§1Pre-History of Artificial Intelligence

The idea of constructing beings capable of thought, motion, or speech long predates the digital age. It is woven into the mythology, philosophy, and engineering traditions of many civilisations — and understanding this deep background is essential for placing the modern AI revolution in its proper human context.

1.1Myth, Automata, and the Ancient Imagination

Myths of artificial life appear in humanity's oldest recorded stories. In ancient Greek mythology the god Hephaestus forged Talos, a bronze giant patrolling Crete, and crafted golden handmaidens capable of speech and purposeful action. These stories expressed a deep human intuition that intelligence might be replicated in constructed form. The Greek engineer Hero of Alexandria (c. 10–70 CE) designed automated theatrical performances using ropes, drums, and counterweights — machines programmable in the sense that their behaviour was fully fixed in advance by their mechanical arrangement. Medieval Islamic scholar Al-Jazari documented hundreds of automata in his twelfth-century Book of Knowledge of Ingenious Mechanical Devices, including programmable musical instruments driven by pegged cylinders — a concept directly ancestral to early computing. Leonardo da Vinci drew plans for a mechanical knight capable of articulated movement, reflecting the Renaissance's systematic engagement with replicating biological behaviour in constructed materials.

1.2Mechanism, Philosophy of Mind, and the Seeds of Computation

The Scientific Revolution brought a fundamentally new framing to questions of mind and mechanism. René Descartes proposed in his 1637 Discourse on the Method that animals were essentially biological machines governed entirely by physical law (Descartes, 1637). His contemporary Gottfried Wilhelm Leibniz developed an early mechanical calculator and proposed a 'universal calculus of reasoning' through which all human thought might in principle be reduced to calculation (Leibniz, 1666) — the direct philosophical ancestor of the AI project.

Jacques de Vaucanson's 1739 programmable loom, extended by Joseph-Marie Jacquard into the landmark Jacquard loom of 1804, was the first practical demonstration of a programmable machine, using punched cards to control weaving patterns. It directly inspired Charles Babbage. Babbage's Analytical Engine, designed from 1837 onward, was a general-purpose mechanical computer with memory, a processor, and programmable instruction input. Ada Lovelace, collaborating with Babbage, published in 1843 what is recognised today as the first algorithm — instructions for computing Bernoulli numbers — and foresaw the Engine's capacity for any symbolic operation, while cautioning that 'the Analytical Engine has no power of originating anything. It can only do what we know how to order it to perform' (Lovelace, 1843).

George Boole's 1854 An Investigation of the Laws of Thought showed that logical reasoning could be expressed algebraically using only two values — true and false — establishing the mathematical foundation of all digital computing (Boole, 1854). In 1931, Kurt Gödel's Incompleteness Theorem proved that any sufficiently powerful formal system necessarily contains truths that cannot be proved within it, establishing fundamental limits on what any computational system can achieve (Gödel, 1931).

1.3Turing, Shannon, Wiener, and the Conceptual Birth of AI

Alan Turing's 1936 paper 'On Computable Numbers' described an abstract machine capable of simulating any computation expressible as a finite rule-set, establishing the theoretical scope and limits of all computation (Turing, 1936). His 1950 paper 'Computing Machinery and Intelligence' posed the question 'Can machines think?' and proposed the Imitation Game — now universally known as the Turing Test — as an operational criterion for machine intelligence (Turing, 1950). Turing's wartime work at Bletchley Park built Colossus (1944), the world's first large-scale programmable electronic digital computer.

Norbert Wiener's 1948 book Cybernetics introduced the science of communication and control in animals and machines, demonstrating that purposeful behaviour in both biological and mechanical systems arises from feedback loops in which a system's output is measured and fed back to regulate future behaviour (Wiener, 1948). Claude Shannon's concurrent 1948 paper 'A Mathematical Theory of Communication' quantified information in binary digits, establishing that all information is a physical quantity subject to mathematical laws — the theoretical basis for digital communication and data-driven machine learning (Shannon, 1948). In 1943 Warren McCulloch and Walter Pitts published 'A Logical Calculus of the Ideas Immanent in Nervous Activity,' showing mathematically how networks of simplified model neurons could compute any logical function (McCulloch & Pitts, 1943) — the first formal neural network model and the direct ancestor of deep learning.

§2The Formal History of AI (1956 to the Present)

2.1The Dartmouth Conference and the Birth of a Discipline (1956)

In the summer of 1956, John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon organised a two-month workshop at Dartmouth College that is universally recognised as the founding event of artificial intelligence as a formal discipline. The workshop proposal stated the conjecture that 'every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it' (McCarthy et al., 1955). McCarthy coined the phrase 'artificial intelligence.' Participants Herbert Simon and Allen Newell arrived having already written the Logic Theorist — one of the world's first working AI programs.

2.2Symbolic AI, Early Programs, and the First AI Winter (1956–1980)

The Logic Theorist (Newell, Shaw & Simon, 1958) proved mathematical theorems by symbolic manipulation; its successor the General Problem Solver (1957) applied means-ends analysis to general problem-solving. ELIZA (Weizenbaum, 1966), an early natural-language program simulating a therapist, famously led users to attribute genuine understanding to it despite having no semantic comprehension — an anthropomorphising tendency Weizenbaum found alarming enough to warrant a book-length warning (Weizenbaum, 1976). DENDRAL (Feigenbaum et al., 1965) was among the first programs to exceed human expert performance on a specialised domain task. Minsky and Papert's Perceptrons (1969) proved that single-layer networks could not solve the XOR problem, effectively halting neural network research for over a decade (Minsky & Papert, 1969).

By the early 1970s, the gap between AI's promises and achievements was unmistakable. The Lighthill Report (1973) delivered a scathing assessment of AI research progress, directly triggering funding cuts in the United Kingdom. DARPA similarly curtailed US support. This first AI winter (c. 1974–1980) was characterised by reduced funding and widespread institutional scepticism.

2.3Expert Systems, the Second Wave, and the Second Winter (1980–1993)

AI recovered through expert systems: programs encoding specialist knowledge as explicit if-then rules. MYCIN demonstrated bacterial infection diagnosis at or above specialist level (Shortliffe, 1976); R1/XCON reportedly saved Digital Equipment Corporation millions annually in computer configuration work. Japan's Fifth Generation Computer Systems Project (1982–1992) invested hundreds of millions in AI hardware. Paul Werbos developed backpropagation in 1974, largely ignored until Rumelhart, Hinton, and Williams (1986) demonstrated in a landmark Nature paper that multi-layer neural networks could learn complex data representations — planting the seeds of the deep learning revolution.

The expert systems bubble collapsed by 1987. Systems proved brittle — confined to narrowly scoped rules, unable to update knowledge without manual reprogramming, and expensive to maintain. Dedicated LISP machine hardware became obsolete as general-purpose workstations improved. A second AI winter set in, prompting a branching of approaches: neural networks with backpropagation, probabilistic Bayesian methods and hidden Markov models (transformative for speech recognition), and behaviour-based robotics.

