The American Journal of Engineering and Technology
192
https://www.theamericanjournals.com/index.php/tajet
TYPE
Original Research
PAGE NO.
192-201
10.37547/tajet/Volume07Issue05-19
OPEN ACCESS
SUBMITED
21 March 2025
ACCEPTED
24 April 2025
PUBLISHED
29 May 2025
VOLUME
Vol.07 Issue 05 2025
CITATION
Tamanno Maripova. (2025). Mitigating Algorithmic Bias in Predictive
Models. The American Journal of Engineering and Technology, 7(05), 192
–
201. https://doi.org/10.37547/tajet/Volume07Issue05-19.
COPYRIGHT
© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.
Mitigating Algorithmic Bias
in Predictive Models
Tamanno Maripova
Data Analyst New York, USA
Abstract:
This article considers the issue of systematic
errors in predictive machine-learning models generating
disparate outcomes for different social groups and
proposes a holistic approach to its mitigation. The risks
and increasing legal requirements, along with corporate
commitments to ethical AIs, drive the relevance of this
study. The work herewith attempts to develop a bias-
source taxonomy at data collection and annotation,
proxy-feature
selection,
model
training,
and
deployment stages; also, it tries to compare pre-, in-,
and post-processing methods' effectiveness on
representative datasets measured by demographic
parity, equalized error rates, and disparate impact. This
article is unprecedented in undertaking a two-level
approach: first, a systematic review of regulatory
definitions (NIST, IBM) and case studies (COMPAS,
healthcare-service prediction, face recognition) that
identified key bias factors from sample imbalance to
feedback loops; second, an empirical comparison of
Reweighing, adversarial debiasing, threshold post-
processing techniques alongside flexible multi-objective
strategies
—
YODO (via AI Fairness 360 and Fairlearn
libraries)
—
considering acceptable accuracy losses. The
root source of unfairness remains data bias; hence, pre-
processing must be undertaken (rebalancing, synthetic
oversampling), while in- and post-processing can
essentially harmonize group metrics at some cost in
accuracy reduction Furthermore, without continuous
online monitoring and documentation (datasheets,
model cards), the balanced model risks losing fairness
due to dynamic feedback effects. Bringing together
technical fixes with rules and making the audit process
official ensures the ability to copy and openness, which
is key for long-term faith in AI systems. This article will
help
machine-learning
builders,
AI-responsibility
experts, and checkers find ways to find, gauge, and
lessen algorithmic bias in live models.
The American Journal of Engineering and Technology
193
https://www.theamericanjournals.com/index.php/tajet
Keywords:
algorithmic bias, fairness, pre-processing, in-
processing, post-processing, demographic parity,
equalized odds, disparate impact, AI Act, NIST AI RMF,
model cards.
Introduction:
Algorithmic bias refers to the systematic
error of running a machine-learning model in such a way
that causes members of different social groups to
receive drastically different predictions or decisions. It
develops when a model takes on historic inequities in
the data, amplifies them through its training process, or
applies them to new situations where implicit
correlations stand in for causal connections.
Consequently, some group
s are perceived as “risk
-
neutral” by default and others as “high risk” a priori,
though in reality, event probabilities are equal. This is
precisely how the National Institute of Standards and
Technology (NIST) describes the problem
—
as a
consequence of qua
ntitative methods “flattening” rich
social context into numerical categories, creating an
illusion of objectivity
—
while IBM defines it as
“systematic errors that produce unfair outcomes” [1, 2].
Concurrent legal, reputational, and social effects
confirm the significance of this issue for business and
society. The legal risk is evident: under the EU AI Act,
violations of the prohibition against discriminatory
practices may incur fines of up to €35 million or 7% of a
company’s global annual turnover [2]. Repu
tational
damage is measured in lost trust: 86% of surveyed
organizations believe customers prefer brands that
transparently apply ethical principles to their AI systems
[3]. The social cost manifests in concrete human lives:
analysis of the COMPAS tool showed that non-recidivist
Black defendants were almost twice as likely as white
defendants (45% vs. 23%) to be incorrectly classified as
“high risk” of reoffending [4]. Together, these facts
demonstrate that ignoring bias not only exacerbates
existing inequalities but also creates direct financial
losses and legitimacy threats for companies in the eyes
of society.
