Generative AI and foundation models are no longer buzzwords limited to the tech elite. They are fast becoming core technologies powering products, services, and decisions across sectors. From writing emails to designing new medicines, generative AI and its foundational underpinnings are shaping how organisations work, solve problems, and stay competitive.
In this pillar page, we cover everything business leaders, product teams, and AI practitioners need to know about generative AI and foundation models — how they work, why they matter, and what the future holds.
What are generative AI and foundation models in artificial intelligence?
Generative AI: Content Creation through Learning
Generative AI refers to machine learning models that can generate content such as text, images, audio, video, and code. Instead of only analysing existing data, these models learn patterns from training data and then produce new outputs based on what they have learned.
For example, a generative AI trained on English news articles can generate a news-style article about a current topic when prompted. The model does not retrieve or remix old content. It creates something new that mimics the structure, tone, and information style of its training data.
Foundation Models: Scalable, Adaptable, General-Purpose
Foundation models are large-scale models pretrained on a broad dataset to develop a general understanding of data. They are called "foundation" because they can be used as the base for many downstream tasks, such as classification, summarisation, translation, and question answering.
Instead of training a new model for every individual task, a foundation model is built once using a massive, diverse dataset. Then it can be fine-tuned or prompted to solve specific problems. Foundation models often serve as the engine behind generative AI systems.
The Shift from Task-Specific to General-Purpose
Earlier generations of AI models were narrow in scope. A sentiment classifier could only detect positive or negative sentiment. A translation model could only translate between specific languages.
Generative AI and foundation models have changed this approach. They are trained to perform many tasks out-of-the-box using patterns learned from large datasets. This shift has led to:
Greater flexibility in deployment
Faster time-to-value
More accessible AI capabilities for non-technical users
Organisations no longer need separate models for every function. A single foundation model can support content creation, summarisation, extraction, and analysis — all through prompt inputs.
How do generative AI models work?
Generative AI models use deep learning architectures to generate new content based on patterns they have learned from large datasets. These models operate by predicting what comes next — whether it’s the next word in a sentence, the next pixel in an image, or the next musical note in a melody. Here's how they work, broken down by architecture and process.
Model Architectures That Power Generation
1. Transformers Transformers are the most widely used architecture in modern generative AI. They rely on self-attention mechanisms, which help the model focus on relevant parts of the input while generating each output token. GPT, Claude, and LLaMA are all transformer-based.
2. Variational Autoencoders (VAEs) VAEs are used for generating structured data like images. They encode input into a latent space, sample from that space, and then decode the sample back into a new image. The randomness of sampling allows for generating novel variations.
3. Generative Adversarial Networks (GANs) GANs consist of two neural networks — a generator and a discriminator. The generator tries to create fake data, while the discriminator attempts to detect fake from real. Over time, the generator learns to produce highly realistic content.
4. Diffusion Models Diffusion models begin with pure noise and gradually refine the image or audio through a denoising process. Models like Stable Diffusion and DALL·E use this method to produce high-quality visuals from text prompts.
Encoder-Decoder Structure and Attention
In many generative models, the encoder processes the input into a dense representation, while the decoder uses this representation to generate new content. Attention mechanisms inside the model assign weights to different input tokens, allowing the model to focus more on relevant inputs during generation.
How Input Tokens Lead to Probabilistic Outputs
The model processes input as tokens — subwords or characters — and converts them into embeddings. Each embedding passes through several neural layers where attention and transformations happen. At each generation step, the model predicts the next token based on a probability distribution.
Instead of always picking the most likely next token (which can lead to repetitive or bland text), the model uses sampling strategies like top-k or nucleus sampling to maintain diversity and creativity in the output.
Examples of Generative AI Models in Use
GPT (OpenAI) – Generates coherent, contextually relevant text for writing, coding, summarising, and more.
Stable Diffusion – Generates images based on natural language prompts.
DALL·E – Creates original images from detailed descriptions.
Claude (Anthropic) – Conversational model trained for safety and nuance.
LLaMA (Meta) – Lightweight and adaptable large language model used for research and production fine-tuning.
These models rely on different mechanisms and training techniques, but the end goal is the same — to produce new, plausible outputs based on learned patterns.
What defines a foundation model in AI?
Foundation models are large-scale, pretrained models that serve as a base for multiple tasks across domains. They are trained on massive datasets and are built to be general-purpose. This makes them highly flexible, scalable, and efficient for real-world deployment.
