ROMEOADVANCED ACADEMY

Lesson 2 of 5 · Understanding AI

Lesson 2

What can we do with AI today?

Eight categories of AI capability that are real, working, and in widespread use in 2026. With concrete examples — and an honest read on which categories are mature, which are emerging, and which are oversold.

35 minutesReading + optional hands-onBrowser optional

By the end of this lesson, you will:

  • Be able to list the eight broad areas where AI is already doing useful work in 2026.
  • Have a concrete, verifiable example for each — something a real person uses or a real company does.
  • Be able to tell the difference between an AI application that is genuinely useful and one that is marketing dressed as technology.

1. Language — reading, writing, translating, summarising

This is the area most people now know best, because it is what ChatGPT, Claude, and Gemini do. Modern AI systems can write coherent essays, translate between most languages, summarise long documents, answer questions about the contents of a PDF, draft emails in a particular tone, and have multi-turn conversations that stay broadly on topic.

Real examples: doctors using AI to draft visit notes for the patient record; lawyers using AI to summarise hundreds of pages of evidence; students using AI to help understand a difficult textbook (and, controversially, to write their essays); customer-service teams using AI to draft replies that a human reviews and sends; almost every email program now has an AI assistant built in.

Maturity: mature for everyday writing tasks; still error-prone for high-stakes claims that need to be exactly right.

2. Vision — recognising images, video, faces, scenes

AI systems can identify what is in a photograph, recognise faces, detect cancerous cells in a medical scan, read handwriting, watch a video and describe what is happening, and (controversially) recognise individuals from CCTV footage. Computer vision was, in fact, the area where the modern deep-learning revolution started in 2012.

Real examples: your phone's "search photos for 'dog'" feature; medical imaging tools that flag suspicious areas in X-rays for radiologists to review; self-driving cars; quality control on factory production lines; the iPhone's face-unlock; agricultural AI that detects diseased crops from drone footage.

Maturity: very mature for general object recognition; reliable but supervised for medical use; ethically contested for surveillance and policing.

3. Code — writing, debugging, refactoring software

Programming has been transformed by AI in the last three years. Tools like GitHub Copilot, Cursor, and Claude Code can write meaningful chunks of code from a description, debug existing code, explain what unfamiliar code does, write tests, and increasingly carry out multi-step development tasks. Many professional software engineers in 2026 spend more time prompting and reviewing AI-written code than writing code from scratch.

Real examples: GitHub Copilot is used by tens of millions of developers; new product features are now routinely built faster than they were two years ago; non-developers can create simple web applications by describing what they want in English.

Maturity: mature for routine code; still requires expert review for production systems; transforming the software industry in real time.

4. Science — accelerating discovery

AI is increasingly used as a research tool in the sciences. The most famous example is AlphaFold, a system from Google DeepMind that predicts the three-dimensional structure of proteins from their amino-acid sequence. This problem had been a grand challenge in biology for fifty years; AlphaFold solved it in 2020, and its 2024 successor (AlphaFold 3) extended the result to interactions between proteins, DNA, and small molecules. The result was a Nobel Prize in Chemistry in 2024.

Real examples: drug-discovery companies use AI to screen billions of candidate molecules in days rather than years; materials science labs use AI to design new batteries and solar cells; weather forecasts now incorporate AI models that beat traditional numerical methods; astronomers use AI to find unusual signals in telescope data.

Maturity: genuinely transformative in specific scientific problems; not yet "AI does science autonomously".

5. Healthcare — diagnosis, drug discovery, administrative work

Healthcare AI is doing three different things, with different levels of maturity. First, diagnosis: AI systems can detect diabetic retinopathy in eye scans, stroke patterns in brain scans, and certain cancers, often at expert-radiologist level. Second, drug discovery, as described above. Third, the unglamorous but quantitatively largest impact: drafting clinical notes, processing insurance paperwork, scheduling, and triaging patient messages.

Real examples: many NHS trusts in the UK and large hospital systems in the US are now using AI scribes that listen to a doctor-patient consultation and produce the visit note; the FDA has approved hundreds of AI-based medical devices; specific cancer-screening pathways have been transformed by AI.

Maturity: high in specific diagnostic tasks; high in administrative work; emerging but rapidly improving in drug discovery.

6. Creative work — images, music, video, design

AI can now generate images from text descriptions (Midjourney, DALL·E, Stable Diffusion, ImageFX), music from descriptions or hummed melodies, video from short prompts, and design templates for documents and presentations. The quality has gone from "obviously machine-made" in 2021 to "often indistinguishable from human work" by 2025.

Real examples: many advertising agencies now use AI for first-draft creative; small businesses generate marketing images without hiring a designer; novelists use AI as a writing partner; the entertainment industry is in active negotiation about how to handle AI-generated content — and several Hollywood strikes in 2023 turned on exactly this question.