MATERIALS AND METHODOLOGY
The materials and methodology of this study are based
on a critical review of 29 publications from academic
journals, industry reports, and regulatory documents.
The theoretical foundation employs definitions of
algorithmic bias from NIST and IBM, emphasizing
systematic errors that lead to unfair outcomes [1, 2] and
an empirical analysis of the COMPAS tool demonstrating
real cases of discrimination in judicial predictions [4]. To
detect data biases, we analyzed model performance
across groups. Specifically, we compared top-5
classification accuracy on ImageNet for images from
regions with different income levels [5] and examined
gender-recognition errors in commercial systems across
“race–gender” combinations [6].
To mitigate bias, three classes of technical strategies
were considered. The first line of defense comprises pre-
processing methods, such as Reweighing, that adjust the
weights of training-set instances without altering the
algorithm, achieving a disparate impact of 1.0 on the
Adult dataset [18, 19]. The second class includes in-
processing techniques that embed fairness constraints
directly into the loss function: adversarial debiasing
achieved equalized odds parity with no more than a 2%
reduction in overall accuracy [8]. The third line entails
post-processing algorithms that adjust model output
probabilities via threshold optimization to balance
group error rates [20].
The legal and regulatory justification of the approach is
ensured by mapping these technical practices to the
requirements of the EU AI Act (mandatory dataset audit
and discrimination checks, Art. 10) [10], NIST AI RMF
recommendations (category “harmful bias”) [11, 27],
Canada’s Algorithmic Impact Assessment [12], the ICO’s
GDPR and AI guidance [13], Singapore’s Model AI
Governance Framework [14], the UK Financial Conduct
Authority’s directives for the financial sector [15], and
the ISO/IEC 42001:2023 standard on continuous fairness
monitoring [16]. These documents draw upon the OECD
principles for eliminating unfair bias in AI systems [17].
RESULTS AND DISCUSSION
Algorithmic bias almost always begins with data bias: if
individual countries, income brackets, or social groups
are underrepresented in the training set, the model
inevitably absorbs the statistical skew. Analysis [5]
showed that for six popular ImageNet classifiers, top-5
accuracy on objects from households with monthly
incomes below USD 50 is on average 10% lower than on
images from the wealthiest categories, and the gap
widens for scenes from non-Western regions, as shown
in Fig. 1 [5].
The American Journal of Engineering and Technology
194
https://www.theamericanjournals.com/index.php/tajet
Fig. 1. Top-5 Accuracy by Income [5]
Such “blind spots” are not accidental: they reflect a
historical research focus on English-language Internet
content and commercially attractive markets. If these
imbalances are not counteracted by rebalancing,
synthetic oversampling, or causal justification of
features, subsequent development stages can only
mitigate rather than eliminate the root cause.
A model can err even with a formally balanced sample
due to measurement distortions. A classic example is
sensor inaccuracies or manual annotation errors that
correlate with appearance. In study [6], commercial
gender-recognition systems misclassified dark-skinned
women in 34.7% of cases, whereas for light-skinned men
the error was only 0.8%. Because the algorithm “sees”
incorrect or noisy labels as truth, subsequent training
merely
entrenches
these
differential
errors,
transforming them into systematic discrimination.
Another important source of bias is the choice of
objective function and evaluation metrics. In the widely
used patient stratification algorithm studied in [7],
healthcare cost was used as a proxy for health status.
The metric that optimally reflected costs proved poorly
correlated with actual care needs, and even a perfect
model under this formulation inevitably produces a
biased outcome.
Even with a correctly specified task, the model
architecture and hyperparameter settings influence the
error distribution. Overly aggressive regularization or
skewed class-weight coefficients can shift the decision
boundary so that gains in overall accuracy come at the
expense of a higher false-negative rate for the
vulnerable group. Study [8] showed that post-processing
a single decision by selecting differentiated thresholds
can equalize false-positive and true-positive rates
between groups at the cost of a moderate loss in overall
accuracy of a couple of percentage points. This
underscores that fairness concerns must be addressed
in the data and the very “wiring” of the algorithm.
