Switch language / تبديل اللغة
You are here

AI Leadership Insights

What Demis Hassabis says about the future of intelligence

By Prof. May El Barachi, Dean of Computer Science, University of Wollongong in Dubai. Published 22 December 2025. 

Demis Hassabis, co-founder of Google DeepMind and 2024 Nobel laureate in Chemistry, argues that progress toward AGI is not a straight line. In a recent interview with Prof. Hannah Fry, he describes today's AI as "jagged intelligence" - strong at narrow peaks, unreliable in the gaps - and bets that the next leap will come from world models, simulation, and reasoning research, not from scale alone.

Who is Demis Hassabis?

Demis Hassabis is a British AI researcher, entrepreneur, and neuroscientist best known as the co‑founder and CEO of Google DeepMind. He led the teams behind AlphaGo and AlphaFold, behind AlphaGo and AlphaFold, work that reshaped both AI and biology and culminated in his 2024 Nobel Prize in Chemistry for AI-driven protein structure prediction. He has been vocal on AI safety and governance, contributing to global policy discussions including the UK AI Safety Summit.

What makes Google Deepmind different?

Google DeepMind has been a major force in AI research for over a decade. Much of the work it began ten to fifteen years ago laid the groundwork for later advances, including transformer models that were eventually commercialized by organizations such as OpenAI. What differentiates DeepMind, in my view, is its focus on producing world-class research and science rather than a purely commercial focus. Hassabis' background in neuroscience clearly shapes this direction: drawing ideas from how humans think, learn, and plan, and applying them to machine learning. He often describes DeepMind's mission as "solving intelligence," then applying it to hard problems in science and society. 

What is the future of intelligence? Six takeaways from the interview 

After listening to Hassabis reflect on intelligence, one thing became clear: progress toward AGI is not a straight line, and it is not just about building bigger models. The reality is messier, and more interesting. 

1. A decade of progress in a year exposed real weaknesses 

Multimodal systems have advanced at remarkable speed. At the same time, this progress revealed an uncomfortable truth: models that can solve Olympiad-level problems may still fail at basic reasoning. Instead of smoothing intelligence, recent gains made its uneven nature obvious. 

2. The main obstacle is consistency, not raw ability 

Hassabis describes today's systems as "jagged intelligences" - impressive at certain peaks, unreliable in the gaps. Until AI can reason steadily across domains and recognize when it does not know something, general intelligence remains out of reach. 

3. Bigger models alone are not enough 

DeepMind is placing its bets evenly: half on scaling compute and data, half on new system designs. Scale helps, but progress also depends on better reasoning, handling uncertainty, and learning over longer time horizons. 

4. Language does not equal understanding 

Some parts of intelligence cannot be learned from text alone. Physical intuition, spatial reasoning, and interaction with the world require experience. This is why world models and simulation are becoming central to current research. 

5. Simulation may teach us why intelligence exists at all 

One of the most striking ideas discussed was the use of large-scale simulations to study how intelligence, social behavior, and even consciousness might arise. Running millions of controlled experiments could help explain not just how intelligence works, but why it emerged. 

6. AI may be overstated now, and still underestimated later 

Hassabis holds two views at once: parts of today's AI ecosystem are clearly inflated, yet the deeper, long-term effects - especially in science and energy - are still widely misunderstood. The biggest changes may arrive later, but cut much deeper. 

What were the new announcements and concepts? 

Beyond the broader reflections, six specific announcements or framings stood out in the interview. 

Deepened partnership with Commonwealth Fusion Systems 

Hassabis revealed that the collaboration with Commonwealth Fusion Systems is now much deeper than previously understood. The work goes beyond advisory roles into plasma containment and advanced materials, positioning fusion research as a real testbed for AI-driven scientific discovery. 

Genie and SIMA: an infinite training loop 

For the first time, Hassabis publicly described how Genie (DeepMind's world-model system) and SIMA (DeepMind's embodied-agent system) are intended to form an infinite self-improving loop. World models generate environments, agents act within them, outcomes refine the world model, and the cycle repeats. This frames embodied learning as central, not auxiliary, to AGI progress. 

Physics benchmarking via game engines 

A genuinely novel methodological detail: DeepMind is developing A-level physics benchmarks inside game engines. The goal is to test whether models actually respect Newtonian laws, not just predict outcomes statistically. This signals a shift from language-centric evaluation to grounded physical correctness. 