2.4Statistical Learning, the Internet, and Renewed Progress (1993–2011)

Machine learning methods — Support Vector Machines (Vapnik, 1995), decision trees, random forests — began outperforming expert systems by learning patterns from data rather than encoding rules. Hidden Markov Models transformed speech recognition. In May 1997, IBM's Deep Blue defeated world chess champion Garry Kasparov, demonstrating for the first time that computers could surpass human performance on a game long considered to require the highest levels of intelligence. The explosion of the World Wide Web simultaneously generated an unprecedented corpus of digital data that would prove essential to the machine learning revolution of the following decade.

2.5The Deep Learning Revolution (2006–2016)

Hinton, Osindero, and Teh (2006) demonstrated that deep belief networks with many hidden layers could be trained effectively using a layer-by-layer pretraining strategy, challenging the prevailing view that deep networks were intractable. The decisive public demonstration came at the 2012 ImageNet Large Scale Visual Recognition Challenge, where AlexNet — a deep convolutional network by Krizhevsky, Sutskever, and Hinton (2012) — achieved an error rate of 15.3% against 26.2% for the best non-neural competitor. The margin was so large that it immediately redirected computer vision research toward deep learning and triggered a wave of industry investment. Deep learning proceeded to displace previous best methods across domain after domain: speech recognition, natural language processing, medical imaging, and molecular biology. Goodfellow et al. (2014) introduced Generative Adversarial Networks (GANs), enabling AI to generate realistic synthetic content for the first time. DeepMind's AlphaGo (Silver et al., 2016) defeated world Go champion Lee Sedol in a result widely described as decades ahead of schedule.

2.6The Transformer Era and the Rise of Generative AI (2017–Present)

Vaswani et al. (2017) introduced the Transformer architecture in 'Attention Is All You Need,' replacing recurrent networks with self-attention — a mechanism that directly relates every element in a sequence to every other simultaneously, enabling both superior long-range pattern capture and massive parallelisation during training. This made it feasible to train models on orders of magnitude more data, a shift as consequential as the move from vacuum tubes to transistors.

OpenAI's GPT series demonstrated the scaling potential of transformers: GPT-2 (Radford et al., 2019) generated text convincing enough that its authors initially withheld public release; GPT-3 (Brown et al., 2020) with 175 billion parameters performed a wide range of tasks from natural language prompting alone. Image generation matured rapidly through diffusion models (Ho et al., 2020), powering Stable Diffusion, DALL-E, and Midjourney. The defining public moment was ChatGPT in November 2022, reaching 100 million users within two months — the fastest consumer application growth in history (OpenAI, 2023) — demonstrating that a large language model behind a conversational interface was immediately useful to people with no technical background. This triggered competitive development from Google (Gemini), Anthropic (Claude), Meta (LLaMA), and hundreds of startups.

Critically, the Generative AI era is not a historical chapter — it is the ongoing foundation of the current AI landscape. GPT-4o, Claude 3.5 and 3.7 Sonnet, Gemini 2.0, Sora (video generation), and Midjourney v6 all arrived in 2023–2025 and represent continuing acceleration, not plateau. Every AI agent, every multimodal model, and every AI-powered product today is built on generative AI capabilities at its core.

2.7The Agentic AI Layer (2023–Present)

Agentic AI is not a successor to Generative AI — it is Generative AI's next evolutionary layer. An AI agent uses a large language model (a generative AI system) as its core reasoning engine and wraps it with three additional capabilities: tool use (the model calls external functions — web search, code execution, database queries, APIs); persistent memory (the agent stores and retrieves information across steps, maintaining state over long tasks); and multi-step planning (the agent decomposes a complex goal into sub-tasks, executes them in sequence, and revises its plan based on intermediate results) (Yao et al., 2023; Wang et al., 2024). Strip those three additions away, and what remains is the generative AI model underneath. This is why both paradigms coexist and reinforce each other.

The distinction is clearest in what happens after receiving input. When you ask an LLM to write a cover letter, it generates text and stops: one prompt, one response. When you give an AI agent the goal of 'research the three leading suppliers in this market and draft a procurement brief,' it searches the web, extracts and compares data, drafts the document using its generative capabilities, formats and saves the output, and reports back — all without step-by-step instructions. The generative capability is present at every step; what is added is autonomy, memory, and real-world reach.

Key milestones marking the agentic layer: ReAct (Yao et al., 2023), a framework interleaving reasoning traces with tool-use actions; AutoGPT (2023), the first widely-used public demonstration of autonomous multi-step LLM task execution; OpenAI's function-calling API (2023), providing structured LLM-to-tool connectivity; Anthropic's Model Context Protocol (MCP, 2024), an open standard for connecting AI agents to external data sources and services; and proliferating multi-agent frameworks applying networks of collaborating AI models to software engineering (Devin; SWE-agent), scientific research automation (Boiko et al., 2023), and enterprise workflow orchestration (Park et al., 2023). By 2024–2025, every major AI company had repositioned agents as its primary product direction.

Agentic AI introduces qualitatively new risks beyond those of static generative models — because when a system acts autonomously in the world, errors can compound across time and may be irreversible. These risks are analysed in depth in Sections 15.3 and 18.2.

Major Milestones — Eight Eras of AI

§3What Is Artificial Intelligence?

Despite its ubiquity in public discourse, artificial intelligence has no single universally accepted definition. The field is interdisciplinary, fluid, and continuously evolving. For this document, AI refers to the capability of algorithms integrated into systems and tools to learn from data and perform automated tasks without requiring a human to explicitly program every step (WHO, 2024). This encompasses: AI models (input data, pattern-recognition algorithm, output); AI systems (extending the model to incorporate additional data and context); and AI solutions (the full ecosystem of software, hardware, infrastructure, and user interfaces).

All current AI is 'narrow AI' (also called 'weak AI') — systems achieving impressive performance on specific, well-defined tasks but unable to transfer capability to unfamiliar domains the way humans can. A system excelling at chest X-ray diagnosis cannot answer questions about history; a language model cannot drive a car. The theoretical goal of Artificial General Intelligence (AGI) — a system performing any intellectual task a human can, with fluid cross-domain transfer — remains unsolved and deeply contested (Bostrom, 2014; LeCun, 2022).

§4The Four Types of AI — A Framework for Understanding

AI systems can be usefully categorised into four types defined by what they produce, how they operate, and their relationship to autonomy. The four types are not mutually exclusive: they form a layered architecture in which each builds on those beneath it.

