Wednesday, 22 July 2026

From Algorithms to Justice: Foundations of AI and Automated Decision‑Making for AI Governance


Why these foundations matter

Artificial intelligence (AI) and machine learning (ML) now influence core public functions—from welfare allocation and credit scoring to hiring, criminal justice, and content moderation—making conceptual clarity essential for any AI governance or legal oversight role.
For judges, regulators, and policymakers, understanding AI, ML, deep learning, neural networks, generative AI, and automated decision‑making systems is a prerequisite to evaluating legality, fairness, transparency, and accountability in technology‑mediated decisions.

Understanding artificial intelligence (AI)

Artificial intelligence is broadly defined as computer systems that simulate human intelligence processes such as reasoning, learning, problem‑solving, perception, language understanding, and decision‑making.
NASA’s working definition (based on US federal law) describes AI as artificial systems that perform tasks under varying and unpredictable circumstances, learn from experience, and solve tasks requiring human‑like perception, cognition, planning, learning, communication, or physical action.
In practice, AI is an umbrella field that includes multiple sub‑disciplines such as machine learning, natural language processing (NLP), robotics, and computer vision, all aimed at producing systems that act intelligently or rationally toward specified goals.

Narrow, general, and superintelligent AI

Contemporary systems are almost entirely “narrow AI,” designed to execute specific tasks like spam filtering, facial recognition, or recommendation of movies on a streaming platform.
“Artificial general intelligence” (AGI) refers to hypothetical systems that could perform any cognitive task a human can perform, and “artificial superintelligence” (ASI) refers to systems surpassing human intelligence on most or all domains—both remain speculative concepts that mainly feature in research and ethics debates.

Machine learning (ML) basics

Machine learning is a subset of AI that focuses on algorithms which enable computers to learn patterns from data and improve performance over time without being explicitly programmed for every rule or decision.
A widely cited characterization is that ML uses data and algorithms to train models so they can detect regularities and then make classifications, predictions, or recommendations about new, unseen data.
ML is now widely used across domains such as fraud detection in banking, autonomous vehicles, medical image analysis, game‑playing agents (e.g., chess, Go), and credit scoring, demonstrating its central role in practical AI systems.

Core learning paradigms in ML

ML methods are commonly grouped into three main paradigms: supervised learning, unsupervised learning, and reinforcement learning.
In supervised learning, models are trained on labeled data—pairs of input and correct output—so the system learns to predict outputs (e.g., “spam/not spam,” “approved/rejected”) from new inputs.

In unsupervised learning, models receive unlabeled data and try to discover hidden structures such as clusters or latent factors (e.g., grouping customers by purchase behavior), while reinforcement learning trains an agent through trial‑and‑error interactions, rewarded or penalized based on its actions (e.g., learning optimal control strategies in robotics or game environments).

Neural networks and deep learning

Artificial neural networks (ANNs) are a family of ML models inspired by the interconnected neurons in the human brain, consisting of nodes organized in layers (input, one or more hidden layers, and output) with weighted connections that are tuned during training.
Each neuron receives inputs, applies a transformation, and passes an output to the next layer; through iterative training using methods such as backpropagation, the network learns to map inputs (images, text, tabular data) to outputs (labels, decisions, scores).
Neural networks underpin many modern AI applications, particularly in tasks like image recognition, speech recognition, and complex pattern analysis that are difficult to capture with manually designed rules.

What is deep learning?

Deep learning is a subfield of machine learning that uses neural networks with multiple hidden layers (“deep” architectures) to learn highly abstract and complex representations from data.
Google Cloud and other technical sources describe deep learning as a subset of ML that leverages large neural networks to process and analyze information, often outperforming traditional ML on tasks involving unstructured data like images, audio, or natural language.
Because deep learning can automatically discover relevant features instead of relying on manual feature engineering, it is particularly powerful when trained on very large datasets, but its models are often opaque (“black boxes”) and computationally intensive—issues that are central to AI governance debates.

Generative AI explained

Generative AI (GenAI) refers to AI systems that can create new content—text, images, audio, video, or code—that resembles the style or statistical properties of the data on which they were trained.
IBM defines generative AI as systems that produce original content in response to a user’s prompt or request, while Intel describes it as deep‑learning‑based AI that generates new and original content such as images, music, or text.
These models learn patterns and structures in massive training datasets and then generate novel outputs that are statistically consistent with those patterns, using architectures such as transformers (for large language models), generative adversarial networks (GANs), and diffusion models.

