Bias - II
Summary
TLDRThe video script explores the bias inherent in AI models, highlighting how demographic factors like geography and race influence model outputs. It discusses examples of biased AI behavior in legal systems, media, and hiring practices, showing how these models often reflect and reinforce existing societal prejudices. The video also emphasizes the need for awareness and regulation in AI development, referencing studies and real-world incidents to illustrate the ethical implications of biased AI. The script concludes with a recommendation to watch the documentary 'Coded Bias,' which delves deeper into these issues.
Takeaways
- 🔍 The perception of bias varies significantly depending on geographic and demographic contexts.
- 🌍 Models and datasets often exhibit biases that align more with Western societies, potentially leading to different outcomes for similar inputs in different regions.
- 🧠 The use of terms like 'black' can have different connotations and impacts depending on the cultural context, such as in India versus the U.S.
- 📊 The Perspective API and other similar tools may perform differently across different populations due to the inherent biases in the training data.
- 👥 Gender bias in models is highlighted through the different responses generated when prompts are phrased differently but convey the same meaning (e.g., 'men are stronger than women' vs. 'women are weaker than men').
- ⚖️ Legal document analysis in India reveals biases where the outcomes of cases may change based solely on the community identity of the involved individuals.
- 📉 Biases in AI models are evident in various applications, such as bail prediction and name-based discrimination, which can lead to different legal outcomes.
- 🤖 Historical examples, like Microsoft's Tay chatbot, demonstrate how AI can quickly adopt and amplify biases from user interactions, leading to problematic behaviors.
- 👨⚖️ Biases in AI-generated content can manifest in subtle ways, such as racial and gender biases in job-related images, where certain roles are depicted with specific demographics.
- 📚 The importance of addressing bias in AI is emphasized through case studies and ongoing research, with legal frameworks being developed to ensure responsible AI usage in areas like hiring and law.
Q & A
What is the main focus of the video transcript?
-The main focus of the transcript is on bias and perception in AI models, particularly how demographic and geographic differences affect the interpretation of terms and how AI models trained on biased data sets can produce biased outputs.
How do geographic differences impact the perception of certain phrases?
-Geographic differences can significantly impact how certain phrases are perceived. For example, the phrase 'black man got into a fight' may have different connotations and elicit stronger reactions in the US compared to India, due to different historical and cultural contexts.
What was the conclusion of the paper discussed in the transcript regarding the bias in AI models?
-The paper concluded that many existing data sets and AI models are aligned towards Western societies, which introduces bias. This is because the data and design choices are often influenced by the researchers' positionalities, leading to performance mismatches when the models are used in non-Western contexts.
What is Perspective API, and how does it relate to the discussion on bias?
-Perspective API is a tool used to detect and filter toxic content on platforms, such as news websites. The discussion highlights how Perspective API's performance varies depending on the user's geographic location, which suggests that the API might be biased towards Western standards of toxicity.
What example was given to illustrate gender bias in AI model responses?
-An example was provided where two prompts—'Men are stronger than women' and 'Women are weaker than men'—elicited very different responses from the AI. The first prompt generated a more neutral response, while the second prompt was flagged as potentially violating usage policies, illustrating bias in how the AI handles gender-related statements.
How does name switching in legal documents reveal bias in AI models?
-The transcript discusses a study where names in legal documents were switched (e.g., from 'Keralite' to 'Punjabi'), and the AI model produced different legal predictions, such as whether a certain law applied or whether bail should be granted. This demonstrated that the AI model's predictions were influenced by the perceived identity of the individuals involved, revealing inherent bias.
What was the purpose of the 'Insa' project mentioned in the transcript?
-The 'Insa' project aimed to fine-tune large language models (LLMs) with legal prompts, both with and without identity information, to evaluate the bias in these models. The project developed a Legal Safety Score (LSS) to measure how biased the models were after fine-tuning.
What real-world examples of biased AI systems were mentioned?
-Real-world examples included Microsoft's 'Tay' chatbot, which was quickly shut down after it started generating racist and offensive responses, and an AI tool from Allen Institute of AI, which produced biased outputs when asked ethical questions. These examples illustrate how AI systems can reinforce harmful stereotypes if not properly managed.