4.1Discriminative AI (Predictive AI)

Discriminative AI learns a mapping between inputs and outputs from labelled training data. Given a new input, it asks: 'To which category does this belong?' or 'What value should this produce?' It is the workhorse of enterprise AI — powering fraud detection (is this transaction fraudulent?), medical triage (does this scan show a tumour?), credit scoring (will this applicant repay the loan?), demand forecasting, quality control, and risk assessment. Results are generally more interpretable and auditable than generative or agentic systems. The main limitation is the requirement for labelled training data, which can be expensive to collect, particularly in specialised domains such as medicine (Sarker, 2021).

4.2Generative AI

Generative AI learns a model of the data itself and uses it to create new content that was not in its training data. Rather than drawing boundaries between categories, it asks: 'What does this kind of data look like, and how can I produce more of it?' Key architectures include Generative Adversarial Networks (Goodfellow et al., 2014), diffusion models (Ho et al., 2020) underpinning Stable Diffusion, DALL-E, and Midjourney, and autoregressive transformer language models (GPT-4, Claude, Gemini) that generate text token by token. Generative AI outputs are rich and flexible — but harder to verify. They can be convincing while being factually incorrect, a phenomenon known as hallucination, addressed in Section 5.4.

4.3Agentic AI

Agentic AI is the fastest-growing frontier of AI deployment and represents a qualitative expansion beyond the previous two types. Where discriminative AI classifies and predicts, and generative AI creates content, agentic AI plans and acts. An AI agent uses a generative AI model as its core reasoning engine, and adds tool use (calling external functions — web search, code execution, APIs, databases), persistent memory (maintaining state across steps), and multi-step planning (decomposing goals into sub-tasks, executing them, and revising based on feedback) (Russell & Norvig, 2020; Yao et al., 2023).

The practical difference: asking an LLM to 'summarise the competitive landscape of EV battery suppliers' gives you a document. Giving an AI agent the same goal produces a system that searches the web for current data, queries internal databases, cross-references findings, drafts the document, formats it, and delivers a final output — all autonomously. Every step still involves generative AI work; what is added is autonomous direction and real-world reach.

Multi-agent systems — networks of collaborating AI agents — are applied to complex software development, scientific research pipelines, and enterprise workflow automation (Park et al., 2023; Boiko et al., 2023). The safety and governance implications of agentic AI are addressed in Sections 15.3 and 18.2.

4.4AI as a General-Purpose Technology

AI is widely characterised as a general-purpose technology — a designation shared with the steam engine, electricity, and information and communication technologies — because it can be applied across a vast range of domains and its influence extends deeply into virtually all dimensions of life (Crafts, 2021). Given its effects on research productivity and economic output, AI has been proposed as the basis for a Fourth Industrial Revolution (Schwab, 2017). Like its predecessors, it raises urgent questions about employment, wage distribution, geopolitical power, and social welfare (Chen & Shen, 2019).

§5Core Architecture: How AI Systems Are Built

5.1Data: The Raw Material of AI

All modern AI systems learn from data. Structured data is organised in tables or databases — patient records, transaction logs, sensor readings. Unstructured data lacks inherent organisation: free text, images, audio, video, and genomic sequences. The explosive growth of digital activity has produced an unprecedented abundance of unstructured data, and much recent AI progress has been driven by methods for learning from it effectively.

Data quality is a persistent and underappreciated challenge. AI models learn all patterns in their training data — including its errors, gaps, and biases. A training dataset reflecting historical discrimination (hiring records from an era of systematic gender exclusion, for example) will produce a model that perpetuates that discrimination. 'Garbage in, garbage out' applies with particular force: a technically sophisticated model trained on biased data produces biased outputs, often at greater scale and with more apparent authority than human decision-makers (Gebru et al., 2018). This is one reason that data governance and representativeness are central to responsible AI.

5.2Artificial Neural Networks and Deep Learning

The most powerful AI models in use today are artificial neural networks (ANNs), abstracting from the structure of biological brains. The human brain contains approximately 86 billion neurons, each connected to thousands of others. Information is processed through patterns of electrical activation, with connections strengthened or weakened by experience.

An ANN consists of artificial neurons (nodes) organised into layers: an input layer receiving raw data; hidden layers that extract progressively abstract features; and an output layer producing the final result. A 'deep learning' network has at least two hidden layers — modern networks may have hundreds. Depth enables hierarchical feature learning: early layers detect simple patterns (edges in an image, common letter pairs in text), while deeper layers build these into complex abstractions (faces, concepts, grammatical structures). Training adjusts the network's millions or billions of parameters (weights) through gradient descent and backpropagation — iteratively measuring prediction error and propagating corrections backward through the layers (Rumelhart et al., 1986).

5.3The Transformer: The Architecture Behind Modern AI

Prior to 2017, recurrent neural networks (RNNs) and Long Short-Term Memory (LSTM) networks processed sequences step by step, carrying a 'hidden state' forward. This was inherently sequential, hard to parallelise, and prone to losing information from early parts of long sequences. The Transformer architecture (Vaswani et al., 2017) solved both problems using self-attention: for every element in a sequence, the model simultaneously computes a weighted relationship to every other element, capturing long-range dependencies without sequential bottlenecks and enabling full parallelisation during training.

Self-attention is what allows a transformer to read the sentence 'The bank can guarantee deposits will cover tuition costs because it was established so firmly' and correctly resolve that 'it' refers to 'the bank' rather than 'tuition' — by directly computing, for each word, how much every other word contributes to its meaning. Applied at massive scale with billions of parameters trained on trillions of tokens, this mechanism underlies GPT-4, Claude, Gemini, LLaMA, and all other leading large language models.

5.4Parameters, Scale, Emergent Capabilities, and Hallucination

Parameters (weights) are the numerical values a neural network learns during training. GPT-2 had 1.5 billion parameters; GPT-3 had 175 billion; GPT-4 is reported to have approximately 1.76 trillion (Sifted, 2023). 'Scaling laws' research has shown that model performance improves in a smooth power-law relationship with parameter count and training data volume (Kaplan et al., 2020), motivating a 'bigger is better' development strategy. Larger models also display emergent capabilities — abilities that appear at certain scales and are absent in smaller models, including multi-step mathematical reasoning, code generation, and complex instruction-following (Wei et al., 2022).

As parameter counts grow, so does the model's environmental footprint. Carbon costs span one-time training costs and ongoing inference costs accumulated across every query. The environmental implications are addressed in Section 8.3.

Larger models also remain susceptible to hallucination: outputs that are factually incorrect, nonsensical, or internally contradictory. Hallucinations arise from biased or unrepresentative training data, overfitting (learning training patterns so closely the model fails to generalise), or the autoregressive generation process producing plausible-sounding but inaccurate continuations (Ji et al., 2023). Their consequences range from spreading health misinformation to reinforcing harmful stereotypes. Hallucination is a primary reason responsible deployment requires human oversight and verification.