Core Attributes of Foundation Models
1. Scale Foundation models have billions (or even trillions) of parameters. This scale allows them to understand deep patterns across language, vision, and other modalities.
2. Generality They are not built for a single task. A single foundation model can handle summarisation, translation, classification, reasoning, and content generation with minimal or no retraining.
3. Pretraining on Large Datasets These models are trained on vast and diverse corpora. For example:
Common Crawl – Large-scale web content
C4 (Colossal Clean Crawled Corpus) – Cleaned version of web data
LAION – Open image-text dataset used for training image generators
The size and variety of the dataset help the model learn a wide range of linguistic and semantic patterns.
4. Fine-Tunability and Transferability Once pretrained, foundation models can be fine-tuned on smaller, task-specific datasets or prompted effectively with examples. This makes them efficient for quick deployment and reduces the need for building models from scratch.
How They Differ from Task-Specific Models
Task-specific models require extensive retraining when the task or domain changes. Foundation models, in contrast, are reusable with minimal adaptation.
Examples of Foundation Models
BERT – Trained using masked language modelling. Excellent for understanding tasks like classification and extraction.
GPT-4 – Autoregressive model capable of generating and understanding long text.
PaLM 2 (Google) – Trained on multilingual and multitask data, supports reasoning and coding.
Claude 2 (Anthropic) – Focuses on safety, usability, and reasoning quality in conversation.
These models act as AI backbones, enabling faster innovation, better performance, and broader use-case coverage.
What are the key components of generative AI systems?
Generative AI systems are built from several key components that work together to learn, generate, and refine outputs. Understanding these building blocks helps organisations make informed choices when evaluating or adopting generative AI technologies.
1. Neural Networks
Neural networks are the foundation of all generative AI models. They are composed of layers of interconnected nodes (neurons) that process input data and learn representations. Most modern generative models use deep architectures — particularly transformers — with dozens or hundreds of layers.
These networks learn patterns from data, adjust weights through backpropagation, and generate new outputs by predicting one element at a time.
2. Training Datasets
Data quality directly affects model quality. Generative AI systems require large, diverse, and well-structured datasets. Common sources include:
Text corpora from the web, books, and research papers
Image-text pairs from datasets like LAION or COCO
Code repositories from platforms like GitHub
Speech and audio clips for voice synthesis models
Good data coverage leads to better generalisation and less bias.
3. Latent Space Representation
Models learn to encode inputs into a latent space — a compressed representation where similar inputs are placed near each other. In this space, the model can manipulate ideas, blend styles, or perform arithmetic on features (e.g., “Paris – France + Italy = Rome”).
Latent spaces are crucial for generating novel outputs. VAEs, GANs, and diffusion models all rely on effective latent representations.
4. Tokenization
Tokenization splits raw input into smaller units (tokens), which the model can understand. In text, this could be subwords (e.g., “un+believ+able”). In images, tokenisation might involve dividing a picture into patches or pixels.
Common tokenisation techniques include:
Byte Pair Encoding (BPE)
WordPiece
SentencePiece
Choosing the right tokenizer affects efficiency and accuracy.
5. Prompt Engineering
Prompt engineering is the practice of crafting input instructions that guide model behaviour. A well-designed prompt can help extract specific outputs without needing to retrain the model.
For example:
“Summarise this article in 100 words.”
“Write Python code that calculates Fibonacci numbers.”
Prompt design plays a major role in performance for zero-shot and few-shot tasks.
6. Loss Functions
Loss functions measure how far the model’s output is from the expected answer. They guide learning during training.
Common loss functions:
Cross-entropy loss – Used in language models for next-token prediction.
Mean Squared Error (MSE) – Common in image and audio generation.
KL divergence – Used in VAEs to measure divergence between probability distributions.
Each type is suited to different model objectives and data types.
7. Sampling Strategies
When generating outputs, models use sampling methods to decide which tokens or pixels to select from the predicted probability distribution.
Popular strategies include:
Top-k sampling: Restricts choices to the top k most probable tokens.
Nucleus sampling (top-p): Chooses from the smallest set of tokens whose total probability exceeds p.
Temperature: Controls randomness. Higher values make output more diverse; lower values make it more predictable.
Sampling affects creativity, coherence, and relevance in output generation.
How are foundation models trained?