Whole-statement confidence scoring 

Hassabis outlined a concrete path to addressing model reliability: confidence is assessed across thinking steps and planning, validating entire statements rather than token-by-token probabilities. This is an important evolution toward trustable reasoning systems rather than fluent text generators. 

"Jagged intelligence" as a first-class concept 

He explicitly used the term "jagged intelligence" to describe how current systems excel in some areas while failing badly in others. This terminology formalizes a widely felt but rarely named limitation of state-of-the-art models. 

World models reaffirmed as his core obsession 

While not new in isolation, Hassabis reinforced that world models remain his longest-standing passion. He sharply contrasted spatial, embodied learning with today's LLM-dominant paradigm, calling out a fundamental gap that still blocks general intelligence. 

Closing thought 

What I appreciated most about the interview was its tone: ambitious but realistic, hopeful but honest. The future of intelligence will not arrive in a single dramatic moment. It will come from working through many difficult, often unglamorous problems that we are only beginning to grasp. 

FAQ

Q. What is "jagged intelligence"?
A. A term used by Demis Hassabis to describe AI systems that perform brilliantly on some tasks (Olympiad problems, protein folding) while failing on basic reasoning in the gaps between those peaks. He treats consistency, not raw ability, as the central remaining obstacle to AGI.

Q. Is AGI close, according to Hassabis?
A. He holds two views at once. Parts of today's AI ecosystem are clearly inflated, but the long-term effects - especially in science and energy - are still widely underestimated. He does not commit to a date; he commits to a research programme split evenly between scaling and new system design.

Q. What are Genie and SIMA?
A. Genie is DeepMind's world-model system, which generates simulated environments. SIMA is DeepMind's embodied-agent system, which acts inside those environments. Hassabis describes them as forming a self-improving loop: world models generate environments, agents act in them, outcomes refine the world model, repeat.

Q. Why are world models important?
A. Because some parts of intelligence - physical intuition, spatial reasoning, interaction with the world - cannot be learned from text alone. World models give AI a simulated environment in which to acquire experience that language data cannot supply.

About the author 

Prof. May El Barachi, Dean of Computer Science and Full Professor at the University of Wollongong in Dubai. Academic leader in digital innovation, applied AI and industry-aligned technology education. 

 

Computer Vision - advancements, applications, and future trends

By Prof. May El Barachi, Dean of Computer Science, University of Wollongong in Dubai. Published 2 December 2025. 

Computer vision is the subfield of AI that lets machines interpret images and video. Deep learning has pushed image recognition from around 50% to nearly 99% accuracy in under a decade, and the market is on track to exceed $50 billion by 2028. CNNs remain the workhorse, Vision Transformers are catching up fast, and generative models are reshaping how training data and content are produced. 

What is computer vision?

Computer vision (CV) is a subfield of artificial intelligence that enables machines to process, analyze, and interpret visual inputs such as images and videos. In essence, CV algorithms strive to replicate human vision - recognizing objects, people, and scenes in digital imagery and extracting meaningful information. 

Modern CV covers a range of tasks: image classification (identifying what an image contains), object detection (locating and labeling multiple objects), segmentation (precisely outlining objects or regions), and scene understanding and action recognition. These capabilities have advanced dramatically in the last decade thanks to deep learning and big data. Breakthroughs in neural networks have boosted image recognition accuracy from around 50% to nearly 99% in less than ten years - a quantum leap that showcases the field's potential. 

This progress, coupled with widespread industry adoption, has produced a booming market - valued at about $22 billion in 2023 and projected to exceed $50 billion by 2028. Computer vision is now not only a technical field but a major driver of business value in the AI era.

How do CNNs analyze images?

A major catalyst for the rise of computer vision has been the convolutional neural network (CNN). CNNs are specialized deep learning models designed for image analysis: they automatically learn hierarchies of visual features from raw pixel data. Lower layers detect simple patterns like edges or textures; deeper layers combine these into higher-level features such as shapes or object parts, ultimately recognizing complex objects or scenes.