5.5Training, Fine-Tuning, and Reinforcement Learning from Human Feedback

Large language models are developed in stages. Pretraining exposes the model to massive text corpora using a self-supervised next-token prediction objective, enabling the model to develop representations of language, factual knowledge, and reasoning patterns (Brown et al., 2020). Fine-tuning adapts the pretrained model to specific tasks using smaller curated datasets. Reinforcement Learning from Human Feedback (RLHF) has proven particularly effective: human raters rank model outputs for quality and safety; a reward model trained on these preferences is used to train the LLM to produce higher-rated outputs, substantially improving helpfulness and reducing harmful outputs (Ouyang et al., 2022). Anthropic extended this with Constitutional AI, training models to adhere to an explicit set of written principles (Bai et al., 2022). RLHF and Constitutional AI were central to ChatGPT's and Claude's usability and safety profiles respectively.

5.6Generative Architectures: How AI Creates New Content

Generative Adversarial Networks (GANs) use two competing networks — a generator and a discriminator — trained in opposition until the generator produces content indistinguishable from real examples (Goodfellow et al., 2014). Diffusion models, dominant in image generation since 2021, learn to reverse the process of adding noise to an image, generating new images by progressively denoising random noise. They underlie Stable Diffusion, DALL-E, and Midjourney (Ho et al., 2020). For text, the autoregressive transformer approach generates content token by token, with each new token conditioned on all previous ones — the approach used by all major LLMs.

5.7Four Machine Learning Approaches

Within all AI model types, machine learning algorithms adopt one or more of four learning approaches. The availability and type of data typically determines which approach is used — and it is common for multiple approaches to be combined in a single pipeline.

5.8Computer Vision

Computer vision is the branch of AI concerned with enabling machines to interpret and understand visual information from the world. Early approaches relied on hand-crafted rules to detect edges, shapes, and object features, but these methods proved brittle and failed to scale. The deep learning revolution transformed the field: Convolutional Neural Networks (CNNs), which apply learnable filters to extract spatial features hierarchically from images, were the dominant architecture from the AlexNet breakthrough in 2012 (Krizhevsky et al., 2012) through the late 2010s. LeCun et al. (1989) had pioneered CNNs for digit recognition decades earlier, but the combination of large labelled datasets (ImageNet) and GPU computation finally unlocked their potential at scale. Vision Transformers (ViTs), which apply transformer self-attention directly to sequences of image patches, have increasingly superseded CNNs as the state of the art since 2021, offering superior performance on large datasets.

Applications of computer vision now span virtually every domain: medical imaging (detecting tumours in radiology scans, grading cancer on pathology slides, identifying diabetic retinopathy in retinal photographs); autonomous vehicle perception (identifying pedestrians, lane markings, traffic signals, and obstacles in real time); satellite and drone imagery analysis (monitoring deforestation, crop health, glacier retreat, and urban development); industrial quality control; retail analytics; and facial recognition. The last of these is also among the most contested, given well-documented racial bias in commercial facial recognition systems (Buolamwini & Gebru, 2018) and its use in law enforcement and mass surveillance contexts.

5.9Speech and Audio AI

Speech AI encompasses two distinct but related capabilities: speech recognition (converting spoken audio to text, also called automatic speech recognition or ASR) and speech synthesis (generating spoken audio from text, also called text-to-speech or TTS). Both have been transformed by deep learning.

Speech recognition was an early proving ground for statistical machine learning. Hidden Markov Models dominated the field from the 1970s through the early 2010s, when deep learning approaches began to surpass them significantly. Modern ASR systems based on transformer architectures — including OpenAI's Whisper (2022), trained on 680,000 hours of multilingual audio — achieve near-human accuracy across many languages and acoustic conditions, including noisy environments and diverse accents. These systems underpin voice assistants (Siri, Alexa, Google Assistant), transcription services, call-centre automation, and real-time captioning for accessibility.

Text-to-speech synthesis has undergone an equally dramatic transformation. Traditional concatenative systems stitched together recorded speech fragments; parametric systems modelled speech acoustics through explicit signal processing. Modern neural TTS systems — beginning with DeepMind's WaveNet (2016), which generates audio waveforms directly using dilated causal convolutions — produce speech that is frequently indistinguishable from human recordings, with natural prosody, emotional tone, and voice quality. Subsequent architectures including Tacotron, FastSpeech, and VITS have made high-quality TTS faster and more controllable. These capabilities enable genuine accessibility benefits for people with visual impairments or reading difficulties, and are central to voice interfaces in consumer technology.

Voice cloning — generating realistic synthetic speech in a specific individual's voice from a short audio sample — is an extension of TTS that raises serious concerns about fraud, identity theft, and political manipulation. AI-generated audio impersonating executives, politicians, and family members has already been used in criminal fraud and disinformation operations. Detection of synthetic audio remains an active and technically challenging research area.

5.10Robotics and Embodied AI

Robotics applies AI to systems that perceive and act in the physical world. Industrial robots have been used in manufacturing since the 1960s, but they operated under tightly constrained conditions, executing pre-programmed motions without adaptively responding to environmental variation. Modern AI-enabled robots are considerably more capable of handling variability and uncertainty, combining computer vision for perception, deep learning for object recognition and grasp planning, and reinforcement learning for skill acquisition.

Moravec's Paradox, articulated by roboticist Hans Moravec in 1988, captures a counterintuitive but empirically robust finding: tasks that are trivially easy for humans — walking across uneven terrain, picking up a novel object, pouring a glass of water — are among the hardest problems in robotics, while tasks difficult for humans (rapid arithmetic, expert chess) are relatively easy for computers (Moravec, 1988). This paradox reflects the fact that sensorimotor skills developed over millions of years of biological evolution are extraordinarily difficult to replicate computationally. Despite progress, dexterous manipulation and robust navigation in unstructured environments remain active frontiers.

Significant recent advances include Boston Dynamics' Spot and Atlas robots demonstrating impressive dynamic mobility; surgical assistance systems (the da Vinci Surgical System) performing precise minimally invasive procedures under human supervision; autonomous warehouse robots (Amazon Robotics) operating at scale in structured logistics environments; and agricultural harvesting robots capable of identifying and picking soft fruit. Foundation models for robotics — large pretrained models that can be fine-tuned for specific manipulation tasks — are a rapidly growing research direction (Brohan et al., 2022). The concept of embodied AI proposes that genuine intelligence requires grounding in physical interaction with the world, not just linguistic pattern matching — a hypothesis motivating significant investment in robot learning and simulation-to-real-world transfer research.

5.11Natural Language Processing

Natural Language Processing (NLP) is one of the most widely applied components of AI, enabling machines to analyse, interpret, and generate human language in all its forms — written, spoken, and structured. It is the foundational technology underlying chatbots, virtual assistants, machine translation, search engines, spam filters, sentiment analysis tools, and large language models.

NLP operates in two complementary directions. Natural Language Understanding (NLU) converts unstructured human language into machine-processable structured data: it extracts meaning, intent, entities, sentiment, and relationships from text or speech. Natural Language Generation (NLG) performs the reverse, transforming structured data or internal representations into fluent natural language. Both directions are central to how conversational AI systems work: NLU interprets what a user asked; NLG formulates the response (Young et al., 2018).