Training foundation models is a large-scale, multi-phase process involving massive datasets, powerful hardware, and advanced optimisation techniques. The goal is to create models that can generalise across many tasks, even those they haven’t seen before.
Training Phases
1. Pretraining The model is trained on broad, unlabelled data using self-supervised learning. For example, in a language model, the task might be to predict the next word in a sentence or to fill in a missing word (masked language modelling).
Key features of pretraining:
No need for labelled datasets
Focus on pattern recognition and representation learning
Covers many domains (e.g., news, books, conversations, code)
2. Fine-Tuning Once pretrained, the model is fine-tuned on smaller, domain-specific datasets. This helps the model specialise in tasks like sentiment analysis, customer support, or medical summarisation.
Fine-tuning can involve:
Full model adjustment (training all weights)
Parameter-efficient tuning (e.g., LoRA, adapters)
Prompt-tuning (learning soft prompts while freezing the model)
Learning Methods
Self-supervised learning: Uses parts of the input to predict other parts. Common in pretraining of models like BERT and GPT.
Unsupervised learning: Learns from unlabelled data without specific outputs.
Reinforcement Learning from Human Feedback (RLHF): Used in models like ChatGPT to align outputs with human preferences, especially during fine-tuning.
Hardware and Infrastructure Requirements
Training foundation models requires high-performance compute environments:
GPUs (e.g., NVIDIA A100, H100): Common in commercial and research labs.
TPUs (Google’s Tensor Processing Units): Used for large-scale training jobs.
Distributed systems: Spread computation across multiple nodes using techniques like data parallelism and model parallelism.
Other critical components:
High-bandwidth networking: For synchronising gradient updates
Large RAM and storage: To handle checkpoints, datasets, and batch processing
Optimised software frameworks: Such as DeepSpeed, Megatron, or FairScale
Training Challenges
Bias in data: Models can learn stereotypes or incorrect assumptions from training data.
Overfitting: Too much exposure to specific data patterns reduces generalisation.
Alignment: Ensuring model outputs reflect desired values and behaviours.
Compute cost and energy usage: Training models at scale consumes significant electricity and raises environmental concerns.
Data cleanliness: Duplicate, irrelevant, or noisy data reduces model quality.
Training foundation models is resource-intensive but delivers powerful capabilities that can be applied across industries and domains.
What technologies enable generative AI and foundation models?
Generative AI and foundation models rely on a stack of advanced technologies that make large-scale learning, generation, and deployment possible. These core enablers span architecture, infrastructure, algorithms, and data processing.
Transformer Architecture
The transformer architecture is the backbone of most modern generative and foundation models. Introduced in 2017, transformers use a mechanism called self-attention, which allows models to focus on different parts of the input at different times.
Key benefits:
Handles long-range dependencies better than previous models like RNNs
Scales effectively with data and parameters
Supports parallel processing, speeding up training
Transformers form the basis of models like GPT, BERT, LLaMA, PaLM, and Claude.
Scalable Compute Infrastructure
Training and serving large models demand high-performance hardware. Key components include:
GPU clusters: Thousands of GPUs working in parallel across multiple servers
TPU pods: Google’s specialised hardware built for deep learning workloads
High-bandwidth interconnects: For fast communication across distributed systems
Cloud platforms: Azure, AWS, Google Cloud offer scalable compute on demand
The ability to scale compute resources is critical for training trillion-parameter models efficiently.
Optimisation Techniques
Training foundation models at scale requires optimisation methods that manage huge parameter sets and large batches. Common techniques include:
Adam (Adaptive Moment Estimation): Most widely used optimiser for deep learning
LAMB (Layer-wise Adaptive Moments): Designed for large-batch training
Gradient clipping: Prevents exploding gradients during training
Mixed precision training: Uses 16-bit and 32-bit operations to speed up training and reduce memory usage
These techniques improve training speed, stability, and convergence.
Large-Scale Data Curation Pipelines
Data curation is essential to ensure quality and diversity in training. Pipelines involve:
Web crawling: Collecting text, images, and code from public sources
Filtering: Removing spam, low-quality content, toxic language
Deduplication: Ensuring no repetitive data inflates learning
Multilingual balancing: Preventing dominance of any single language or style
The better the data pipeline, the more reliable and general the resulting model.
How do generative AI models differ from traditional AI models?
Generative AI models and traditional AI models differ in purpose, structure, capabilities, and adaptability. Understanding these differences helps businesses choose the right solution for their goals.