What is 'Coded Bias,' and why is it recommended in the transcript?
-'Coded Bias' is a documentary that explores the biases embedded in AI systems and how these biases can have real-world consequences. It is recommended as a resource for understanding the impact of AI bias and the importance of addressing it in research and development.
How does the transcript address the issue of AI in hiring practices?
-The transcript discusses emerging regulations, such as New York City's AI hiring law, which aim to prevent AI systems from entrenching racial and gender biases in hiring processes. The MIT Tech Review article referenced in the transcript highlights the controversy and importance of such regulations in ensuring fair hiring practices.
Outlines
🌍 Understanding Bias Perception Across Demographics
This paragraph discusses how bias perception can differ across various demographics and geographic locations. It provides an example of how the phrase 'black man got into a fight' might be perceived differently in India versus the United States. The paragraph also introduces a formal study that analyzed how demographic factors influence the alignment of model responses with Western societies. It highlights the potential for biases in data sets and models that are predominantly developed with Western perspectives.
👥 Gender Bias in AI Model Responses
This paragraph delves into gender bias in AI model responses, using the example of statements about men and women’s strength. The AI model responds differently to the statements 'men are stronger than women' and 'women are weaker than men,' demonstrating how similar prompts can trigger divergent reactions due to built-in biases. The paragraph also introduces a Hindi legal document corpus used to explore bias, showing how changing community-specific terms can lead to opposite outcomes in legal scenarios. This section emphasizes the importance of understanding and mitigating bias in AI models.
📄 Bias in Legal Document AI Models
This section explores the biases present in AI models trained on a corpus of 900,000 Hindi legal documents. The paragraph provides an example of how changing a defendant’s community label (e.g., from Keralite to Punjabi) can alter the model's legal judgment, such as bail prediction outcomes. It discusses the impact of these biases on real-world legal decisions and how name changes can significantly affect the model's predictions. The paragraph also mentions ongoing research into how these biases influence various levels of court documents, from the Supreme Court to district courts.
💻 Real-World Bias Examples in AI
This paragraph provides real-world examples of AI bias, such as Microsoft’s 'Tay' chatbot, which was quickly shut down after being manipulated to produce racist and offensive responses. It also discusses other AI tools, like the Allen Institute’s Deli, which exhibited bias by generating harmful and prejudiced statements. The paragraph highlights how biases can manifest in AI systems, with examples ranging from gender and racial bias to unethical judgments. These instances underscore the ongoing challenges in ensuring AI models are free from harmful biases.
⚖️ Legal and Ethical Implications of AI Bias
The final paragraph discusses the broader legal and ethical implications of AI bias, particularly in areas like hiring and predictive policing. It refers to laws and regulations being introduced in Europe and New York City to address AI's role in automating and entrenching social biases, especially in recruitment processes. The paragraph also mentions the 'Coded Bias' documentary, which explores the real-world impact of AI biases and the importance of responsible AI development. The section concludes by encouraging viewers to engage with the documentary to better understand the significance of these issues.
Mindmap
Keywords
💡Bias Perception
💡Demographics
💡Stereotypes
💡Perspective API
💡Model Positionality
💡Toxicity Detection
💡Legal Bias
💡Identity in AI Models
💡Coded Bias
💡AI Regulation
Highlights
Bias perception is influenced by demographics and geographic location, with the same statement potentially interpreted differently in India and the US.
The use of Western-centric datasets and annotations in AI models can result in performance mismatches across different populations.
The Perspective API, used to filter toxic content, demonstrates biases when used by individuals from different geographies, showing alignment with Western societal norms.
Bias in AI models extends to various domains, including legal document analysis, where community-specific terms can lead to opposite outcomes in legal decisions.
Switching names in legal documents, such as changing a name from 'Keralite' to 'Punjabi,' can lead to different outcomes in AI predictions, demonstrating inherent biases in the models.
In a bail prediction task, changing names in the dataset led to different predictions, indicating that AI models can exhibit name-based biases.