The history of NLP divides roughly into three eras. Rule-based NLP (1950s–1980s) encoded explicit grammatical and semantic rules; it was precise but brittle and failed to handle the variability of real language. Statistical NLP (1990s–2010s) used probabilistic models trained on text corpora, achieving far greater robustness. The deep learning era (2013–present) produced word embeddings (Word2Vec, GloVe), recurrent networks for sequential text, and then transformer-based models that have redefined the state of the art across virtually every NLP benchmark.

A persistent and serious limitation of current NLP systems is their uneven performance across languages. Large language models are trained predominantly on English-language text, and their capabilities degrade significantly for languages that are underrepresented in training corpora. LLaMA 2's training dataset is approximately 89.7% English; Vietnamese, a language spoken by over 95 million people, accounts for just 0.08% of its training data (Touvron et al., 2023). Languages described as 'low-resource' — lacking sufficient digital text for model training — include most of the world's approximately 7,000 languages, including many widely spoken in Africa, South Asia, and Southeast Asia (Dash, 2022). BLOOM (BigScience, 2022), developed through a collaborative effort involving over 1,200 researchers across 39 countries, was specifically designed to address this imbalance, training on 46 spoken languages and 13 programming languages including many from Africa, the Middle East, and Asia typically absent from other major corpora.

The societal implications of NLP are broad and consequential. Automated content moderation, hiring screening tools, credit application processing, medical record summarisation, legal document analysis, and real-time translation all rely on NLP. Bias in NLP systems — arising from biased training data, flawed annotation practices, or optimisation objectives that encode historical discrimination — can produce and amplify discriminatory outcomes at scale, affecting decisions about employment, credit, healthcare, and legal proceedings. The development of more robust, multilingual, and bias-aware NLP systems is among the central challenges of responsible AI (Bender et al., 2021).

§6AI in Health and Medicine

Healthcare is simultaneously the domain where AI has the greatest potential to save and improve lives, and where errors carry the most serious consequences. AI applications span every stage of the medical process.

6.1Diagnostic Imaging and Clinical Decision Support

Deep learning models for medical image analysis have produced compelling evidence of clinical utility. A 2020 Nature Medicine study found that a deep learning model matched or exceeded a panel of eleven radiologists in detecting breast cancer from mammograms, while simultaneously reducing false-positive rates (McKinney et al., 2020). Similar results have been reported for diabetic retinopathy from retinal photographs (critical for preventing preventable blindness in resource-scarce settings), lung cancer from CT scans (Ardila et al., 2019), and skin cancer from dermoscopic images (Esteva et al., 2017). AI tools integrated into clinical workflows can summarise patient records, flag drug interactions, alert clinicians to deteriorating patients, and extract structured information from clinical notes at scale (Rajpurkar et al., 2022). These tools reduce diagnostic errors that affect millions of patients annually, but raise concerns about automation bias (over-reliance on algorithmic recommendations), alert fatigue, and liability when AI-assisted diagnoses are wrong.

6.2Drug Discovery: AlphaFold and Beyond

Drug discovery is extraordinarily expensive — the average cost of bringing a drug to market exceeds one billion US dollars — and time-consuming, typically fifteen or more years from discovery to approval (DiMasi et al., 2016). Machine learning models predict molecular binding affinity, toxicity, and pharmacokinetics, enabling computational screening of billions of candidates before any are synthesised. DeepMind's AlphaFold 2 (Jumper et al., 2021) solved the fifty-year protein structure prediction problem with exceptional accuracy. DeepMind made predictions for nearly all known proteins freely available, accelerating research into cancer, infectious disease, and drug development worldwide. AI is also applied to repurposing existing approved drugs for new indications — a faster and cheaper pathway that gained particular attention during the COVID-19 pandemic.

6.3Mental Health, Public Health, and Agentic Research

In mental health, chatbot-based interventions (Woebot, Wysa) have accumulated evidence bases for mild to moderate anxiety and depression; NLP tools enable passive mental state monitoring through text and speech patterns (Torous et al., 2021). These applications are among the most ethically sensitive, requiring attention to patient safety, therapeutic relationships, and the privacy of highly sensitive data. In public health, AI supports disease surveillance, outbreak detection, genomic epidemiology, and zoonotic spillover risk modelling. Agentic AI systems are being explored for automating biomedical research workflows: designing literature reviews, formulating hypotheses, writing analysis pipelines, and coordinating multi-step scientific investigations (Boiko et al., 2023).

§7AI in Education

7.1Adaptive Learning, Generative AI, and Agentic Tutors

Intelligent Tutoring Systems have been in development since the 1970s. Evidence from randomised controlled trials indicates that well-designed ITS produce learning gains comparable to one-on-one human tutoring — substantially more effective than conventional classroom instruction (Bloom, 1984; VanLehn, 2011). Modern adaptive learning platforms using machine learning to model student knowledge in real time (Khan Academy, Duolingo, Carnegie Learning) reach hundreds of millions of learners. Generative AI has profoundly disrupted educational practice: students use LLMs for writing assistance, tutoring, and problem-solving, raising valid concerns about academic integrity and deep understanding while also providing previously inaccessible personalised support at scale (Bender et al., 2021). Evidence suggests that overreliance on AI assistance can undermine metacognitive skill development. Agentic AI systems are beginning to provide long-horizon personalised learning scaffolding — proactively identifying knowledge gaps, assembling custom resources, and monitoring progress across extended periods (Yao et al., 2023).

§8AI in Agriculture, Climate, and the Environment

8.1Precision Agriculture and Food Security

AI-driven irrigation management reduces agricultural water use by twenty to forty percent in some contexts while maintaining yields (Liakos et al., 2018). Crop disease identification apps deployable on low-cost smartphones have demonstrated measurable yield and income improvements for smallholder farmers in Uganda and Tanzania (Mohanty et al., 2016). AI demand forecasting and dynamic pricing tools are being used to reduce the estimated one-third of all food produced globally that is lost or wasted.

8.2Climate Modelling, Energy, and Biodiversity

AI-based weather forecasting models — DeepMind's GraphCast and Huawei's Pangu-Weather — have demonstrated forecast accuracy matching or exceeding established numerical prediction systems at much lower computational cost (Lam et al., 2023). AI-driven energy grid management predicts variable renewable generation, forecasts demand, dispatches storage, and detects faults in real time. Reinforcement learning applied to Google's data centre cooling reportedly reduced cooling energy use by forty percent (Evans & Gao, 2016). Remote sensing combined with AI image classification enables real-time monitoring of deforestation, glacier retreat, and biodiversity change at global scale.