This ability to discern intricate patterns has made CNNs the dominant architecture for tasks like image classification and object detection. Since AlexNet's breakthrough in 2012, CNN-based models - VGG, ResNet, EfficientNet - have continuously pushed the state of the art, enabling machines to classify images and detect objects with superhuman accuracy in some cases. In industry, CNN-powered solutions are everywhere, from real-time face recognition in smartphones to defect detection on assembly lines. They have been the primary deep learning model for image processing for much of the 2010s and remain fundamental building blocks today.

Are Vision Transformers replacing CNNs? 

Not yet, but the landscape is shifting. Vision Transformers (ViTs) and other attention-based architectures have recently emerged as powerful alternatives. ViTs apply transformer techniques - originally developed for language - to image patches, modeling global relationships in an image through self-attention. 

Thanks to this global context modeling, Vision Transformers often match or exceed CNN performance on image recognition tasks. The practical implication: CNNs still power most production CV applications, especially edge deployments requiring efficient inference, but new model innovations on the horizon could further enhance image analysis. For practitioners and businesses, CNNs remain core workhorses, with transformers and hybrid architectures rising fast.

How do GANs generate images? 

Beyond analyzing existing images, computer vision can also generate entirely new ones. The landmark innovation here is the Generative Adversarial Network (GAN), introduced in 2014. 

A GAN consists of two neural networks: a generator that creates synthetic images, and a discriminator that evaluates whether images are real or artificially generated. The two are trained together in a competitive "game": the generator tries to fool the discriminator by producing increasingly realistic images, and the discriminator learns to better distinguish fakes from genuine images. Over time, this adversarial process yields a generator capable of outputs so realistic that even the discriminator - or a human eye - can hardly tell they are fake. 

The GAN approach unleashed a wave of image-generation and creative AI applications. Early GANs produced blurry handwritten digits or faces; modern GANs like NVIDIA's StyleGAN generate hyper-realistic human faces, artwork, and even video frames. In industry, GANs and their variants are used for creating synthetic training data (for example, generating rare defect images to train inspection systems), enhancing image resolution (super-resolution), and producing photorealistic virtual try-on visuals or game scenery. The technology has also given rise to deepfakes - AI-generated imagery or video impersonations - highlighting both the power and the ethical challenges of image generation. 

More recently, the field has expanded beyond GANs to include diffusion models and transformer-based generators. Text-to-image systems such as OpenAI's DALL-E 3 and Stable Diffusion XL have dramatically improved the quality and realism of images generated from text descriptions, enabling new creative workflows in design, advertising, and entertainment. Businesses are already auto-generating product images and marketing graphics tailored to campaigns. GANs pioneered the era of AI image synthesis; ongoing advances continue to push the boundaries of what computers can imaginatively create.

Where is computer vision being used today? 

Computer vision is being adopted across many industries. The most prominent sectors and use cases are below. 

Manufacturing 

Automated visual inspection systems use CV for quality control on production lines, spotting defects or irregularities far more reliably and quickly than the human eye. CV also assists in inventory management by scanning and tracking stock items in warehouses. These applications help manufacturers improve yield, reduce waste, and ensure consistency. 

Healthcare 

In medical imaging, CV algorithms aid doctors by detecting diseases and anomalies in X-rays, MRIs, and CT scans with high accuracy. CV models can highlight potential tumors or pneumonia indicators on X-rays for radiologists. By automating image analysis, computer vision helps diagnose conditions earlier and with fewer errors, and it guides surgeons in precision robotics and treatment planning. 

Retail and e-commerce 

Computer vision enables innovative retail experiences such as Amazon's "Just Walk Out" stores, where cameras track what items customers pick up so they can be charged automatically without a checkout line. In e-commerce, CV powers virtual try-on tools (using augmented reality and pose estimation) that let shoppers see how clothing or accessories would look on them before buying. These applications boost customer engagement and sales while reducing return rates. 

Transportation and autonomous vehicles 

Self-driving cars and advanced driver-assistance systems rely heavily on CV to perceive their surroundings. Cameras - alongside lidar and radar - feed models that detect lane markings, traffic signs, signals, pedestrians, and other vehicles in real time. This enables safe driving decisions (steering, braking). Drones and unmanned aerial vehicles similarly use onboard vision for navigation and obstacle avoidance. CV is literally the "eyes" of the autonomy revolution. 