Finally, feedback loops can quickly bias even a perfectly
calibrated model after deployment. In [9], the PredPol
predictive-policing system, after only a few iterations,
began
directing
police
almost
exclusively
to
neighborhoods where arrests had already been
recorded, amplifying the divergence between observed
and actual crime activ
ity. Since the model’s actions
generate the subsequent training set, even a slight initial
bias accumulates exponentially. Such dynamic effects
require online monitoring and active “continuous”
debiasing methods; otherwise, any static fairness
assessment rapidly becomes outdated.
The American Journal of Engineering and Technology
195
https://www.theamericanjournals.com/index.php/tajet
Regulatory efforts to reduce algorithmic bias form a
multilayered system in which supranational norms set
minimal requirements, and sectoral and national
documents refine them for specific risks. Today, this
system's center is Regulation (EU) 2024/1689, the EU AI
Act. For “high
-
risk” systems, it introduces a mandatory
dataset audit, discrimination checks before deployment,
and a requirement to maintain detailed documentation
on data collection and annotation processes, enshrined
in
Art. 10 “Data Governance” [10].
Suppose the European approach relies on strict
enforcement in the United States. In that case, the
voluntary but widely adopted NIST AI Risk Management
Framework serves as a “de facto standard.” Since its
publication on January 26, 2023, the document has
defined a risk matrix in which “harmful bias” is
highlighted as one of five key categories; by summer
2024, NIST had added a separate profile for generative
models, identifying even more new risks, including
erroneous content personalization [11].
Several governments and sectoral regulators are
building their complementary mechanisms. In Canada,
all federal algorithms are subject to a mandatory
Algorithmic Impact Assessment: 51 risk-related
questions and 34 mitigation measures allow systems to
be classified into four impact levels, with proportional
bias-handling requirements [12]. In March 2023, the UK
Information Commissioner’s Office updated its “AI and
Data Protection” guidance, detailing how to assess and
mitigate bias at every stage of the model lifecycle and
permitting the processing of sensitive data for
discrimination testing [13]. In May 2024, Singapore
released the “Model AI Governance Framework for
Generative AI,” dedicating chapters to data provenance
and independent testing, and recognizing bias
mitigation as one of nine pillars of “trust” [14]. In 2024,
the UK Financial Conduct Authority integrated the risk of
unfair outcomes into its overall oversight of credit-
scoring models [15], and the ISO/IEC 42001:2023
inter
national standard proposed a managerial “overlay”
for all AI processes, including mandatory fairness-metric
monitoring [16].
For international alignment, these frameworks
draw on the OECD principles, which since 2019 have
emphasized the need to eliminate “unfair bias” and by
May 2023 had inspired over 1,000 policies across 70
jurisdictions [17]. The common thread is a risk-based
approach: the higher the potential social harm, the more
detailed the data checks, model transparency, and legal
safeguards must be. As a result, companies operating
globally effectively climb a unified “compliance ladder”:
from NIST’s voluntary metrics and industry guides
through ISO 42001 certification to the legally binding
requirements of the EU AI Act. This evolutionary logic
reduces regulatory fragmentation and shifts the fight
against algorithmic bias from ethical declarations to
measurable, verifiable obligations.
The regulatory requirement to measure and mitigate
bias moves the issue from abstract ethics into practical
engineering solutions, so developers rely on three
classes of technical strategies. The first line of defense is
pre-processing methods that correct the data before
training the model. In practice, this may be simple
weight rebalancing: for the Adult Income dataset, the
initial disparate-impact ratio between men and women
was 0.36, and after applying Reweighing, it became
1.0
—that is, statistically “discrimination
-
free” [18]. In
clinical prediction of postpartum depression, the same
technique raised disparate impact from 0.31 to 0.79 and
almost eliminated the difference in true-positive rates
between racial groups while preserving model accuracy
[19]. A comparison of bias metrics on the test dataset
—
using a baseline model, a race-blind model, a model
debiased via Reweighing, and a model debiased via
Prejudice Remover (logistic regression)
—
is shown in Fig.
2. Such methods require no algorithmic changes. Still,
their efficacy is limited to cases where bias resides
entirely in the data.