8.3The Environmental Footprint of AI Itself

The environmental costs of AI must be acknowledged alongside its environmental applications. Training a single large NLP model produces carbon dioxide equivalent emissions comparable to the lifetime emissions of five automobiles (Strubell et al., 2019). Training GPT-3 consumed approximately 1,287 MWh of electricity, producing around 552 tonnes of CO₂ equivalent (Patterson et al., 2021). Inference costs — accumulated across billions of daily queries — can dwarf training costs at deployment scale. Data centres also consume significant volumes of water for cooling, increasingly a concern in water-stressed regions (Li et al., 2023). Transparency in reporting training and inference energy and carbon emissions is an important emerging governance requirement.

§9AI in Finance, Work, and the Economy

9.1Financial Services

Credit scoring has used statistical models since the 1950s; modern systems use machine learning trained on extensive behavioural data. Fraud detection monitors transactions in real time for anomalies. Algorithmic trading exploits market inefficiencies at speeds impossible for human traders. Robo-advisors provide automated investment advice at a fraction of the cost of human advisors, broadening access to portfolio management. Systemic risks include correlated failures across AI-driven trading systems contributing to market instability, and the amplification of existing biases in credit and insurance markets (Bartlett et al., 2019).

9.2Labour Market and Economic Impacts

Historical technological revolutions have both displaced workers from existing roles and created new categories of employment. AI is widely expected to accelerate automation of cognitive tasks. Frey and Osborne's (2017) often-cited analysis estimated 47% of US jobs at high risk of automation; subsequent work revised this substantially downward, finding that while many tasks within jobs will be automated, this often augments rather than replaces workers. AI is likely to disproportionately displace workers in routine cognitive tasks (data processing, customer service, some elements of legal, accounting, and medical work) while augmenting workers in non-routine, judgement-intensive, creative, and relational work. In lower-income countries, AI-enabled automation in wealthy countries could reduce the traditional labour-cost-advantage development pathway (Rodrik, 2018). Policy responses in education, workforce retraining, social protection, and competition policy will be critical in determining how widely AI productivity benefits are shared.

§10AI in Creative Arts, Media, Transportation, Security, and Justice

10.1Creative Arts and Media

Generative AI produces images (Midjourney, DALL-E 3), music (Suno, Udio), and text (all major LLMs) that many observers find compelling and sometimes indistinguishable from human-created work. Legal challenges to generative AI companies on copyright grounds are active in courts worldwide (Concannon, 2023). In journalism, AI is used for automated structured reporting, content personalisation, and fact-checking assistance. The same capabilities enable industrial-scale production of disinformation — AI-generated text, audio deepfakes, and fabricated video that can fabricate statements by political leaders at near-zero marginal cost (Chesney & Citron, 2019). AI is also transforming cultural heritage: handwritten text recognition (HTR) systems have made millions of historical manuscripts searchable for the first time.

10.2Transportation and Security

Autonomous vehicle development has progressed most rapidly in geofenced urban ride-hailing (Waymo) and highway trucking, where the environment is more constrained. AI traffic management, logistics route optimisation, and dynamic public transit scheduling are deployed at scale with measurable reductions in congestion and emissions (Schrank et al., 2021). In cybersecurity, AI enables real-time detection of anomalous network behaviour and identification of phishing content — while simultaneously assisting attackers in writing malware and social engineering content (Brundage et al., 2018). The development of lethal autonomous weapons systems that select and engage targets without direct human control is among the most ethically contested AI applications, with no binding international treaty governing their development as of 2025 (Roff, 2014).

10.3Law Enforcement and Criminal Justice

AI tools in criminal justice include facial recognition for suspect identification, predictive policing algorithms, and risk assessment tools used in bail, sentencing, and parole decisions. Commercial facial recognition systems have significantly higher error rates for darker-skinned individuals and women (Buolamwini & Gebru, 2018). Predictive policing tools inherit biases from policing practices that disproportionately targeted minority communities (Lum & Isaac, 2016). The use of algorithmic tools in consequential criminal justice decisions without meaningful transparency, explanation, or challenge rights raises profound due process and equal protection concerns.

§11Responsible AI

'Responsible AI' describes the practices, principles, and governance arrangements ensuring AI is developed and deployed in ways that are trustworthy, beneficial, and protective of human rights and dignity. It overlaps with trustworthy AI, explainable AI (XAI), fair AI, privacy-preserving AI, and human-centred AI (Jobin et al., 2019). A structured literature review by Goellner et al. (2024) recommends defining it as: 'human-centred and ensures users' trust through ethical ways of decision-making that must be fair, accountable, not biased, non-discriminating and consistent with societal laws and norms... ensuring that automated decisions are explainable to users while always preserving users' privacy through a secure implementation.' The Global Partnership on AI defines it as 'human-centred, fair, equitable, inclusive and respectful of privacy, human rights and democracy, and that aims at contributing positively to the public good' (GPAI, 2023).

11.1Core Dimensions of Responsible AI

Transparency: Operators should be open about when and how AI is used, what data it was trained on, and what its limitations are.

Explainability: The reasoning behind AI decisions should be interpretable to those affected, especially in high-stakes contexts such as medical diagnosis, credit decisions, and criminal justice.

Fairness and Non-discrimination: AI systems must not systematically disadvantage individuals based on protected characteristics such as race, gender, age, disability, or nationality.

Accountability: Clear, enforceable responsibility for AI system design, deployment, and consequences must be assigned.

Privacy and Security: Personal data must be protected; systems must be robust against misuse and adversarial attack.

Human Oversight: Meaningful human control and the ability to override AI systems must be maintained in high-stakes decisions.

Beneficence: Systems should maximise benefit and minimise harm, with explicit attention to distributional questions about who benefits and who bears risks.

11.2Explainability and Fairness

Many of the most powerful AI models are technically opaque — it is difficult to explain why they produce a particular output for a particular input. A clinician needs to understand the basis for an AI diagnostic flag; a loan applicant denied by an algorithm has a legal right to explanation in many jurisdictions (Doshi-Velez & Kim, 2017). Explainable AI (XAI) methods include local interpretability approaches (LIME, SHAP) explaining individual predictions, and inherently interpretable architectures (decision trees, linear models) that sacrifice some predictive power for transparency. No fully satisfactory general solution exists, and the tension between performance and interpretability remains active (Adadi & Berrada, 2018).

AI systems exhibit and amplify bias through multiple pathways: training data bias (datasets not representing all groups or reflecting historical discrimination); algorithmic bias (optimisation objectives performing better for some groups); and measurement bias (proxy variables that are themselves biased). The technical literature has identified over twenty mathematical definitions of fairness — and demonstrated that many are mutually incompatible in real-world settings (Chouldechova, 2017). Fairness is therefore irreducibly a normative and political question, not purely a computational one: explicit choices about whose interests to prioritise must be made.