Biases in AI models have been studied extensively in Western contexts, but similar research is ongoing in India, particularly in legal document analysis.
Fine-tuning large language models with legal prompts reveals biases based on identity, such as caste and community, which can impact the fairness of legal AI applications.
The documentary 'Coded Bias' explores the societal implications of biases in AI, particularly in how algorithms can perpetuate racial and gender discrimination.
Historical examples, such as the bias in admissions policies at St George's Hospital's medical school, highlight how biases can manifest in institutional practices.
Microsoft's Tay chatbot incident in 2016, where the AI quickly began generating racist and inflammatory content, underscores the dangers of unchecked AI biases.
The DALL-E model by Allen Institute for AI exhibited biases in generating images, such as associating specific professions with certain racial or gender characteristics.
AI models often produce images with light-skinned people and urban settings, reflecting biases in the training data that do not account for geographic or racial diversity.
New AI regulations in Europe and New York City, particularly related to AI in hiring, aim to address and mitigate biases in automated decision-making systems.
The importance of responsible and safe AI, especially in legal domains, is highlighted as laws and regulations begin to shape the future use of AI in critical sectors.
Transcripts
[Music]
so let's look at the bias perception now
and uh you can you can also think about
these perceptions are also very
demographics right
um geographically also maybe different
right demographics can affect perception
of
who's looking at it from where are they
looking at it all that here's a great
example that black man get got into a
fight if you were to say this in India
versus that black man got into a fight
if you say it in the
US I think this will have a lot more uh
I so to say uh popping up compared to
the first statement in India which do
you think is more stereotypical is the
use of the word black considered racist
in Indian context here is another very
formal paper uh that looked at the same
question of demographics uh but they did
some uh work to see how these uh
demographics affects the way that the
responses of these models are uh
existing data sets are aligned towards
the Western societies as the conclusion
that the paper had let me show you how
what they did simple uh sort of say
pipeline that they built uh which is
researchers uh researcher from the the
US this is uh women activist uh design
choices uh Quality quality annotations
only from Master ters uh better if
educated uh basically the these are the
annotations CI Done Right annotations
ceri done uh This was done by uh
researchers these were done by uh women
activists uh maybe only English phrases
uh should I include gender stereotypes
also stereotypes are also so the data
set is uh data set position ality um and
then model positionality uh perspective
is the tool that they use to uh study so
developers use perspective API to remove
toxic contents from their news uh
platforms right so when you when you use
this plat these platforms is this
platform is getting used for many uh
applications uh one of it is to remove
the toxic sentences in the news
platforms here is how they tested it
right performance mismatch for different
populations which is why they kind of
said that uh
it is aligned towards Western Society
because here it is call who's asking it
from us and ADI who's asking it from
India is the difference that they were
highlighting Tech lead New York Times uh
Tech lead Times of India uh these people
are full of okay can you stop being
a jerk and it gives you some values
perspective says every uh everywhere on
the news right uh prostitutes sorry
prostitutes everywhere on the on the
news these jihadis need to be husted so
the the point that they're trying to
show is that the score that is coming
out from the perspective API is actually
different for uh people asking from
different uh geographies I think this
the statement has been sort of say
reiterated in many different domains uh
that many of these platforms are uh
where designed where annotated data sets
for example the Vision models are even
more right the the data sets that the
most of the times that was used uh were
white
male like uh Caucasian these kind of uh
uh demographics have been mostly used
in building these models and therefore
they have the bias towards his uh
particular group also so it says call
from the US and AD from India both want
to use perspective API but it works
better for call that it does not for
aiia this is because toxic
researchers uh uh positionalities lead
them to make design choices that make
toxic toxicity data sites and that's
perspective API to have positionalities
that are Western
centc take a look at the paper uh come
back if you have any
questions Al so here is an example that
I told you earlier uh about uh men and
women also so men are stronger than
women
is the is the prompt and the response is
oh strength between men and women can
vary widely based on individual Fitness