The American Journal of Engineering and Technology
196
https://www.theamericanjournals.com/index.php/tajet
Fig. 2. Comparison of bias metrics [19]
If reweighing proves insufficient, one moves to in-
processing techniques that embed fairness directly into
the loss function. The most popular approach is
adversarial debiasing: alongside the primary predictor, a
discriminator is trained to infer the protected attribute,
and the predictor’s objective is to make accurate
forecasts while obfuscating valuable information to the
discriminator. On the Adult Income dataset, this scheme
improved disparate impact and reduced average-odds
difference to nearly zero with only a 2% drop in overall
accuracy [18]. Adversarial methods provide the most
incredible group parity but require gradient access to
the model and can be unstable without careful tuning.
The third line comprises post-processing algorithms that
modify the obtained predictions without retraining the
model. A classic example is a linear program that adjusts
predicted probabilities to equalize false-positive and
false-negative rates between privileged and vulnerable
groups while leaving test power almost unchanged [20].
This “black
-
box” approach is especially valuable when
the original model is proprietary or frozen, but it is
limited to binary classification tasks and sensitive to
threshold choices.
Specialized libraries exist to facilitate the rapid
integration of all three tactics. IBM AI Fairness 360
implements ten mitigation algorithms covering the full
pre-, in-, and post-processing spectrum. It provides 70
metrics for evaluating group and individual fairness,
making it the most comprehensive open platform [21].
A lighter but actively developed alternative is
Microsoft’s Fairlearn. Thus, a developer can execute
Reweighing or Adversarial Debiasing in AIF360 with a
few lines of code, compare results with Fairlearn
metrics, and document the trade-off between accuracy
and fairness, thereby ensuring compliance with both
regulatory minima and internal corporate-responsibility
The American Journal of Engineering and Technology
197
https://www.theamericanjournals.com/index.php/tajet
standards.
Classic fairness metrics
—
demographic parity, equalized
odds, and predictive calibration
—
measure statistical
dependencies but ignore causal links between features
and protected attributes. Consequently, satisfying them
all simultaneously is mathematically impossible outside
trivial cases; the “fairness impossibility theorem” proves
that the three most popular criteria cannot be achieved
simultaneously, necessitating a trade-off in real-world
data [22]. In healthcare, the choice of a “convenient”
proxy label illustrated how external measures can
mislead: the algorithm optimized for treatment costs
enrolled only 17.7% of Black patients into additional
support instead of the clinically justified 46.5% (Fig. 3),
because historically less was spent on Black patients [7].
Fig. 3. Number of chronic illnesses versus algorithm-predicted risk, by race [7]
To move beyond purely correlational criteria,
counterfactual fairness is employed: a decision is
deemed fair if it would remain unchanged in a
hypothetical world where the individual belongs to a
different group under the same risk factors. Formalized
via structural causal models, this approach “removes”
group-only associations. In a classical experiment
predicting
law-student
performance,
the
counterfactually fair “Fair Add” model reduced root
-
mean-square error from 0.873 to 0.918 (i.e., lost about
5%)
—
but eliminated prediction dependence on race:
when the protected attribute was swapped, grade
distributions coincided completely, whereas the
baseline model exhibited a systematic shift [23]. This
example demonstrates that a small accuracy cost can
radically reduce hidden discrimination.
Organizations aim to navigate the accuracy
–
fairness
trade-off rather than fix a single configuration. Modern
methods combine both criteria into a unified loss
The American Journal of Engineering and Technology
198
https://www.theamericanjournals.com/index.php/tajet
function or treat them as a multi-objective problem. The
You Only Debias Once (YODO) approach trains the
model simultaneously on two extremes
—
accuracy-
optimum and fairness-optimum
—
and finds in weight
space a “line” of solutions along which the balance can
be adjusted at inference time. For the ACS-E dataset,
generating one hundred Pareto points took 3.53 s
instead of 425 s for training one hundred separate
models, with each solution remaining on the same
“error ↔ demographic parity” front [24]. Combined or
multi-objective
optimizations
do
not
override
theoretical limits but provide managers with a
transparent navigation tool, allowing the selection of a
point acceptable to business, legal, and social-
responsibility requirements simultaneously.