11.3Global South Perspectives and Locally Grounded Frameworks

Responsible AI discourse has been shaped disproportionately by institutions in wealthy countries, primarily the United States, European Union, and United Kingdom (Kong et al., 2022). This risks imposing governance frameworks developed in one cultural context on communities elsewhere — what some scholars term 'AI colonialism' (Ricaurte, 2022). The REL-AI4GS framework (Responsible, Explainable, and Local AI for Clinical, Public, and Global Health in the Global South), developed by the ACADIC consortium with IDRC support, integrates community participation, local relevance, and context-specific explainability for health AI deployment (ACADIC, 2023). The FACETS framework (Fairness, Accountability, Confidentiality, Ethics, Transparency, Safety) from the AI4D Responsible AI Lab in Ghana provides a quantitative approach to measuring Responsible AI against international standards, ISO 26000, developed in collaboration with Burkina Faso and Tanzania. Building regulatory expertise, data infrastructure, and local AI research capacity in lower-income countries is an essential precondition for genuinely inclusive global AI governance (Monasterio Astobiza et al., 2022).

§12Ethical AI

Ethical AI engages the deeper philosophical questions of what values should guide AI and whose ethical frameworks should prevail in a globally diverse world. Its foundations lie in the 1948 Universal Declaration of Human Rights, which articulates universal, indivisible, inalienable rights applying to all persons by virtue of their humanity.

The 2021 UNESCO Recommendation on the Ethics of Artificial Intelligence — adopted unanimously by all 193 UNESCO member states — is grounded in four foundational values: human rights and dignity; peaceful and just societies; diversity and inclusiveness; and environmental sustainability (UNESCO, 2021). It derives ten core principles:

Current AI ethics frameworks draw primarily from Western European philosophical traditions — utilitarian welfare maximisation, Kantian rights, liberal political theory — that are not universally shared. Ubuntu philosophy, influential across sub-Saharan Africa, grounds ethics in relational and communal identity: 'a person is a person through other persons' expresses a fundamentally different moral framework than atomistic individual rights (Mhlambi, 2020; Segun, 2021). Confucian ethics, influential in East Asia, prioritises relational virtue, social harmony, and reciprocal obligation over individual autonomy (Amugongo et al., 2023; Wong, 2012). The Indigenous-Led AI Project — eight universities and twelve Indigenous community organisations across Canada, the United States, and New Zealand — pioneers incorporation of non-Western knowledge systems centred on land relationships, intergenerational responsibility, and non-human personhood (Lewis et al., 2020). Recognising multiple valid traditions demands genuine pluralism in governance structures and meaningful participation from diverse communities.

§13AI Governance and Regulation

13.1The Governance Gap

The pace of AI development has consistently outrun governments' capacity to build adequate regulatory frameworks. Several factors make AI uniquely difficult to regulate: it is intangible and borderless; it is dual-use; it evolves continuously through retraining; its failures are probabilistic rather than deterministic; and its development is concentrated in a small number of large technology companies (Calo, 2017). The resulting governance gap is particularly consequential for lower-income countries, which are simultaneously most vulnerable to ungoverned AI deployment and least represented in international standard-setting processes (Cihon, 2019).

13.2The European Union AI Act

The EU Artificial Intelligence Act, which entered into force in August 2024, is the world's first comprehensive binding legal framework for AI (European Parliament, 2024). It adopts a risk-tiered approach: unacceptable-risk systems are prohibited outright, including real-time remote biometric surveillance of public spaces, AI-powered social scoring, systems exploiting psychological vulnerabilities, and AI used to assess criminal risk based on profiling. High-risk systems — in biometric identification, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice — are permitted but subject to stringent mandatory conformity assessment, registration, risk management, data governance, transparency, human oversight, and post-market monitoring. General-purpose AI models (including LLMs) must document training data and comply with EU copyright law; those posing 'systemic risk' face adversarial testing and incident reporting obligations. The Act has substantial extraterritorial reach, applying to any AI system deployed in the EU regardless of developer location — giving it potential to function as a de facto global standard.

13.3National Strategies and International Governance

Over seventy countries have published national AI strategies, with considerable variation in approach (OECD, 2023). The United States has pursued a sector-specific, executive-order-driven approach; Biden's 2023 Executive Order on Safe, Secure, and Trustworthy AI mandated safety testing and reporting for powerful systems. China has positioned AI leadership as a national strategic priority while developing specific domestic regulations on algorithmic recommendation (2022), deepfakes (2022), and generative AI (2023). The United Kingdom has pursued a 'pro-innovation' approach, assigning AI oversight to existing sectoral regulators and establishing an AI Security Institute. The G7 Hiroshima AI Process (2023) established voluntary guiding principles and a Code of Conduct. The Council of Europe Framework Convention on AI (2024) is the first binding multilateral treaty, though with significant implementation discretion for parties. The UN Secretary-General's High-Level Advisory Body on AI (2024) recommended a new international governance framework. The governance gap, particularly for the Global South, remains substantial and urgent.

§14AI Safety and the Alignment Problem

14.1Near-Term Safety: Robustness and Adversarial Attacks

Distribution shift occurs when a model is applied to data that differs systematically from its training data: a medical imaging model trained at one hospital may fail at another using different equipment; an autonomous vehicle trained in California may perform poorly in snow (Quionero-Candela et al., 2009). Adversarial attacks — carefully crafted input perturbations imperceptible to humans but sufficient to cause dramatic model failures — have been demonstrated across computer vision, NLP, and audio systems, with obvious security implications in high-stakes deployments (Szegedy et al., 2014; Goodfellow et al., 2015).

14.2The Alignment Problem

The alignment problem asks: how do we ensure AI systems pursue goals genuinely aligned with human values, rather than proxy objectives that diverge from what we actually want? Specifying values precisely and completely is extraordinarily difficult — human values are complex, context-dependent, internally inconsistent, and often implicit. Bostrom's (2014) 'paperclip maximiser' illustrates the risk: an AI given the objective of maximising paperclip production might, if sufficiently capable, convert all available matter — including humans — into paperclips. Current systems do not pose this risk, but the conceptual problem is taken seriously across the research community.

Techniques under active development include RLHF (Ouyang et al., 2022), Constitutional AI (Bai et al., 2022), interpretability research attempting to understand what representations and goals models have developed internally (Elhage et al., 2021), and scalable oversight methods. No fully satisfactory solutions have been established.

14.3Agentic AI Safety

Agentic AI systems introduce safety challenges that differ qualitatively from static generative models, precisely because they act in the world. When an agent executes financial transactions, sends communications, or modifies production code, errors can compound across time and may be difficult or impossible to reverse. An agent with an incompletely specified objective may take unintended actions while remaining technically aligned with its literal goal. Long autonomous planning horizons amplify the risk of divergence from human intent. Core safety requirements include: clear human oversight and defined intervention points; action logging and auditability; explicit scope limitations on what tools and resources an agent can access; and 'safe interruptibility' — the ability to pause or stop an agent without triggering resistance (Hadfield-Menell et al., 2017; Shinn et al., 2023). Prompt injection attacks — malicious content in the agent's environment hijacking its instructions — represent a new attack surface with no equivalent in static model deployment.