Security and surveillance 

CV enhances security by enabling automated surveillance and detection. Intelligent CCTV cameras can recognize faces or identify suspicious activities without human monitoring. In public safety, CV aids in spotting intruders, detecting weapons or accidents, and alerting authorities in real time. While these applications raise privacy concerns, they are increasingly used in airports, stadiums, and smart cities. 

Agriculture 

Advanced farming uses CV via cameras on drones, robots, or tractors to monitor crop health and farm conditions. CV systems analyze aerial images of fields to identify pest infestations, detect nutrient deficiencies through leaf color and texture, and estimate crop yields. Targeted actions like precision spraying of herbicides on weeds become possible, making agriculture more efficient and reducing chemical use. 

Robotics 

Many modern robots incorporate vision to interact with the world. Industrial robots use CV to locate and grasp objects on assembly lines, sorting systems recognize and route items, and delivery robots and warehouse AGVs navigate using vision-based SLAM (simultaneous localization and mapping). In healthcare, robotic assistants leverage vision for delicate tasks such as surgical robots that "see" the operative field. Computer vision gives robots the sensory input they need to operate autonomously and safely alongside humans. 

Each of these areas illustrates how CV is driving tangible value - cutting costs through automation and enabling entirely new products and experiences. Companies across sectors are investing in computer vision to gain competitive advantage. 

What is next for computer vision? 

Several trends are poised to shape the field over the next few years. 

Augmented and mixed reality everywhere 

With tech giants releasing consumer-grade AR devices such as Apple Vision Pro and Meta AR glasses, CV is expected to become even more prevalent in daily life. Computer vision will enable these devices to understand the environment - mapping surfaces, recognizing objects and people - so digital content can be overlaid believably onto the real world. This will enhance retail (interactive shopping), education (immersive learning), gaming, and professional training by blending virtual visuals with reality. 

Vision-language and multimodal AI 

The frontier of AI is moving toward multimodal systems that combine vision with other data types, particularly natural language. By integrating visual understanding with language comprehension, AI agents can interact more intuitively. Robots or home assistants with vision-language models can see an object and understand spoken instructions about it ("grab the red book on the table"). Generative models like CLIP and GPT-4's vision component allow zero-shot recognition of new objects from text descriptions. This convergence will enable AI customer service that can see a problem via camera, or AR glasses that respond to voice commands and visual cues. 

Enhanced 3D perception 

After conquering 2D images, computer vision is tackling 3D understanding. New techniques like neural radiance fields (NeRFs) allow AI to construct detailed 3D models of scenes from 2D images. Better depth perception and 3D object recognition will improve autonomous driving (more accurate distance and spatial awareness), robotics (better navigation and manipulation), and digital twins for industry. CV systems will not only detect what is in an image, but understand an object's shape, size, and position in the world — a crucial step for immersive AR/VR and realistic virtual simulations. 

Edge computing and real-time vision 

There is a push to run CV on the edge - directly on cameras, smartphones, and IoT sensors - rather than in the cloud. On-device processing reduces latency and improves privacy because raw images never leave the device. Techniques such as model quantization, pruning, and efficient CNN architectures enable high-performance CV in resource-constrained environments. This is vital for time-sensitive use cases: factory robots and self-driving cars cannot afford cloud delays. Expect more optimized vision AI chips and embedded CV software powering smart cameras, drones, AR glasses, and other edge devices. 

Generative AI for synthetic data and content 

Generative models (GANs, diffusion models) can now produce highly realistic images. A major emerging trend is using generative AI to create synthetic training data for computer vision. When real data is scarce or sensitive, companies can generate simulated images - thousands of synthetic medical scans or factory defect images - to train CV models without costly manual data collection. Synthetic data can also help address biases and privacy by augmenting datasets in a controlled way. Generative AI is also used for on-the-fly image augmentation, editing (removing objects, changing backgrounds), and generating entire virtual worlds for simulation. This will accelerate model development and unlock new creative applications. 

Advanced vision architectures and foundation models 

We are entering an era of foundation models in vision - large pretrained models that can be adapted to many CV tasks. Vision Transformers and hybrid models lead this charge by offering robust performance across classification, detection, and segmentation. Tech companies are developing massive vision-language models (multimodal GPT-style) that understand images in the context of text, and universal segmentation models like Meta's Segment Anything Model that generalize to segment any object. These foundation models can be fine-tuned for specific applications with relatively little data, making CV development more accessible and scalable. Expect more "generalist" vision AI models that can describe images, answer questions about them, and detect anomalies - analogous to how large language models function. 