1. Architecture and Purpose
Traditional AI models are often built for a specific task. For example, a logistic regression model might be trained to detect fraud, or a decision tree might classify customer churn. These models rely on engineered features and are narrow in scope.
Generative AI models, on the other hand, are designed to learn from patterns and generate new data. They use deep learning architectures like transformers, allowing them to generate text, images, audio, and code from scratch.
2. Flexibility and Adaptability
Traditional models are limited to the task they were trained for. Changing the task means retraining or building a new model.
Generative AI models are more flexible. They can be adapted to new tasks using prompt engineering, fine-tuning, or zero-shot/few-shot learning. For example, a single language model can translate, summarise, write essays, or answer questions — depending on the prompt.
3. Generalisation Capabilities
Generative models are better at generalising across domains, especially when trained on diverse, large-scale data. They can handle edge cases and ambiguous input more gracefully than traditional models.
This generalisation allows them to:
Perform unseen tasks (zero-shot)
Learn from a few examples (few-shot)
Combine knowledge across domains
Traditional models typically fail or degrade in accuracy outside their trained dataset.
4. Multi-Modal Abilities
Generative AI supports multi-modal inputs and outputs — for example, combining text and images (e.g. caption generation, visual question answering).
Traditional models usually handle one type of input (text, image, or numbers), with limited cross-modal capabilities.
5. Emerging Capabilities
Large generative models exhibit emergent behaviour, such as reasoning, analogy, coding, and multi-step planning. These behaviours are not explicitly programmed but arise from training at scale.
Traditional AI lacks this depth. It excels in well-defined environments but struggles with tasks needing context, creativity, or abstraction.
What are current use cases of generative AI and foundation models?
Generative AI and foundation models are driving a wave of innovation across industries. They are not limited to labs or tech companies anymore — real-world applications are growing rapidly in both enterprise and consumer environments.
1. Text Generation
Generative language models can write, summarise, rewrite, translate, and analyse text. This is the most mature application area.
Examples:
ChatGPT, Claude, Gemini: Virtual assistants and conversational agents
Notion AI, GrammarlyGO: Writing support for professionals and students
Jasper, Copy.ai: Marketing copy and email generation
2. Image Synthesis
Text-to-image generation tools allow users to create high-quality visuals from simple prompts.
Examples:
MidJourney: Artistic and creative illustrations
DALL·E: Realistic and stylised images from descriptive text
Stable Diffusion: Open-source model used for image design, product mockups, advertising
3. Code Generation and Assistance
Foundation models trained on programming languages can suggest, complete, and even debug code.
Examples:
GitHub Copilot: AI pair programmer that suggests code in real time
Amazon CodeWhisperer: Context-aware code recommendations
Tabnine: Auto-complete for developers
These tools improve developer productivity and lower the entry barrier for non-programmers.
4. Speech and Audio Generation
AI systems can convert text into lifelike speech and generate new audio clips or voices.
Examples:
VALL-E (Microsoft): High-fidelity voice cloning from just a few seconds of input
Descript’s Overdub: Voice replication for podcast editing and narration
Audio synthesis is useful in customer service, media production, and accessibility tools.
5. Drug Discovery and Scientific Research
Generative models help propose new molecules, simulate protein folding, and optimise chemical interactions.
Applications:
Faster identification of potential drug candidates
Reducing cost and time in pharmaceutical R&D
Discovering novel materials for medical and industrial use
6. Finance and Business Intelligence
AI can generate reports, predict trends, and summarise complex data into actionable insights.
Use cases:
Generating executive summaries from financial statements
Analysing customer feedback at scale
Creating risk models using unstructured data
7. Customer Service Automation
Chatbots powered by foundation models can handle large volumes of queries with context awareness and natural language understanding.
Benefits:
Reduced wait times
Consistent and personalised support
24/7 availability without additional headcount
8. Legal and Document Automation
Legal teams use generative models to review, summarise, and draft legal documents, saving hours of manual effort.
Tasks:
Contract analysis
Compliance checks
Legal research and summarisation
These use cases demonstrate how foundation models can be adapted across domains with high value and low effort.
How do foundation models contribute to AGI research?
Foundation models play a central role in advancing research into Artificial General Intelligence (AGI) — systems that can perform a wide range of intellectual tasks with human-like flexibility. While current models are not AGI, their capabilities and behaviours are bringing us closer to that goal.