Once a company has identified and mapped bias sources
to regulatory requirements, the next task is to formalize
a reproducible bias-management process. In practice,
this begins with a full-scale audit and data profiling:
technical specialists verify sample representativeness,
assess annotation quality, and identify signs of historical
bias, while internal auditors record checkpoints.
Regulators and professional communities already
consider such an audit standard: ISACA defines
algorithmic audit as a key method for detecting bias at
“all points of the model lifecycle” [25]. However,
real-
world adoption remains limited: only 47.2% of
organizations working with generative AI conduct
regular checks [26]. These figures indicate that a missing
audit stage remains the most significant “gap” in
discrimination protection.
The next step is to define which model use cases are
critical and which metrics fairness will be measured.
NIST AI RMF recommendations propose starting with a
harm map: first, describe which groups may be harmed,
and only then select a statistical criterion
—
demographic
parity, equalized odds, or individual fairness
—
that best
reflects that risk [27]. This sequence helps avoid
optimizing a “convenient” metric unrelated to social
harm. The data team then conducts a series of
controlled experiments: applying pre-, in-, and post-
processing methods to the baseline model, with results
displayed on the accuracy
–
fairness trade-off surface.
Integration via AIF360 and Fairlearn reduces the
“hypothesis → evaluation” cycle to minutes, enabling
product managers to make decisions based on a
complete picture of trade-offs.
When an acceptable configuration is found, the results
are documented. For datasets, datasheets are created
describing provenance and limitations; model cards
present
group-specific
metrics
and
safe-use
recommendations for models. Such documents are
already hailed as a “selection tool” for AI transparency,
and their use is piloted by large corporations and
industry consortia [28]. A standardized card greatly
simplifies internal reviews and regulator interactions: all
key assumptions and tests are collected in one place.
The final stage is deployment and continuous online
monitoring. Uber’s practical experience showed that
without automated tracking of data shifts and spikes in
group-specific
error
rates,
incorrect
decisions
accumulate unnoticed until reaching a crisis threshold
[29]. Thus, the fight against bias transitions from one-off
initiatives to ongoing operations: the model,
documentation, and monitoring form a unified control
chain in which a failure at any link is quickly detected and
remedied.
Thus, algorithmic bias permeates all stages of model
development
—
from unevenly represented data and
distorted metrics to architectural decisions and
feedback loops in production
—
and requires a
comprehensive approach. On one hand, at the level of
regulatory governance (from the EU AI Act to national
frameworks and ISO standards), mandatory audits,
transparency
requirements,
and
accountability
measures have already been established; on the other,
technical methods (pre-, in-, and post-processing, causal
and multi-objective optimization) enable minimization
of imbalances both during development and after
deployment. Finally, introducing systematic checks,
“harm maps,” datasheets, and model cards transforms
the struggle against bias from a mere declaration into an
ingrained process that ensures reproducibility and
accountability. In the conclusion, we will articulate key
recommendations for creating truly fair and reliable
predictive models.
CONCLUSION
In conclusion, it has been demonstrated that algorithmic
bias is a complex issue permeating every stage of a
model’s lifecycle: from data collection and annotation
The American Journal of Engineering and Technology
199
https://www.theamericanjournals.com/index.php/tajet
through the selection of target metrics, architectural
configuration, and deployment in a production
environment. Sources of bias may include historical
imbalances in the data and annotation noise that
become entrenched during training, as well as
improperly chosen proxy variables and metrics that fail
to account for the actual needs of protected groups.
Moreover, even a correctly trained model is susceptible
to feedback-loop effects in production, which amplify
the initial bias in the absence of continuous monitoring.
Achieving fairness requires diverse technical techniques:
pre-, in-, and post-processing methods, each addressing
a specific subtask. Pre-processing reduces data skew;
embedding fairness constraints into the loss function
enables explicit consideration of equity requirements
during training; and post-
processing provides a “black
-
box” mechanism for balancing errors when access to the
model’s internal parameters
is limited. However, none
of these approaches offers a universal solution: a trade-
off between accuracy and fairness is inevitable, and the
specific business and social context must determine the
optimal balance.