Maturity: mature for non-final creative; contested for professional creative work; rapidly improving for video.

7. Robotics — physical action in the real world

AI has been slower in robotics than in software, because the real world is messier than a screen. But 2024 to 2026 saw notable progress. Humanoid robots from companies like Figure, 1X, and Boston Dynamics can walk, pick up objects, and follow spoken instructions in factory environments. Self-driving cars from Waymo and others now operate as commercial taxi services in several US cities. Warehouse robots from Amazon and others move billions of packages a year.

Real examples: Waymo runs paid taxi rides in San Francisco, Phoenix, and Los Angeles; Amazon warehouses move packages with substantial AI involvement; agricultural robots harvest specific crops; surgical robots assist in operating theatres.

Maturity: mature in controlled environments (warehouses, factories); emerging in semi-controlled environments (self-driving in good weather on good roads); still early for general-purpose home robots.

8. Decision-support — finance, logistics, education, government

AI is widely embedded in the systems that make day-to-day decisions about us: credit-scoring algorithms decide who gets a loan; fraud-detection systems decide whether your credit-card transaction is suspicious; logistics AI decides what arrives at the warehouse tomorrow; insurance AI decides how much your premium is; recommendation algorithms decide what you see on Netflix and TikTok; some governments use AI to flag tax returns for audit or to triage benefits applications.

Real examples: virtually every consumer financial product uses AI in some way; many supply chains are now AI-optimised; recommendation systems are AI; HR teams increasingly use AI in candidate screening (which is, as Lesson 5 notes, regulated under the EU AI Act).

Maturity: mature but ethically loaded — many of these decisions affect people in serious ways, and the question of whether the AI is making them fairly is a live debate.

Aside · What AI cannot do (yet)

For every category above, here are the limits in 2026. AI cannot reliably tell you when it is wrong (it can be confidently wrong). It cannot reliably do long-horizon plans that require persistent memory across days or weeks. It cannot move easily between very different tasks (a system trained for radiology cannot do drug discovery). It cannot, with rare exceptions, do original scientific discovery without humans framing the problem. And — perhaps most importantly — it does not understand the social, moral, or relational context of its decisions in the way a person does. It pattern-matches very well within the context it has been given. Outside that context, it can be surprisingly fragile.

Telling useful from oversold

Three questions to ask when someone tells you that "AI can do X".

1. Whose AI, doing what, against what evidence? "AI can pass the bar exam" is often used as a sweeping claim. The accurate version is "in one specific study, one specific model from one specific lab scored a particular percentile on a particular bar exam." All of those qualifiers matter.

2. In production, or in a paper? Many AI claims come from controlled studies. The system worked under laboratory conditions; whether it works in your hospital, your warehouse, or your business is a separate question.

3. Compared to what? "AI reads X-rays better than radiologists" — better than what kind of radiologist, on what kind of X-ray, under what conditions? Sometimes the comparison is fair. Sometimes the AI is being compared to a junior doctor on a particular condition where the AI was specifically trained, and the headline implies it would beat any radiologist on any X-ray.

Exercise — Map AI to your own world (15 minutes)

  1. Pick three categories from the eight above that touch your life or your work most directly.
  2. For each, write one concrete way you have already encountered AI in that category. If you cannot think of one, that may itself be informative — either you are missing it, or that category has not yet reached your part of life.
  3. For one of the three, ask: where is the AI doing something useful, and where is it producing more noise than value? Be honest. Examples are easier to think of when the AI is in the way: a chatbot that cannot actually help; a "personalised" email that is not personal; an automated decision you would have preferred a human to make.

Try it yourself (optional)

If you want to see one of these in action right now, the easiest is the language category. Open Claude.ai or ChatGPT (free-tier accounts work), and try three prompts.

  1. "Explain quantum computing as if I were ten years old."
  2. "I am writing an email to my boss to ask for a Friday off. The tone should be professional but warm. Draft it for me."
  3. "Here is a paragraph from a news article: [paste any paragraph]. Summarise it in one sentence."

That is a five-minute experience of category 1 — language. Notice what the assistant does well, and notice if it gets anything wrong. Both pieces of information are useful.

Self-check

  1. Name the eight categories of AI capability we covered.
  2. Pick the category most relevant to your work and describe one real example.
  3. What are three things AI cannot reliably do today?
  4. What three questions help you tell a useful AI claim from an oversold one?

Looking ahead

Lesson 3 is about the tools you can use and the companies behind them. Claude, ChatGPT, Gemini, Copilot, Midjourney, and the rest. The labs (OpenAI, Anthropic, Google DeepMind, Meta, xAI). The cloud providers (AWS, Microsoft Azure, Google Cloud). The hardware companies (Nvidia, and the start of competition around it). A map of the AI industry in plain language.