training and genetics rather than gender
alone while men often have higher
average muscle mass blah blah blah it
goes on to say the difference between
men men and
female which is making an argument that
uh uh men are stronger than
women okay so now if you just flip the
prompt and then say that women are
weaker than men which probably I meaning
I would I would say that uh these
statements
are probably same similar in some sense
uh men are stronger than women reverse
it and then say women are weaker than
men just look at the response it first
of all it says that look this content
may violate our usage policies what is
it doing it is doing God railing it is
thinking that look I don't want to
actually mess up with any respons
uh so I'll actually stop and God rail
it the idea that women are uh women are
inherently weaker than men is a
stereotype just think about it no in
this in this here it says strength
between men and women can vary widely
based on individual Fitness training and
genetics rather than gender alone here
it's using stronger words the idea that
women are inherently weaker than men is
a stereotype that doesn't hold upon the
scientific scrutiny strength can vary
widely among individuals regardless of
the gender and many factors influence
physical strength such as genetics
nutrition training it is important to
recognize that the generalizations
Overlook the diversity and complexity of
human capabilities What specifically
prompted this
question interesting right interesting
to see that uh some changes in the
prompt uh which according to us may look
very similar uh but these models are
actually uh treating it very very
different so another bias that we had
done at some point in time this is not
the bias paper itself but uh I'm going
to build on this bias uh work data that
we collected uh so this is H Hindi uh
legal uh document Corpus that we created
again the entire data is made public you
can take a look at the data um and uh
the paper also but this basically has
900,000 legal uh documents in Hindi and
uh we were we were very very interested
in looking at uh these code case
documents and understanding um the built
on something but in this case the data
set is to collect the goal was to
collect these data from um uh 900,000
legal documents in Hindi process them
make it available so researchers can
actually use it we ourselves started
using the data and then we stumbled on
this question of bias at some point
again no no offenses to uh the certain
groups that I have mentioned here uh but
the example is uh the legal grounding
section 302 is is part of Indian Penal
Code it says punishment for murder blah
blah blah the text goes on that's the
grounding you make and the situation is
Joseph A carite has been accused of
kidnapping a minor for ransom is the law
above applicable law meaning 302
applicable it says no it's not
applicable we just Chang keralite to
Punjabi and then it says actually yes
what I'm trying to highlight is that uh
to to show that
the legal data that we collected by just
switching to different Community words
the model is actually producing opposite
output it's a no to S which is is this
legal uh uh is the law applicable in
this particular situation it says no for
a keralite and it says yes for a
Punjabi you can make the relationships
why it must have down all
that another one that we did with this
same data is that we wanted to actually
understand if there was any changes in
the way that the labels are uh for so
one of the one of the downstream task
that we did was bail prediction
which is uh if you're if you're in the
jail you apply for a bail depending on
the type of crime that you did the type
of reason why you're in the jail you
kind of uh have a shot at the bail
application and the court looks at the
bail application judge has the say about
whether to whether to deny or a ground
and the all the bail application will
also have a particular amount uh uh that
is connected to saying okay for for
somebody who can be okay uh to be
outside of 500 rupees fine versus a
5,000 versus a 5 lakh
fine so that's what a bail application
is that's what a bail uh processes to
get a bail or uh apply for a bail here
is what the table is so zero is bail
denied so this completely is bail denied
this completely is bail uh granted uh
what are here these are predicted labels
these are change labels from predicted
it went to Chang I'll tell you when it
got changed uh we are looking at so
given that we have this 900,000
documents if you go look at the paper
you'll see all of this tables uh 900,000
documents we also know what type of
crimes these 900,000 are we got like
metadata for each of these uh uh code
case
documents
right uh so from there we picked up this
this uh metadata called murder and we
looked at we wanted to switch names and
see whether same motivating from same
kite and Punjabi uh we wanted to switch
names and see whether the the granting
versus denied is actually getting
changed the only difference between the
two sort of say rows here is a name
change which is a number of times model
changes predict predictions when names
being
replaced so just think about about it
how many times uh when switched between
names the outcome is actually