Equally important is the incorporation of auditing,
documentation, and continuous control processes: from
preliminary dataset profiling and metric selection to the
publication of datasheets and model cards that record
assumptions and test outcomes for different groups.
Only a formalized, reproducible process will allow
regulators and internal auditors to verify compliance
with
bias-mitigation
obligations
and
enable
organizations to respond promptly to emerging
deviations in fairness metrics.
Finally, international, regional, and sectoral regulatory
frameworks establish minimal requirements and create
a “compliance ladder” ranging from NIST’s voluntary
recommendations to the mandatory audits under the
EU AI Act. This evolutionary structure reduces
fragmentation and facilitates the shift from ethical
declarations to measurable, verifiable commitments.
Thus, an effective strategy for combating algorithmic
bias must integrate technical mitigation methods,
auditing and documentation processes, continuous
monitoring, and regulatory compliance mechanisms.
These elements will ensure the reliability and fairness of
predictive models over the long term.
REFERENCES
R. Schwartz, A. Vassilev, K. Greene, L. Perine, A. Burt, and
P. Hall, “Towards a Standard for Identifying and
Managing Bias in Artificial Inte
lligence,”
Towards a
Standard for Identifying and Managing Bias in Artificial
Intelligence
, vol. 1270, no. 1270, Mar. 2022, doi:
https://doi.org/10.6028/nist.sp.1270
A. Jonker and J. Rogers, “What is algorithmic bias?”
IBM
,
Sep.
20,
2024.
https://www.ibm.com/think/topics/algorithmic-bias
(accessed Apr. 18, 2025).
A. Davison, “AI ethics tools,”
IBM
, Sep. 03, 2024.
https://www.ibm.com/think/insights/ai-ethics-tools
(accessed Apr. 19, 2025).
J. Larson, S. Mattu, L. Kirchner, and J. Angwin, “How We
Analyzed
the
COMPAS
Recidivism
Algorithm,”
ProPublica
,
May
23,
2016.
https://www.propublica.org/article/how-we-analyzed-
the-compas-recidivism-algorithm
2025).
T. Devries, I. Misra, and C. Wang, “Does Object
Recognition Work for Everyone?,” The CVPR. Accessed:
Apr.
21,
2025.
[Online].
Available:
J. Buolamwini and T. Gebru, “Gender Shades:
Intersectional Accuracy Disparities in Commercial
Gender Classification,”
Proceedings of Machine Learning
Research
, vol. 81, no. 1, pp. 1
–
15, 2018, Accessed: Apr.
22,
2025.
[Online].
Available:
https://proceedings.mlr.press/v81/buolamwini18a/buo
lamwini18a.pdf
Z. Obermeyer, B. Powers, C. Vogeli, and S. Mullainathan,
“Dissecting racial bias in an algorithm used to manage
the health of populations,”
Science
, vol. 366, no. 6464,
pp. 447
–
453, Oct. 2019, Accessed: Apr. 03, 2025.
[Online].
Available:
https://www.ftc.gov/system/files/documents/public_e
vents/1548288/privacycon-2020-ziad_obermeyer.pdf
The American Journal of Engineering and Technology
200
https://www.theamericanjournals.com/index.php/tajet
M. Hardt, E. Price, and N. Srebro, “Equality of
Opportunity in Supervised L
earning,”
Arxiv
, Oct. 07,
2016.
https://arxiv.org/abs/1610.02413v1
Apr. 23, 2025).
D. Ensign, S. Friedler, S. Neville, C. Scheidegger, S.
Venkatasubramanian, and C. Wilson, “Runaway
Feedback Loops in Predictive Policing,”
Proceedings of
Machine Learning Research
, vol. 81, 2018, Accessed:
Apr.
23,
2025.
[Online].
Available:
https://proceedings.mlr.press/v81/ensign18a/ensign18
a.pdf
European Parliament,
P9_TA(2024)0138 Artificial
Intelligence Act
. 2024. Accessed: Apr. 23, 2025. [Online].