1. Scaling Laws and Emergent Abilities
Research shows that as model size, dataset size, and compute power increase, model performance improves predictably. These relationships are known as scaling laws.
At scale, models begin to show emergent abilities:
Multi-step reasoning
Chain-of-thought logic
Math and coding skills
Compositional understanding
These abilities were not seen in smaller models and were not directly programmed, suggesting deeper generalisation.
2. Multi-Modal and Reasoning Benchmarks
Foundation models are now evaluated on multi-modal and reasoning benchmarks — tasks that simulate real-world cognitive flexibility.
Examples:
MMMU, MMLU: Tests for language understanding and logic
VQA, VizWiz: Vision-language reasoning
BIG-Bench: Measures few-shot performance across diverse domains
Models that perform well on these tasks are closer to general-purpose intelligence. Their success pushes the boundary between narrow AI and AGI.
3. Open vs Closed Development Paths
The race toward AGI includes two major strategies:
Open models: LLaMA, BLOOM, Falcon, and Stable Diffusion allow researchers to study and build on top of the technology. These models contribute to transparency, safety, and collective innovation.
Closed models: GPT-4, Claude, Gemini are developed with proprietary methods and data. While high-performing, they limit external validation.
Both paths contribute to AGI progress, but with different priorities — open models foster collaboration, while closed models focus on control and alignment.
4. Foundation Models as AGI Platforms
Current models already perform many tasks that were once considered AI-complete:
Understanding abstract instructions
Solving problems in unfamiliar domains
Learning new tasks from a few examples
Though not fully general, foundation models are now viewed as the starting point for AGI. Many research teams are working on enhancing memory, planning, reasoning, and tool use — capabilities essential for general intelligence.
What are the risks and limitations of generative AI and foundation models?
While generative AI offers remarkable capabilities, it also comes with significant risks and limitations that organisations must address before large-scale adoption. Ignoring these can lead to legal, ethical, operational, and reputational consequences.
1. Hallucination and Factual Inaccuracy
Generative models sometimes produce outputs that sound convincing but are factually incorrect or entirely made up — a phenomenon called hallucination.
Risks:
Misinformation in reports or summaries
Errors in customer communication
Misleading legal, medical, or technical content
Hallucination is a known limitation of current model architectures, especially when operating without access to real-time, verified knowledge sources.
2. Ethical and Social Bias
Models reflect the biases present in their training data. If the data includes stereotypes, offensive language, or skewed representation, the model may reproduce or amplify them.
Consequences:
Discriminatory outcomes in hiring or lending tools
Offensive or harmful content generation
Unequal performance across user groups or languages
Bias mitigation needs to be part of both model training and deployment practices.
3. Environmental Impact
Training large foundation models consumes significant energy. Data centre usage, cooling systems, and power-hungry GPUs lead to high carbon emissions.
Impact:
One large model can consume the equivalent of hundreds of homes' yearly energy
Environmental cost increases with retraining and fine-tuning cycles
Efforts like model optimisation, smaller variants, and green computing are important steps toward reducing this impact.
4. Misuse and Security Risks
Generative models can be misused to create:
Deepfakes
Disinformation campaigns
Spam or phishing content
Identity theft or impersonation
Bad actors can exploit open-access models without strong safeguards. Responsible use and proper access controls are critical.
5. Lack of Interpretability
Foundation models often function as “black boxes.” It's difficult to trace why a model gave a particular answer or to understand how it arrived at its decisions.
Risks include:
Inability to explain decisions to stakeholders
Difficulty identifying errors in reasoning
Legal and compliance challenges, especially in regulated sectors
Interpretability tools are still maturing and often lag behind the complexity of modern models.
What companies and institutions lead development in generative AI?
Several companies and research groups are at the forefront of generative AI and foundation model development. They differ in approach — some prioritise openness and community collaboration, while others focus on commercial scale and control.
1. OpenAI
Flagship models: GPT-3, GPT-3.5, GPT-4, DALL·E
Focus: Text and image generation, alignment, safety research
Collaborations: Deep partnership with Microsoft (Azure OpenAI Service)
Known for: Driving public awareness and adoption of generative AI
OpenAI’s models are widely used across industries for writing, summarising, coding, and automation.
