AI Regulation and Ethics

Written by Yasin TANIŞ

Originally published on Medium

Scales of justice resting on a robotic hand, with circuit lines and a courthouse in the background

1. Why AI regulation matters

AI technologies, among the greatest innovations of the 21st century, are transforming both individual lives and business. The ethical and legal problems that come with this transformation are a critical factor in shaping the technology’s future. Companies, governments and individuals all have to deal with AI’s risks as well as its benefits.

A few examples:

In this article we look at where AI regulation stands today, how to develop AI ethically, and how companies can adapt to the new rules.

2. The EU AI Act up close

The European Union’s AI Act stands out as the world’s first comprehensive AI regulation. Adopted in 2024, it changes the game both for companies in Europe and for companies anywhere that sell into the EU market. It takes a risk-based approach and sorts AI systems into four tiers: prohibited, high-risk, limited-risk and minimal-risk applications.

High-risk systems

These systems are used in critical areas that directly affect people’s lives, and they must meet strict oversight and transparency requirements. Examples:

  • Healthcare: algorithms that diagnose disease
  • Finance: credit scoring systems
  • Hiring: CV screening and candidate assessment tools

Oversight mechanisms include data records, model explainability and continuous monitoring. A bank’s credit scoring algorithm, for example, must be able to document clearly why a customer was refused credit.

Prohibited applications

The Act strictly bans AI applications that run against society’s core values, including:

Limited and minimal risk

For limited-risk systems such as chatbots, the main duty is transparency: users must know they are talking to an AI. Minimal-risk systems such as in-game AI or spam filters are largely left free; ethical principles still matter here, but regulation is much lighter.

Compliance requirements for companies

For high-risk systems, the Act requires companies to take these steps:

  1. Documentation and transparency: for every model, documents describing the data used, how the algorithm works and its possible risks.
  2. Independent audit: high-risk systems should be audited regularly by independent bodies.
  3. Ethical and technical training: staff, especially technical teams, should be trained on regulation and ethical principles.

Penalties and oversight

Companies that do not comply can face serious penalties:

  • Turnover-based fines: up to 7% of global annual turnover for the most serious breaches
  • Suspension: in a serious breach, use of the AI system concerned can be stopped
  • Withdrawal: models that fail ethical or technical standards can be taken off the market

These penalties can cause heavy financial and reputational losses. A diagnostic algorithm that produces wrong results, for instance, can shake patients’ trust and turn into a serious reputational crisis.

3. Two examples from practice

To show how regulation and ethics play out in practice, here are two examples from finance and healthcare. They describe illustrative scenarios, not real organisations.

Example 1: AI in finance (fictional “XYZ Bank”)

XYZ Bank starts using an AI-based algorithm for credit decisions. In its first rollout, the model turns out to be biased against certain demographic groups.

The problem: the model rejects applications from women entrepreneurs under 35 23% more often than those from men. That means both legal risk and a loss of customer trust.

The fix:

  1. Data analysis and cleaning: historical data is audited, the bias identified, and the gap in data on women entrepreneurs filled.
  2. Bias mitigation: fairness metrics are used in training.
  3. Continuous monitoring: a weekly dashboard tracks drift in the algorithm’s decisions.

The results:

  • The demographic parity score rises from 0.82 to 0.96.
  • The gender gap in credit approval falls to 4%.
  • Model performance (AUC-ROC) holds at 87%.

Example 2: AI in healthcare (fictional “ABC Hospital”)

ABC Hospital uses an AI model for diagnosis. It emerges that patient data is not adequately protected, in breach of the GDPR (General Data Protection Regulation).

The problem: patient data is held on a central server, an open risk for data breaches.

The fix: the hospital moves to federated learning:

  • Data is processed on local devices, removing the need for central storage.
  • The model learns from data on different devices, but no data ever leaves a device.

The results:

  • Privacy breaches are prevented.
  • Model accuracy rises from 91% to 93%.
  • The system becomes GDPR-compliant.

These examples show how compliance with regulation and ethics can work in the real world. Next, how to repeat results like these.

4. A framework for ethical AI development

An ethical AI system should rest on a well-defined framework covering data collection, model development, monitoring and explainability.

Ethical checkpoints in model development

Data collection and processing standards: data should be unbiased, current and representative, and personal data must be protected during processing. Sensitive attributes such as gender or age can, for example, be removed from the data set.

Detecting and preventing bias: bias checks should run during training. The code below sketches what such a check might look like:

# Bias test framework
from fairness_check import calculate_bias_metrics
bias_metrics = calculate_bias_metrics(data, protected_features=["gender", "ethnicity"])

Good practice for explainability: models should be easy for users and auditors to understand.