Key generative AI techniques and applications

Large language models (LLMs) like GPT‑style systems are trained on large corpora of text and can understand prompts, answer questions, draft documents, and perform reasoning tasks in natural language.
Generative image models like DALL‑E or similar systems can create new images from textual descriptions, while other generative models can compose music, synthesize voices, or generate code—uses that are increasingly relevant to content creation, education, and software development.
At the same time, generative AI can produce deepfakes and synthetic media that convincingly impersonate individuals, raising serious legal and ethical concerns related to fraud, defamation, electoral integrity, and evidentiary reliability.

Automated decision‑making and algorithmic decision systems

Automated decision‑making (ADM) and automated decision systems (ADS) describe computational processes—often based on ML, statistical modeling, data analytics, or AI—that generate outcomes such as scores, recommendations, classifications, or decisions that affect people and may assist or replace human discretion.
An ADS can, for example, rank job applicants, prioritize welfare fraud investigations, approve or reject loan applications, recommend medical treatments, or produce risk scores used in criminal justice for bail, sentencing, or parole decisions.
From a governance perspective, these systems are critical because they operate at scale, often without full transparency to those affected, and can embed biases or errors that have direct consequences for fundamental rights, equality, and access to services.

Governance risks and examples

Governance literature highlights that algorithmic decision‑making raises questions about accountability, explainability, fairness, and auditability, especially when used in sensitive domains like justice, healthcare, or employment.
Real‑world controversies include the Netherlands’ SyRI welfare fraud detection system, which was found illegal for violating human rights and disproportionately targeting low‑income neighborhoods, and US criminal justice risk assessment tools like COMPAS, which studies have shown to flag Black defendants as high risk at higher rates.
These examples illustrate why AI governance frameworks such as the EU AI Act and similar initiatives classify many ADS used in public decision‑making as “high‑risk” and insist on documentation, testing, monitoring, and human oversight.

Common model types: classification

Classification models predict categorical outputs such as “spam/not spam,” “approve/reject,” “high/medium/low risk,” or “positive/negative sentiment.”
Technically, classification is a supervised learning task where the model learns from labeled examples and then assigns one of several predefined categories to new data points, using algorithms like logistic regression, decision trees, random forests, support vector machines, or neural networks.
Real‑world applications include email spam detection, disease prediction, creditworthiness assessment, customer segmentation, face detection, and many other tasks where discrete decisions must be made based on observed features.

In governance contexts, classification models may be used to label transactions as “fraudulent/suspicious,” applicants as “eligible/ineligible,” or individuals as “high‑risk/low‑risk,” directly shaping administrative or judicial outcomes.
When such systems are deployed in areas like welfare eligibility, immigration screening, predictive policing, or pre‑trial risk assessment, classification errors or biased training data can lead to discriminatory impacts, making transparency, quality of training data, and appropriate appeals mechanisms essential.

Common model types: regression

Regression models predict continuous numerical values such as prices, probabilities, demand volumes, or monetary amounts, rather than categories.
In supervised learning terms, regression fits a function that maps input variables (e.g., age, income, past behavior) to a continuous output (e.g., predicted loan default probability, predicted insurance claim amount, expected sales revenue), often using linear regression or more complex variants.
Business and public‑sector use cases include forecasting expenses and budgets, predicting property values, modelling electricity consumption, estimating insurance claim costs, or predicting customer lifetime value, all of which inform pricing, resource allocation, and policy analysis.

Regression models can underlie risk‑based pricing in insurance, credit scoring, and dynamic tariffs, affecting access to essential services and raising fairness concerns when correlations mimic or amplify socio‑economic inequalities.
In public administration, regression‑based forecasting can guide staffing, infrastructure planning, and budget allocations, but governance frameworks must ensure that models are validated, periodically updated, and auditable to avoid entrenched errors or opaque assumptions.

Common model types: recommendation systems

Recommendation systems (recommender systems) use AI and ML to analyze large amounts of behavioral and content data and deliver personalized suggestions for products, services, or digital content.
Standard definitions describe recommender systems as a subtype of information filtering that predicts a user’s preferences based on past behavior or item characteristics, using approaches like collaborative filtering (users similar to you) and content‑based filtering (items similar to what you liked).
Platforms such as Netflix, Amazon, Spotify, and YouTube rely heavily on recommender systems to suggest movies, products, songs, or videos that users are likely to engage with, thereby driving retention and revenue.

In the governance discourse, recommender systems are central to debates about algorithmic curation, filter bubbles, and online harms, as they can influence public discourse, political information exposure, and children’s media consumption.
Regulators and courts increasingly scrutinize how platform recommendation logic interacts with obligations regarding harmful content, misinformation, and discrimination, making explainability, user controls, and impact assessments salient aspects of AI governance.