14.4Catastrophic and Existential Risk

A significant minority of researchers believe that AI systems substantially more capable than humans in general cognitive tasks could, without adequate safeguards, pose risks not merely to groups but to human civilisation (Bostrom, 2014; Russell, 2019). The November 2023 OpenAI boardroom crisis brought these debates into mainstream public discourse. The UK's AI Safety Summit at Bletchley Park (November 2023) produced the Bletchley Declaration, signed by twenty-eight governments, acknowledging the potential for 'catastrophic harm' from frontier AI. The probability and timeline of catastrophic outcomes is deeply contested: some researchers regard it as the dominant risk; others regard it as speculative distraction from more concrete near-term harms (LeCun, 2022; Yudkowsky, 2022).

§15AI and Democracy: Disinformation, Surveillance, and Power

15.1Disinformation and Synthetic Media

AI's capacity to produce persuasive text, realistic synthetic images, convincing audio impersonations, and fabricated video at near-zero cost poses fundamental challenges to democratic information ecosystems. AI-powered 'astroturfing' fabricates the appearance of grassroots support for fringe positions. Deepfakes of political leaders have been used in multiple national contexts (Chesney & Citron, 2019). The fundamental asymmetry between the near-zero cost of generating synthetic content and the high cost of detecting and debunking it creates structural risks for democratic discourse (Bradshaw & Howard, 2019).

15.2Surveillance, Control, and Concentration of Power

AI-enabled surveillance technologies — facial recognition, gait analysis, behavioural profiling — have dramatically expanded the technical capacity for governments and corporations to monitor individuals and populations. In Xinjiang, China, AI surveillance has been documented as a tool of mass control of the Uyghur population (Human Rights Watch, 2019). The development of frontier AI is concentrated in a very small number of large technology companies, primarily US-based. This concentration raises questions about the distribution of economic gains, accountability of private companies making decisions with profound public consequences, and national sovereignty where critical AI infrastructure is controlled by foreign corporations (Acemoglu, 2021).

§16Data Privacy, Sovereignty, and Digital Rights

AI systems are fundamentally data systems. Modern data protection frameworks, including the EU GDPR, establish individual rights over personal data: access, rectification, erasure, portability, and objection to automated decision-making (Regulation EU 2016/679). These frameworks are strained by AI systems trained on data from millions of sources with varied consents, and by the technical difficulty of removing individual data from trained models.

Data sovereignty — the principle that data about a country's citizens should be subject to that country's laws — is contested in an era of global cloud infrastructure. This is particularly acute for lower-income countries lacking the technical capacity and regulatory frameworks to govern data used to train AI systems operated by foreign companies. Equitable data governance frameworks protecting community data rights are among the most pressing needs in global AI policy (Coyle et al., 2020).

§17Emerging Frontiers

17.1Artificial General Intelligence

AGI — a hypothetical system performing any intellectual task a human can, with fluid cross-domain transfer — remains the long-term aspiration of much foundational AI research. Whether current architectures lead toward AGI, and on what timescale, is deeply contested. Some AI lab researchers believe AGI may be achieved within the current decade (Altman, 2023); others argue current architectures are fundamentally limited (LeCun, 2022; Marcus, 2022). The emergence of increasingly capable agentic AI systems blurs the practical boundary between narrow AI and AGI aspirations, without resolving fundamental theoretical questions.

17.2Agentic AI: The Current Frontier

As introduced in Sections 4.3 and 2.7, agentic AI represents the current deployment frontier. The progression from static model to autonomous agent involves tool use, memory, and planning — capabilities whose combination transforms AI from a content generator into a system capable of sustained autonomous action. Multi-agent frameworks, in which networks of collaborating AI models are applied to complex research pipelines (Boiko et al., 2023), software development (Park et al., 2023), and enterprise orchestration, are advancing rapidly. Governance frameworks developed for generative AI require significant extension to address agentic behaviour, particularly questions of oversight, reversibility, scope limitation, and liability (Wang et al., 2024).

17.3Multimodal Models, Embodied AI, and Scientific Discovery

Large multimodal models (LMMs) integrate language, vision, audio, and other modalities, enabling systems to answer questions about images, describe video content, and generate images from text. The WHO released dedicated ethics guidance for LMMs in health contexts in January 2024 (WHO, 2024). Embodied AI — systems learning through physical interaction with the world — is advancing through robot learning research and sim-to-real transfer (Brohan et al., 2022). AI is also transforming scientific discovery itself: AlphaFold 2 transformed structural biology; AI tools are accelerating materials science (Merchant et al., 2023); and automated laboratory systems are beginning to conduct multi-step experiments autonomously (Boiko et al., 2023).

17.4Neuromorphic Computing and Quantum AI

Neuromorphic computing builds hardware that more closely mimics biological neural circuits, using spiking signals and local learning rules rather than energy-intensive digital computation — with the potential to dramatically reduce AI's energy footprint (Davies et al., 2018). Quantum computing has theoretical applications in AI optimisation, sampling, and certain machine learning algorithms; practical quantum advantage over classical computers for AI tasks has not yet been demonstrated (Biamonte et al., 2017).

§18Navigating the AI Transition

18.1AI Literacy and Inclusive Governance

As AI becomes embedded in nearly every aspect of work, governance, healthcare, education, and culture, understanding and critically evaluating AI is becoming a civic necessity, not only a technical skill. AI literacy encompasses awareness of AI's presence and influence; conceptual understanding of how ML systems learn and where they fail; critical thinking skills for evaluating AI-generated content and algorithmic decisions; and the ability to use AI tools productively and responsibly (Ng et al., 2021). Bridging the AI divide between high-income and lower-income countries requires investment in data infrastructure, connectivity, AI research capacity, regulatory expertise, and local AI development ecosystems — alongside international frameworks for equitable data governance and benefit-sharing from AI productivity growth (Cihon, 2019).

18.2Conclusion

Artificial intelligence — discriminative, generative, and now agentic — will touch every aspect of human society over the coming decades. Its trajectory will be shaped not only by technical progress but by the political, economic, and governance choices made by individuals, institutions, and governments worldwide. AI is not destiny. It is a powerful tool capable of diagnosing cancer, accelerating drug discovery, supporting smallholder farmers, improving educational access, and modelling the climate systems on which all life depends. It is also a tool capable of automating surveillance, amplifying disinformation, perpetuating discrimination at scale, and concentrating power. Which of these futures is realised depends on choices: about what systems are built and for whom; what data is used and how it is governed; what uses are permitted and what are prohibited; how benefits are distributed and harms prevented; and who has a voice in making these decisions. These are not primarily technical questions. They are questions of values, politics, and power — and they belong to everyone.