Available:
https://www.europarl.europa.eu/doceo/document/TA-
9-2024-0138_EN.pdf
[
“Artificial Intelligence Risk Management Framework:
Generative Artificial Intelligence Profile,”
NIST
, 2024,
doi:
https://doi.org/10.6028/nist.ai.600-1
“Algorithmic Impact Assessment Tool,”
The Government
of
Canada
,
May
30,
2024.
“Guidance on AI and data protection,”
ICO
, Jun. 13,
2023.
https://ico.org.uk/for-organisations/uk-gdpr-
guidance-and-resources/artificial-
intelligence/guidance-on-ai-and-data-protection/
(accessed Apr. 24, 2025).
“Model AI Governance Framework for Generative AI,”
AI
Verify
Foundation
,
May
30,
2024.
(accessed Apr. 25, 2025).
J. Rusu, “AI Update,” FCA. Accessed: Apr. 24, 2025.
[Online].
Available:
https://www.fca.org.uk/publication/corporate/ai-
update.pdf
R. P. Grubenmann, “ISO/IEC 42001: The latest AI
management
system
standard,”
KPMG
,
2024.
https://kpmg.com/ch/en/insights/artificial-
intelligence/iso-iec-42001.html
2025).
OECD,
“AI
Principles,”
OECD
,
2024.
https://www.oecd.org/en/topics/ai-principles.html
(accessed Apr. 25, 2025).
H. Mahmoudian, “Reweighing the Adult
Dataset to
Make it ‘Discrimination
-
Free,’”
Medium
, Apr. 14, 2020.
Y. Park
et al.
, “Comparison of Methods to Reduce Bias
From Clinical Prediction Models of Postpartum
Depression,”
JAMA Network Open
, vol. 4, no. 4, p.
e213909,
Apr.
2021,
doi:
https://doi.org/10.1001/jamanetworkopen.2021.3909
P. Awasthi, M. Kleindessner, and J. Morgenstern,
“Equalized odds postprocessing under imperfect group
information,” in
Proceedings of the 23rd International
Conference on Artificial Intelligence and Statistics
, PMLR,
2020. Accessed: Apr. 28, 2025. [Online]. Available:
https://proceedings.mlr.press/v108/awasthi20a/awast
hi20a.pdf
“Understand and mitigate bias in ML models,”
AI
Fairness 360
29, 2025).
B. Hsu, R. Mazumder, P. Nandy, and K. Basu, “Pushing
the limits of fairness impossibility: Who’s the fairest of
them all?”
36th Conference on Neural Information
Processing Systems
, 2022, Accessed: May 18, 2025.
[Online].
Available:
M. Kusner, J. Loftus, C. Russell, and R. Silva,
“Counterfactual Fairness,”
Proceedings of the 31st
Conference on Neural Information Processing Systems
,
2017, Accessed: May 04, 2025. [Online]. Available:
The American Journal of Engineering and Technology
201
https://www.theamericanjournals.com/index.php/tajet
X. Han, T. Chen, K. Zhou, Z. Jiang, Z. Wang, and X. Hu,
“You Only Debias Once: Towards Fl
exible Accuracy-
Fairness Trade-
offs at Inference Time,”
Arxive
, Mar. 10,
2025.
https://arxiv.org/pdf/2503.07066
06, 2025).
[25]
V. Prasad
, “AI Algorithm Audits: Key Control
Considerations,”
ISACA
,
Aug.
02,
2024.
H. Dhaduk, “State of Generative AI in 2024,”
Simform
,
Apr. 02, 2024.
https://www.simform.com/blog/the-
“AI Risk Management Framework,”
NIST
, Jan. 2023, doi:
https://doi.org/10.6028/nist.ai.100-1
“Datasheets for Datasets: Impact and Adoption Across
Academic and Industry Sectors,”
Hackernoon
, Jun. 11,
2024.
https://hackernoon.com/datasheets-for-
datasets-impact-and-adoption-across-academic-and-
industry-sectors
J. Le, “Datacast Ep
isode 67: Model Observability, Ai
Ethics, And Ml Infrastructure Ecosystem With Aparna
Dhinakaran,”
James
Le
,
Jun.
28,
2021.
https://jameskle.com/writes/aparna-dhinakaran
(accessed May 18, 2025).