Common model types: generative models

Generative models are designed not only to analyze or classify data but to produce new synthetic data or content that are statistically similar to the training data.

Examples include LLMs generating text, GANs creating realistic images, and diffusion models producing high‑quality synthetic media; these models learn joint distributions over inputs and outputs and can sample from those distributions to generate new instances.
Generative models underpin chatbots, code assistants, image and video synthesis tools, and automated report drafting systems, making them powerful productivity aids but also potential sources of misinformation, plagiarism, and authenticity challenges.

Governance concerns around generative AI include provenance and authenticity of digital evidence, consent and personality rights in synthetic depictions, intellectual property implications for training data and outputs, and the risk of large‑scale automated misinformation or harassment.
Frameworks for responsible generative AI emphasize transparency about AI‑generated content, safeguards against deepfakes, mechanisms for consent and redress, and robust controls when generative tools are used within public‑sector or judicial workflows.

Automated decision systems as “decision support” vs “decision replacement”

Literature on AI governance distinguishes systems that support human decision‑making from those that fully automate decisions, with automated decision support systems (ADS) defined as machine‑based systems that can make predictions, recommendations, or decisions influencing real or virtual environments for human‑defined objectives.
These systems may operate with varying levels of autonomy, from providing advisory scores or rankings that humans may accept or reject, to fully automated processes that directly determine outcomes such as benefit approvals, access controls, or content removal.
Best‑practice guidance emphasizes documentation of decision logic, auditability, explainability, monitoring for performance and fairness over time, and keeping humans involved—especially for high‑risk decisions affecting legal rights or access to essential services.

Integrating these concepts for AI governance and law

For AI governance, the conceptual hierarchy—AI as the umbrella, ML as a core technique, neural networks and deep learning as powerful modeling approaches, and generative AI as a frontier for content creation—helps clarify which systems fall under which regulatory obligations.
Automated decision‑making systems built on classification, regression, recommendation, or generative models bring efficiency and scale but also introduce systemic risks of bias, opacity, and error, requiring regulatory safeguards, oversight mechanisms, and judicial readiness to evaluate their legality in concrete disputes.

For legal professionals, mastery of these foundational definitions and model types is essential to engage critically with AI systems in evidence evaluation, administrative review, and constitutional or human‑rights adjudication, rather than treating them as inscrutable “black boxes.”

Think of this chapter as one simple picture: AI is the big umbrella; ML is how it learns; deep learning is powerful learning with neural networks; generative AI is AI that creates; automated decision‑making is AI that decides for or with humans.

Below is an exam‑oriented, easy‑to‑memorize version with analogies and bullet points you can revise quickly.

1. AI, ML, DL, GenAI – the “family tree”

Use this mental picture (very useful in MCQs):

AI
└── ML (subset of AI)
     └── DL (subset of ML, uses neural networks)
            └── Generative AI (subset of DL, creates new content).

  • Artificial Intelligence (AI) – “smart behaviour”

    • Simple definition: Giving a computer the ability to act smart like a human.

    • Examples: Chess‑playing computer, face unlock, Google Maps routes, voice assistants.

  • Machine Learning (ML) – “learning from data”

    • Simple definition: AI that learns patterns from data instead of hard‑coded rules.

    • Example: You give 1,000 labelled emails (spam / not spam); the system learns to classify new emails.

  • Deep Learning (DL) – “brain‑like learning”

    • Simple definition: ML that uses many‑layered neural networks, similar to a mini digital brain.

    • Example: Systems that recognise faces in photos or understand speech (Siri, Google Assistant).

  • Generative AI (GenAI) – “creative machine”

    • Simple definition: DL models that can create new content: text, images, audio, code, etc.

    • Examples: ChatGPT‑style text bots, DALL·E‑style image generators, tools that write code or music.

One‑line memory hook:

AI dreams, ML learns, DL thinks deeper, GenAI creates.

2. Neural networks – the “digital brain”

Neural networks are the core technology behind deep learning and many GenAI systems.

  • Imagine:

    • Input layer = senses (seeing numbers, pixels, words).

    • Hidden layers = brain thinking and transforming the information step by step.

    • Output layer = final decision (cat/not cat, score, label).

  • Key points for exam:

    • Neural networks are inspired by the human brain.

    • Deep learning = neural networks with many hidden layers (“deep” = many layers).

    • Used heavily for images, audio, video, and language because those are complex, unstructured data.

Hook:

Neural network = input → many thinking layers → output. Brain in code.

3. Generative AI – what makes it special

Generative AI is simply: AI that can “invent” or “generate” new content based on patterns it has learned.