  • Explain feature importance with SHAP (SHapley Additive exPlanations) values
  • Prepare a detailed model card for every model:
model_card = {
    "model_name": "Disease Predictor v1.0",
    "description": "Hastalık teşhisi için kullanılan bir AI modeli",
    "fairness_metrics": {"demographic_parity": 0.95},
    "limitations": ["Sınırlı veri temsil durumu"]
}

Documentation requirements

Every AI system needs detailed documentation:

  • Data sources and processing steps
  • Parameters used in training
  • Fairness and performance metrics
  • Monitoring and audit mechanisms in place

Continuous monitoring and updating

AI systems are not static; performance and ethical compliance should be checked regularly:

  • Set up real-time monitoring.
  • Add alerts for drops in performance and fairness metrics.

5. The situation in Türkiye

Türkiye is working to align with international standards on AI regulation and ethics. The current framework is shaped mainly by the KVKK (Personal Data Protection Law), and the EU AI Act is becoming increasingly influential. The alignment process brings both opportunities and challenges.

The current legal framework

Türkiye has no AI-specific law yet, but the KVKK provides an important framework:

  1. KVKK Article 4: principles of accuracy, currency and lawfulness in processing personal data
  2. KVKK Article 6: stricter conditions and explicit consent for special categories of personal data

AI systems used in healthcare, for example, must comply with Article 6 when processing patient data. The banking regulator BDDK and the telecoms regulator BTK also publish guidance on AI in their sectors.

Aligning with the EU AI Act

Turkish companies that offer products or services in the EU market are already subject to the Act, and because of Türkiye’s customs union with the EU, its own legislation is expected to move in the same direction. This calls for:

  • New oversight mechanisms in sectors such as banking, telecoms and healthcare
  • Better transparency and documentation processes in companies

Expected changes:

  • Stricter data protection protocols
  • Mandatory independent audits for high-risk systems
  • Structural reforms such as ethics committees

Sector practice

  • Banking: BDDK rules on data protection and algorithmic transparency stand out in finance; some banks have started having independent third parties audit their credit algorithms.
  • Healthcare: the Ministry of Health has published guidance on processing patient data with AI, and approval processes are being developed for clinical diagnostic algorithms.
  • Telecoms: BTK guidance requires customer data protection and anonymisation standards.

Türkiye’s AI ecosystem is growing fast, but more work is needed on regulation and ethical standards. Next, what companies can do to comply.

6. Recommendations for implementation

Complying with regulation and ethics takes both strategic and operational steps. Here is a guide.

Compliance checklist

CheckpointDescriptionDone?
Data inventory completed?Have all data sources in use been identified?☐
Model documentation in place?Are the models’ technical details on record?☐
Risk assessment done?Has each system’s risk tier been determined?☐
Audit mechanism set up?Are continuous monitoring and audit processes running?☐

A list like this makes compliance systematic.

Risk assessment method

Companies should assess the AI systems they use by risk. An example risk matrix:

Risk tierExample useRequired measures
High riskMedical diagnosisHuman oversight mandatory
Medium riskCredit scoringRegular audit
Low riskRecommendation systemsBasic monitoring

A diagnostic algorithm used in a hospital, for example, should be checked under human supervision every week.

Internal audit and documentation

Companies should document the whole life cycle of their AI systems in detail. A documentation template should cover at least:

  • Data sources used
  • How the algorithm works and decides
  • Ethical compliance and bias mitigation processes

Detailed documentation like this makes audits easier and lowers legal risk.

7. Good practice and tools

Many tools can help make AI systems compliant and ethical. They guide critical areas such as bias detection, explainability and continuous monitoring.

Bias testing tools

  • IBM AI Fairness 360: bias detection and mitigation techniques, especially effective for credit scoring and hiring models.
  • Google What-If Tool: visualises the effect of model decisions by testing scenarios, with a user-friendly interface.
  • Microsoft Fairlearn: analyses fairness metrics and suggests improvements.

Example usage:

from aif360.datasets import BinaryLabelDataset
from aif360.metrics import BinaryLabelDatasetMetric
dataset = BinaryLabelDataset(df, 'target', protected_attribute_names=['gender'])
metric = BinaryLabelDatasetMetric(dataset)
print(metric.disparate_impact())

These tools are critical for finding and removing sources of bias in models.

Explainability tools

Making model decisions explainable is needed for user trust and for compliance.

  • SHAP: shows which features a model’s decision rests on; widely used in finance and healthcare.
  • LIME: produces local explanations for any model.

Example SHAP usage:

import shap
explainer = shap.TreeExplainer(model)
shap_values = explainer.shap_values(data)
shap.summary_plot(shap_values, data)

Monitoring systems

Continuous monitoring is critical for keeping systems performant and ethically compliant.

  • MLflow: tracks training, evaluation and deployment, and catches changes in performance.
  • Weights & Biases: real-time monitoring and visualisation, with alerts for data drift.
  • TensorBoard: tracks performance metrics during training.

Example data drift check:

from sklearn.metrics import roc_auc_score
# Eğitim ve gerçek zaman verileriyle performans karşılaştırma
auc_train = roc_auc_score(y_train, model.predict(X_train))
auc_live = roc_auc_score(y_live, model.predict(X_live))
if abs(auc_train - auc_live) > THRESHOLD:
    print("Uyarı: Veri kayması tespit edildi!")