2. Google DeepMind
Models: Gemini, Flamingo, Chinchilla, PaLM, Gopher
Focus: Reasoning, safety, multi-modal capabilities
Known for: Leading benchmarks in performance and multi-task learning
Research-heavy: Combines large-scale engineering with cognitive science
DeepMind is also contributing foundational work to AGI-related tasks and safety.
3. Meta AI
Models: LLaMA, LLaMA 2, LLaMA 3
Focus: Open research, multilingual support, lightweight architectures
Open access: LLaMA models are shared with researchers and developers under controlled terms
Use cases: Research, academia, fine-tuning for production
Meta’s open strategy has made LLaMA one of the most fine-tuned models globally.
4. Anthropic
Models: Claude 1, Claude 2, Claude 3
Focus: AI safety, constitutional AI, alignment
Known for: Safer and more controllable conversational AI
Funding: Backed by major players including Google and Salesforce
Anthropic aims to make AI systems that are honest, helpful, and harmless.
5. Microsoft
Role: Infrastructure provider, investor, and product integrator
Partnership: Deep integration with OpenAI (ChatGPT in Azure, Copilot in Microsoft 365)
Focus: Embedding AI in productivity tools, cloud AI services
Strength: Enterprise AI delivery at global scale
Microsoft is a key enabler of commercial generative AI through tools and platforms.
6. NVIDIA
Role: Hardware provider, model trainer, inference platform leader
Products: DGX systems, NeMo framework, pretrained models
Contributions: Model optimisation, parallel training, GPU acceleration
NVIDIA powers most foundation model training worldwide and supports research and deployment ecosystems.
7. Cohere
Focus: Language understanding, enterprise AI APIs
Products: Embedding models, generation APIs, retrieval-augmented generation
Differentiator: Developer-first approach and enterprise focus
Cohere aims to provide flexible, accessible AI for business applications.
8. Mistral AI
Speciality: Open-weight large language models
Positioning: European response to US-dominated AI landscape
Known for: Lightweight, performant LLMs for developers and enterprises
9. Open-Source Communities
HuggingFace: Hosts thousands of models, datasets, and spaces. Popular for collaboration and model deployment.
EleutherAI: Built GPT-Neo and GPT-J, open alternatives to GPT-3.
BigScience: Responsible for BLOOM, a multilingual open foundation model.
These communities drive transparency, accessibility, and decentralisation in AI.
What is the future of generative AI and foundation models?
Generative AI and foundation models have already reshaped industries, but the next wave of innovation focuses on making them smaller, safer, more efficient, and deeply embedded in domain-specific workflows. Here are the trends shaping the future.
1. Smaller and More Efficient Models
Not every organisation needs a trillion-parameter model. The future will bring:
Distilled models: Smaller versions of large models that retain most of their performance.
Quantised models: Models that use lower-precision computation to run faster and use less memory.
Edge-ready AI: Running powerful models on devices like smartphones and IoT sensors.
These developments will make generative AI more accessible, affordable, and energy-efficient.
2. Better Alignment and Safety Research
The focus is shifting from “can it generate?” to “should it generate?” Future models will:
Be trained with stronger safeguards (e.g. Constitutional AI, RLHF)
Undergo red-teaming and adversarial testing before release
Respond to explicit user intent and ethical boundaries
AI governance and safety will become standard parts of product development.
3. Domain-Specific Integrations
General-purpose models are being adapted to suit specific industries such as:
Healthcare: Clinical documentation, symptom analysis, trial design
Legal: Contract review, case summarisation, compliance tracking
Education: Tutoring, assessment, lesson planning
Finance: Portfolio summaries, regulatory filings, fraud insights
These vertical models will combine general AI power with domain precision.
4. The Open-Source vs Proprietary Race
There’s a growing divide between:
Open-source models: Widely available, community-verified, and modifiable (e.g. LLaMA, Mistral, BLOOM)
Proprietary models: High performance, tightly integrated into ecosystems (e.g. GPT-4, Claude, Gemini)
Both approaches are expanding fast, pushing innovation, but the tension raises questions about control, access, and value.
5. Emerging Regulatory Frameworks
Governments and industry bodies are responding with frameworks for transparency, fairness, and accountability.
Key initiatives:
EU AI Act: Risk-based regulation of AI systems
NIST AI RMF: US-based framework for trustworthy AI
India’s Digital India Act (proposed): Expected to cover AI risk, ethics, and compliance
Organisations will need to align their AI practices with evolving legal standards and documentation requirements.