  • What it does:

    • Reads huge amounts of data (text, images, audio, etc.).

    • Learns patterns and structure.

    • When you give a prompt, it produces new content that looks similar to what it learned, but is not a copy.

  • Typical examples:

    • Text: answer questions, write emails, draft judgments or summaries.

    • Images: convert a description (“judge in a courtroom”) into a realistic or artistic image.

    • Code: write simple functions or scripts based on instructions.

  • Simple exam line:

    Generative AI = “create new”, not just “classify or predict”.

     4. Automated decision‑making systems (ADM / ADS)

This is the governance‑relevant part: systems that make or support decisions affecting people, using algorithms and data.

  • Definition (in easy terms):

    • Automated decision system = machine‑based system that gives predictions, scores, recommendations or decisions that influence the real world.

    • It may support a human decision (decision support) or replace the human decision (fully automated).

  • Common examples:

    • Credit scoring: approve / reject loans based on ML‑generated risk score.

    • Welfare or tax risk scoring: flag “high‑risk” cases for investigation.

    • Criminal justice risk assessments: risk of re‑offending used for bail or parole decisions in some jurisdictions.

    • Hiring filters: ranking or filtering CVs automatically before human review.

  • Governance keywords to remember:

    • Bias – unfair disadvantage to certain groups due to skewed training data.

    • Transparency / explainability – can we understand why the system gave a particular score or decision?

    • Accountability – who is responsible: system designer, deploying authority, or human decision‑maker?

Hook:

ADM = “algorithm helps or replaces the decision”; law cares about fairness, transparency, and who is accountable.

5. Four common model types – with one clear example each

Think of four “job roles” that models can perform: classify, regress, recommend, generate.

5.1 Classification – “put into buckets”

  • What it does:

    • Predicts a category (spam / not spam, high risk / low risk, yes / no).

  • Easy examples:

    • Spam filter: label emails as spam vs not spam.

    • Creditworthiness: flag loan applicants as “likely to default” or “safe”.

  • Exam keyword:

    Classification = output is a label/category.

    5.2 Regression – “predict a number”

  • What it does:

    • Predicts a continuous value (price, demand, probability, amount).

  • Easy examples:

    • Predict house price from area, location, number of rooms.

    • Forecast monthly sales or revenue amounts.

  • Exam keyword:

    Regression = output is a number/value.

    5.3 Recommendation – “suggest something”

  • What it does:

    • Suggests items/content a user is likely to prefer.

  • Easy examples:

    • Netflix: “Because you watched X, we recommend Y.”

    • Amazon: “Customers who bought this item also bought…”

  • Main techniques:

    • Collaborative filtering (users similar to you).

    • Content‑based filtering (items similar to what you liked).

  • Exam keyword:

    Recommendation = personalized suggestions.

    5.4 Generative – “produce new content”

  • What it does:

    • Generates new text, images, audio, video, or code based on learned patterns.

  • Easy examples:

    • Chatbot that drafts an order or judgment outline from prompts.

    • Image model that draws an illustration from text description.

  • Exam keyword:

    Generative = output is new content, not just a label or number.

    6. Super‑short exam “capsule notes”

You can write or recite these as rapid‑revision points before the AIGP exam:

  1. AI – broad field of making machines act smart like humans (reasoning, learning, perception, language).

  2. ML – subset of AI; learns patterns from data to make predictions or decisions without explicit rules.

  3. DL – subset of ML; uses deep neural networks (many layers) for complex, unstructured data like images and text.

  4. Neural networks – layered nodes inspired by the brain; core of deep learning.

  5. Generative AI – uses deep learning to generate new content (text, images, audio, code) similar to training data.

  6. ADM / ADS – algorithmic systems that give predictions, scores, or decisions affecting people; key issues: bias, transparency, accountability.

  7. Classification – outputs a category (spam/not spam, high/low risk).

  8. Regression – outputs a number (price, probability, demand).

  9. Recommendation – outputs suggestions (movies, products, content).

  10. Generative model – outputs new content (text, images, etc.).


7. How to memorize efficiently 

Two methods will work very well:

Feynman technique: Explain each term (AI, ML, DL, GenAI, ADS, classification, regression, recommendation, generative) in one or two sentences as if you are teaching a junior judge or law student. If you struggle to explain in simple language, revisit that concept and refine your definition.

  • Spaced repetition with flashcards:

    • Create digital/physical flashcards with term on front, simple definition + one real‑world example on back.

    • Revise on Day 1, Day 3, Day 7, then weekly; this schedule greatly improves retention.


Print Page

No comments:

Post a Comment