Ethical and trustworthy vision AI 

As CV permeates high-stakes domains - security, healthcare, automotive - there is growing focus on ethics, bias, and safety. One aspect is detecting and countering deepfakes and manipulated media; CV algorithms themselves are being employed to spot telltale signs of fake images or videos, helping maintain information integrity. Another aspect is addressing bias, for instance ensuring face recognition works fairly across demographics and does not invade privacy. Regulators and societies are increasingly concerned with how vision AI is used (surveillance vs. civil liberties), so expect more guidelines and tools for explainable and responsible CV. Techniques like explainable AI for vision (highlighting which image regions influenced a decision) and privacy-preserving vision (blurring faces, federated learning on device) will become standard. The next phase of CV will not just be about what the technology can do, but how it is implemented - transparently, fairly, and securely. 

Concluding remarks 

Computer vision has grown from a niche research area into a transformative technology fueling innovation across industries. From enabling autonomous machines to unlocking new insights in business data, the ability of AI to interpret visual information is a key component of modern "agentic" AI solutions. The field continues to advance -algorithms are getting more powerful, datasets bigger, and computing hardware faster - creating a positive feedback loop of progress. Industry leaders recognize the opportunity: the global CV market is already tens of billions of dollars and attracting heavy investment as organizations seek to improve efficiency, safety, and customer experience through vision AI. 

Looking ahead, computer vision will become even more ubiquitous and integrated into everyday products. Cameras are everywhere in the modern world; with AI, every camera can become a smart sensor that not only records visuals but also understands and reacts to them. This opens the door to smarter cities, smarter homes, and more adaptive intelligent agents all around us. For business leaders and developers, computer vision is a maturing but still rapidly evolving field - those who stay abreast of the latest CV advancements, from CNNs to transformers, from GANs to generative data augmentation, will be well positioned to build the next generation of AI-driven solutions. Computer vision's journey is far from over; as it converges with other AI disciplines and we address ethics and deployment, CV will continue to redefine how machines see the world, and how we interact with an AI-powered visual one. 

FAQ

Q. What is computer vision used for in business?
A. Quality control in manufacturing, medical imaging and diagnostics in healthcare, frictionless checkout and virtual try-on in retail, perception for self-driving cars and drones in transportation, automated surveillance, crop monitoring in agriculture, and object manipulation in robotics.

Q. Are CNNs still relevant?
A. Yes. CNNs remain the dominant architecture in production, particularly for edge deployments requiring efficient inference. Vision Transformers and hybrid architectures are emerging as powerful alternatives, but CNNs continue to be core workhorses in computer vision systems.

Q. How big is the computer vision market?
A. The global CV market was valued at about $22 billion in 2023 and is projected to exceed $50 billion by 2028.

Q. Have GANs been replaced by newer methods?
A. Not replaced - joined. The field has expanded beyond GANs to include diffusion models and transformer-based generators. Text-to-image systems such as OpenAI's DALL-E 3 and Stable Diffusion XL have dramatically improved the quality and realism of generated images from text, enabling new creative workflows in design, advertising, and entertainment. GANs pioneered the era of AI image synthesis; the newer methods continue to push the boundaries.

Q. What are foundation models in vision?
A. Large pretrained models that can be adapted to many CV tasks with relatively little task-specific data. Examples include Vision Transformers, multimodal vision-language models (GPT-style models that understand images in the context of text), and universal segmentation models such as Meta's Segment Anything Model.

About the author 

Prof. May El Barachi, Dean of Computer Science and Full Professor at the University of Wollongong in Dubai. Academic leader in digital innovation, applied AI and industry-aligned technology education. 

Unlocking the potential of AI in healthcare

By Prof. May El Barachi, Dean of Computer Science, University of Wollongong in Dubai. Published 14 November 2025. 

AI in healthcare is reshaping how care is delivered - automating routine tasks, sharpening diagnostic accuracy, and enabling proactive care from wearables to genomics. Real gains are already documented: ambient scribes cutting documentation time up to 70%, predictive analytics reducing emergency visits by 68%, clinical chatbots cutting readmissions by 30%. 