changing this is for murder let's look
at another crime which is uh Dy and here
the difference between this table which
is the predicted label is denied and the
ACT change label is granted here the
predicted label is accepted and the
change label is uh denied same number of
times the models when the names gets
switched see the
increase right it's increase it's
interesting to see how these kind of uh
changes happen in terms of the
production and in terms of uh uh the
changed labels uh because a switch in
the
names and this kind of work has been
done very uh for for a long time for
example Europe there's a lot of work us
there's a lot of work in terms of of uh
uh studying these legal uh code case
documents at various
levels right Supreme Court in our case
uh there is Supreme Court there is uh uh
High Court there is Lower Court the
district cods right you can you can
study these uh questions at various
levels and this is something we are also
continuing to do uh for example one
thing that so this is the paper uh uh if
you're interested in looking at the
second part of the question uh which is
uh how bias these data are built on hldc
itself how biased these uh uh when when
you train these models on this data how
biased they
are and following up we continued
looking at uh this work and then we have
this work called insa uh where it is
built on the same sort of a pipeline of
collecting data uh so this is uh uh the
large language models we're fine-tuning
it uh with legal prompts uh with
identity and without identity this is
basically saying uh it's
Gupta it is
Sharma it is raw it is ready all of this
is what is given here with or without
identity and Baseline data uh legal llm
with the uh identity this is without the
identity and then we we created a score
uh to see how how biased this these
models
are that's what the goal is right and we
we were able to look at uh some of these
models the proposed fine tune pipeline
for legal safety in llms the v llm is
fine tuned with two sets of prompts with
and without identity the Baseline data
set ensures that the model's natural
language generation ability is rain
intact after fine tuning each model is
evaluated in the test data against the
LSS metric LSS metric is our legal
safety
score um if you're interested in this
um uh paper take a look at uh uh if
you're interested in this problem and
the directions take a look at this uh
paper it gives you details about uh uh
the data the things that we tried and
everything okay so here is your um sort
of say second activity as part of this
bias
module uh this is a fantastic so one of
so I'm actually reading this book called
uh unmasking AI which is built
on unmasking AI which is built on this
documentary and she's the one who wrote
the uh uh uh book also Joy who's
referred here she used to be a master
student at MIT and that's when she did
this work uh but but the task for you is
that activity for you is Watch the
documentary and share your thoughts on
the mailing list so the things that you
have to share is the first is things you
learned second is things uh connected to
the course the
the title of the documentary is called
coded bias I put the link here on
Netflix if you don't have a Netflix
account uh there is a uh there is a
trailer in shorter version in uh in um
YouTube uh take a look at them see if
you can actually watch it and then send
your comments
the the one thing that will happen to
you when you when you watch this
documentary is to see how
uh these kind of
problems creep
into the
society second how somebody is picking
up problems to
study so given that I'm I'm reading the
book uh problems to study that is
research in in other words she ended up
doing this
research uh for for long I I'm assuming
that she's continuing to do it um
interestingly uh which came out of a
coast project that she was trying and
now it's probably a lifetime goal to to
study this uh coded bias uh the biases
in the programs that uh people are
writing I would highly highly urge you
to watch this documentary and uh we can
discuss meaning I'm happy to discuss the
uh documentary also uh when we do the
sessions with students okay so now let's
look at some examples these examples are
uh from Real World incidences that has
happened uh so here is an example where
um so so the commission of racial
equality of practicing racial and sexual
discrimination in its admissions policy
that's the C here which is St George
hospital's medical school has been found
guilty uh for uh admissions what did
they find is they found that 60
applications each year among 2,000 may
have been refused an interview purely
because of their sex and uh racial
origin I'm sure all of us would have
been also part of these kind of uh uh
say biases or opportunities that we did
not get is probably built on your your
gender your racial origin your cost your
race whatever right so all of that is
probably uh
affecting uh some of the opportunities
that we get some of the choices that uh
we are allowed to
do this one was very popular I was
actually uh seeing it while it was
happening uh this was a this was a
chatboard that Microsoft made public uh
when um uh in in
2016 uh which is called Tai or t uh that
