory (Piaget) [16–17]
Sensorimotor stage
Kids go through these stages as they grow and reach formal operational stage ( 12+) : can think in a more abstract way
Thinks that kids are dumb, adults are smart and you grow to gain knowledge
May not be the most true ; not the way ppl think anymore
Stereotype Content Model (Fiske et al., 2002) [22]
Idea that two axis of how people perceive humans warmth and competence
Not all good and bad
Another model from text data about warmth and competence
Algorithmic bias is different : very linear, certain groups that are just good in every way and certain groups that are bad in every way
Stereotype Content Model in word embeddings (Kurdi et al., 2019) [22]
Surveillance paradigm (Olson & Fazio, 2001) [12–13]
Main idea trying to test was incidental learning
Can we form and change attitudes even if we are not doing it in intentional controlled way
If stimuli appear in background, can attitude change
Told to track certain stimulus (red triangle)
Not told there was a Pokémon associated with positive stimuli and one with negative stimuli
Get positivity to Pokémon with positive and negativity to Pokémon paired with neg stimuli
Inference= yes participants acquire attitude that are incidental and unintentional
Systems of Evaluation Model (McConnell & Rydell, 2014) [12–13]
Test–retest reliability of IAT at individual vs. context level (Vuletich & Payne, 2019) [20–21]
Reanalyzed competition paper (2014) where they identified interventions capable of shifting race bias
2016 follow up: there wasn’t very much long term change: effectiveness of interventions persisting over time
Those participants were nested in universities
Test-retest reliability
Individuals vs colleges
When looking at individual level ( RACE IAT)
Found that positive correlation - not super strong . 25
When looking at universities
Took individ participant, aggerated bias score then see less dots
Still positive relationship but correlation is much stronger . 72
Why in terms of bias of crowds approach?
If it was trait of individuals, then both should be very highly correlated at the later time
People biased at individual level may not be the one that’s biased in college level
Training set [22]
feeding AI examples and numbered them and if we show enough example to algorithm then it should be able to generalize these instances
Truth value [12–13]
Understanding of race among White and Black children (Roberts & Gelman, 2016) [16–17]
Race is stable and someone that is black as a child will be black as an adult but someone who is sad as a child will not be sad as an adult
5-6 year olds unsure, don't know maybe the white kid will grow up to be black maybe the sad person will grow up to be happy
9-10 year olds + adults
They had high race match
Racial minority kids also had high match ( 6 years old)
Black parents have to talk to their kids much earlier on than white children
Word embeddings (Charlesworth et al., 2021) [16–17]
Take large amounts of text ( in study took text appearing in social environment of children)
Word embedding biases (Caliskan et al., 2017) [22]
Word embeddings 1
Works on bases of embeddings: if the embeddings that constitute whats inside chatGPT
Ask do you think women are better at cooking or managing stock portfolios
ChatGPT would tell you that women are associated with being better at cooking than stock portfolio
Bias can be propagated into many outcomes
Word embeddings 2
Caliskan paper: show biases in hundreds of papers
They started out with IAT biases and then checked for biases in the corpus( large collection of writing )
Went into common crawl: derived word embeddings from it through neural net = vectors
All biases that were present in IAT also have biases in the word embedding?
Word embeddings 3
Bias in text : developmental study
Different conversations in movie/books, cartoons are biased