What is AI doing in healthcare today? 

Artificial intelligence is reshaping healthcare by automating routine tasks, enhancing diagnostic accuracy, and enabling proactive care, all while aiming to make complex technological solutions accessible and understandable. Drawing from recent advancements as of 2025, AI is not just a tool but a transformative force that predicts, learns, and acts to reinvigorate modern medicine - from linking genetic codes to powering robotic assistants in surgery. 

Across patients, clinicians, and administrators, AI simplifies operations, often completing human-like tasks more efficiently and cost-effectively. 

However, AI offers tremendous promise only if it is implemented with careful consideration of ethical issues such as data bias and privacy, to ensure equitable benefits. The sections below expand on key examples with real-world applications and emerging trends. 

How does AI improve early disease detection? 

AI excels in early detection by processing signals from wearables or imaging devices to flag potential issues before they escalate. 

For someone predisposed to conditions like cardiovascular disease, epilepsy, or diabetes, a smartwatch or sensor can monitor heart rate, blood sugar, or neurological patterns in real time. An AI model then analyzes this data to predict silent heart attacks, strokes, or seizures, alerting users or doctors promptly. 

Beyond wearables, AI interprets brain scans with remarkable precision: trained on thousands of images, it can detect stroke timing twice as accurately as human experts, enabling faster interventions that save lives. In orthopedics, AI reviews X-rays to identify fractures that might be overlooked - up to 10% of cases by radiologists - minimizing errors and reducing the need for follow-up scans. 

In disease surveillance, AI models trained on population data can predict over 1,000 conditions, such as Alzheimer's or kidney disease, years in advance by spotting subtle patterns in MRIs or health records. These capabilities are particularly valuable in under-resourced areas, where AI augments limited specialist availability. 

How does AI support remote monitoring for chronic conditions? 

For patients with ongoing illnesses, AI enables continuous oversight without constant hospital visits. 

Sensors in devices track vital signs, movement, or even cough patterns, feeding data into models that detect anomalies - a sudden drop in oxygen levels for COPD patients, or mood shifts relevant to mental health. Proactive alerts can summon help during episodes such as falls or irregular heartbeats, improving quality of life for the elderly and those in remote locations. 

Tools like AI-powered apps for chronic cough analysis and platforms that reduce emergency visits by 68% through predictive analytics exemplify this approach, allowing healthcare teams to intervene early and optimize resource use. This cuts costs and empowers patients with self-management insights, though integration with human care remains essential to address false positives. 

How is AI accelerating medicine development? 

AI revolutionizes drug discovery by simulating molecular interactions at speeds impossible for humans. 

By inputting vast datasets on chemical behaviors, disease pathways, and past trials, models predict effective compound combinations for specific conditions, slashing development timelines from years to months. Generative AI further innovates by creating synthetic datasets for rare diseases or modeling drug efficacy, as seen in platforms that simulate personalized treatment responses. 

In 2025, this has led to breakthroughs in targeted therapies, such as AI-optimized vaccines or antivirals, with tools reducing clinical trial failures by identifying viable candidates early. Collaboration with traditional research ensures safety: AI complements rather than replaces expert oversight. 

What other applications of AI are emerging in healthcare? 

Beyond detection, monitoring, and drug discovery, AI is reaching into administration, training, and patient support. 

  • Documentation: ambient scribe tools transcribe consultations using speech recognition and NLP, cutting documentation time by up to 70% and giving doctors more face-time with patients. 
  • Clinical chatbots: powered by retrieval-augmented generation, they answer medical queries accurately 58% of the time, guide decisions, and reduce readmissions by 30%. 
  • Medical education: AI simulates patient scenarios for training, providing personalized feedback to students. 
  • Mental health support: 24/7 chatbots offer support, tracking symptoms and suggesting coping strategies. 
  • Ambulance triage: AI assesses patient needs with 80% accuracy based on vitals and history, aiding paramedics. 
  • Traditional medicine integration: AI catalogs indigenous knowledge to discover new compounds, blending ancient wisdom with modern tech. 

How is AI enabling personalized medicine? 