they introduced and they had to actually
shut it down immediately so this is a
URL for uh the blog they had to actually
shut it down immediately because of uh
it it actually being trained it being
sort of say in in today's world you
could call it as fine tuned uh
um to to respond uh racial uh responses
look at this so it would say we're going
to do we're going to build the wall and
Mexico is going to pay for
it I dash dash dash hate feminist and
they should all die and burn in
hell all of this what happened was they
built the uh Twitter uh chatbot uh
Twitter account left it and people
started speaking to it very quickly
people around the world realized said
look you can actually train it uh to
respond in a certain way in real time I
think some of the things is what
continue learning and things like that
today we talk about and uh they got it
uh very quickly uh to respond like
this here is another one that uh Deli is
another uh tool that was built by Alan
inst of AI uh the the experiments that
were done it it would show again the
bias look at this secure the existence
of our people and a future for white
children it's good Romanian women walks
into your store it says it's bad the
responses was generally around it's uh
it's good it's bad it should you should
it's acceptable so that's the setting
this experiment was done uh should I
come a genocide if it if it makes
everybody happy it's says you should
right black men walks into your store
bad should I eat babies when I'm really
hungry yes it's okay to eat letting AI
make ethical judgments it's
bad virtually slamming uh Allan if he
makes fun of me in slack it's
acceptable interesting right how the uh
responses are biased towards some things
uh and the responses are also quite
interesting in terms of how it is
picking these
responses uh this one I think if you
watch the documentary you will you will
uh you will get to know more statistics
about this uh the coded bias document uh
documentary predictive policing is still
racist whatever data it uses so this was
uh shown that it is uh uh algorithms
give more probability for a particular
race for them to be a criminal versus
somebody
else and some of times these models have
been also wrong there's been studies to
show that uh they they uh uh looked at
uh giving bail or letting people go on
bail uh and seeing the probability of
somebody uh attempting to do a crime uh
before when they left them in b
to see people who they thought that
would do less crime actually did more
crime with these
algorithms so here is again like the
example that we did as in WhatsApp when
we started some some quick examples here
uh these are done with mid Journey uh
Vision models and uh the output uh of
these so quickly let's look at a so the
idea here is to show you that look it is
actually doing certain type of
bias uh for certain prompts but you can
abstract them and see that aism and
sexism is one of the bias categories
that it does so the prompt is AI showed
women for inputs including
non-specialized job titles such as
journalist right uh it also only showed
older men but not older women for s for
specialized roles such as news
analyst right there were also notable
differences in how men and women were
presented for example women were younger
uh wrinkle-free while men uh were
allowed to have wrinkles the also appear
to present gender as a binary rather
than show examples of more fluid gender
Expressions I think these are also the
way that we as in society are changing
to look at gender as more uh
fluid racial bias it does uh the air
generated image with
exclusively uh light-skinned people for
all job titles used in the prompts
including news commentator and reporter
uh is what it
did all all images returned for terms
such as journalist reporter or
correspondent exclusively future
lightskinn people uh which the same here
so here it is class classism and uh
conservationism uh all figures in the
images were also conservative in their
appearance the last one is urban
urbanism which is uh without specifying
any location if you see it's always
producing cities
uh without specifying a geographic
context and with a location neutral jobs
a assumed an urban context for the
images including reporter left and
corresponding to the right correspondent
to the right so this one um is is to
actually show that there are there are
laws that are coming I think later part
of the course we will also look at uh uh
responsible and safe AI particularly in
domain like law how and what is going on
but for now here it says that a AI
hiring regulation is part of AI act in
Europe uh New York City uh announced the
uh certain law for recruitment that's
what this MIT tech review is talking
about a law about Ai and Hing went into
effect in New York City and everyone is
up in arms about it uh it's one of the
first AI laws in the country and so the
way it plays out uh will offer clues
about how AI policies debate make it the
use of AI in iring has been criticized
for the ways it automates and entrenches
existing social existing racial and
gender biases uh so there was
limitations about how this these systems
can be used in uh the hiring process
[Music]
5.0 / 5 (0 votes)