Documentation template

Every model needs comprehensive documentation. An example model card:

model_name: Credit Risk Predictor
version: 1.0.0
data_sources:
  - source: Customer data (internal)
  - update_frequency: Monthly
performance_metrics:
  - accuracy: 89%
  - fairness_score: 95%
ethical_considerations:
  - bias_mitigation: Applied
  - privacy_measures: Anonymized data
limitations:
  - Rural data coverage limited
  - Requires manual oversight for high-risk cases

A document like this is valuable both in regulatory audits and in internal reviews.

Open-source frameworks

8. Conclusion and recommendations

AI regulation and ethical principles play a fundamental role in letting the technology fulfil its potential to change lives for the better. Rules such as the EU AI Act are a critical step towards developing systems transparently and responsibly. For companies, compliance is not only a legal obligation but a strategic opportunity to build customer trust and efficiency.

Future trends

  1. Global ethical standards: the EU AI Act will become a model for other countries, and similar rules are expected in countries like Türkiye.
  2. Automated compliance: automation and AI-assisted audit tools will spread.
  3. Ethics as business value: companies that embrace ethical principles will gain a competitive edge through consumer trust.

Steps you can take now

  1. Ensure legal compliance: develop a plan in line with the EU AI Act and the KVKK.
  2. Set up ethics committees: clarify responsibilities from senior management to technical teams.
  3. Use tools and frameworks: build bias checks, explainability and monitoring into your processes.
  4. Run training programmes: raise staff awareness of regulation and ethical principles.
  5. Establish continuous audit: check your systems’ performance and ethical compliance regularly.

Final advice

  • Start now: instead of waiting for a perfect plan, begin with what you have and improve as you go.
  • Value documentation: good documents reduce legal risk and improve your systems.
  • Keep improving: AI systems are not static, so review and update them regularly.
  • Be inclusive: make sure your systems are fair and accessible to everyone.

Afterword and thanks

At the start of a project you often see only the goal; every difficulty and success along the way brings new perspectives. In this piece, which I worked on carefully for a long time, I looked beyond the technical side of AI and beyond the questions of “how” and “what” to the question of “why”. The journey reminded me again and again how valuable the human side of technology is.

I was not alone on this road. My thanks go to everyone in the industry whose work appears in the sources, to visionary leaders and to the unnamed heroes behind the scenes. Real innovation can only come from shared effort and shared passion. For me this work is not an end but a new beginning, and I hope it will be for you too.

References

  1. BBC News (2023). Italy blocks ChatGPT over privacy concerns. bbc.com
  2. Brookings Institution (2023). Algorithmic bias detection and mitigation: best practices and policies to reduce consumer harms. brookings.edu
  3. TechCrunch (2024). EU AI Act: everything you need to know. techcrunch.com
  4. ForeSee Medical. Medical algorithms in healthcare. foreseemed.com
  5. DergiPark. Yapay zekâ ve hukuk. dergipark.org.tr
  6. TestGorilla. Resume screening software guide. testgorilla.com
  7. DergiPark. Yapay zekâ regülasyonları. dergipark.org.tr
  8. Liberties. Pros and cons of facial recognition. liberties.eu
  9. ResearchGate. Subliminal uyarıcıların nöropsikolojisi. researchgate.net
  10. EU Artificial Intelligence Act, Article 13. artificialintelligenceact.eu
  11. B3Lab. Artificial intelligence audit. b3lab.org
  12. Türkiye Yapay Zekâ İnisiyatifi (2024). Yapay Zekâ Etik İlkeleri ve Hukuki Düzenlemeler Raporu. turkiye.ai
  13. BDDK, Bankacılık Düzenleme ve Denetleme Kurumu. bddk.org.tr
  14. T.C. Sağlık Bakanlığı. saglik.gov.tr
  15. BTK, Bilgi Teknolojileri ve İletişim Kurumu. btk.gov.tr
  16. IBM. AI Fairness 360. aif360.res.ibm.com
  17. Google Research. The What-If Tool: interactive probing of machine learning models. research.google
  18. Microsoft Research. Fairlearn: a toolkit for assessing and improving fairness in AI. microsoft.com
  19. C3 AI. LIME: Local Interpretable Model-Agnostic Explanations. c3.ai
  20. TensorFlow. Responsible AI. tensorflow.org
  21. OpenMined. PySyft. openmined.org
  22. Hugging Face. Datasets documentation. huggingface.co
  23. SHAP documentation. shap.readthedocs.io
  24. MLflow documentation. mlflow.org
  25. Weights & Biases documentation. docs.wandb.ai
  26. TensorBoard. tensorflow.org
  27. ELI5 documentation. eli5.readthedocs.io
  28. InterpretML documentation. interpret.ml
  29. AI Fairness 360 documentation. aif360.readthedocs.io