AI and machine learning tackle the overwhelm of vast, varied datasets that doctors face daily. Traditional analysis of records, images, and histories is time-consuming, but deep learning models excel at processing unstructured data - radiology scans, blood tests, EKGs, genomics, and patient histories. This provides real-time insights, such as flagging abnormalities in images or correlating symptoms with underlying causes, enhancing diagnostic speed and accuracy. 

Precision medicine takes this further by customizing treatments to patient subgroups rather than a blanket approach. For ovarian cancer, machine learning analyzed 32 blood markers to identify early-stage patients with poor prognoses, uncovering hidden disease groups and guiding targeted therapies. This data-driven method reveals interactions between variables that hypothesis-based research might miss, especially in multifactorial diseases. 

AI also predicts drug responses via genetic markers, optimizes dosages to minimize side effects, and reshapes clinical decisions for conditions like autoimmune disorders or cancer. Additional advancements include AI for diabetic retinopathy - models that screen retinal images instantly, reducing wait times from weeks to minutes in underserved areas. In genomics, AI interprets sequences to inform therapy, as in pediatric brain tumors where it identifies subgroups amenable to less invasive treatments, avoiding long-term side effects. For cardiovascular risks, AI combines electronic health records with genetics for better predictions. Environmental factors are integrated too, with AI forecasting outbreaks or toxin exposures. 

Challenges like data bias persist, but synergies between AI and human expertise promise more equitable, effective care. 

What guardrails matter so innovation scales safely? 

  • Data quality and governance: unify sources, define stewardship, monitor drift. 
  • Bias and equity: validate across demographics; track outcomes, not just accuracy. 
  • Privacy and security: least-privilege access, auditability, and privacy-preserving options. 
  • Clinical integration: design for the last mile - alert-fatigue control, clear accountability, human oversight. 
  • Change management: upskill clinicians, update SOPs, and align incentives to value-based care. 

What should healthcare leaders do next? 

  1. Start where data is strong (imaging, triage, documentation). 
  1. Pick two or three high-impact use cases with measurable KPIs (time-to-diagnosis, readmissions, clinician minutes saved). 
  1. Build the platform, not one-offs: interoperability, monitoring, model registry, reuse. 
  1. Establish an AI governance council (clinicians, data, legal, ethics, patient reps). 
  1. Invest in skills: clinicians fluent in data; pharmacists and nurses comfortable with AI tools; engineers who understand clinical workflows. 

A 30-day quick win 

  • Pilot an ambient scribe in one clinic; measure note time, after-hours charting, and clinician satisfaction. 
  • Add a radiology-assist model for one indication, for example fracture detection; track sensitivity, specificity, and secondary reads saved. 
  • Stand up a lightweight AI registry and monitoring dashboard (ownership, versioning, metrics, drift alerts). 

Closing thought 

AI in healthcare is not about replacing clinicians - it is about amplifying them. The future of care will be defined not by the machines we build, but by how we use them: to heal, to connect, and to make high-quality care accessible to more people. 

FAQ

Q. Will AI replace doctors?
A. No. The pattern across deployments is amplification, not replacement: ambient scribes free clinician time, imaging assists catch overlooked findings, chatbots triage routine queries. Human judgment and accountability remain at the centre of care.

Q. Where are the easiest wins for hospitals starting now?
A. Three domains where data is strong and ROI is measurable: ambient documentation, imaging assists (e.g., fracture or stroke detection), and triage. Each has documented gains and a clear KPI.

Q. What are the biggest risks?
A. Data bias and inequitable outcomes, privacy leakage, alert fatigue from poor integration, and accountability gaps when an AI recommendation contributes to a decision.

Q. How is AI used in drug discovery?
A. By simulating molecular interactions across vast chemical, biological, and trial datasets to predict promising compounds, shortening discovery timelines from years to months. Generative AI also produces synthetic datasets for rare diseases and models personalized treatment responses.

Q. What is precision medicine?
A. An approach that customizes treatment to patient subgroups defined by data — genetics, biomarkers, clinical history - rather than applying a single protocol to everyone. AI surfaces the subgroups by finding patterns hypothesis-based research would miss.

About the author 

Prof. May El Barachi, Dean of Computer Science and Full Professor at the University of Wollongong in Dubai. Academic leader in digital innovation, applied AI and industry-aligned technology education.