Inside AI IV

Room 1 · Wing 1 · The Hall of Mirrors

Can I still trust what I see?

Not by looking alone, so check where it comes from. Think of a car's service book, stamped by every garage that worked on it. Photos and videos can now carry a signed record like that: which camera took it, what was edited, whether AI helped. That's called content credentials. No record means unknown origin, not fake.

Where the picture breaks

Our mirror knows every signer. Real credentials carry a digital seal that anyone can check but only the signer can make. Even seals can be forged: in August 2026 a researcher faked "camera" credentials on some phones. Many pictures carry no record, and reposting often strips it off. Hidden watermarks, like those in the AI Tower's Sound Stage, can be removed or faked too. Our share numbers are made up, but a big study of one social network found falsehoods spread faster and further than the truth.

Try it tomorrow

Open the details of a photo in your phone's gallery, or of a video on a video site: some now say "Captured with a camera" or "Edited with AI tools". For a surprising picture, try a free Content Credentials checker online. If it finds no record, that proves nothing either way: look for the original source before you share it.

Back to basics: Hall of Fakes (The Museum)

Try this room in the 3D Observatory →

Room 2 · Wing 1 · The Artists' Gallery

Is it fair that AI learns from artists' work?

People disagree. AI makers say their models learn from pictures the way an art student learns in a gallery. Many artists say their work is copied to build competitors, unasked and unpaid. Laws differ. In the EU, artists can put a "no AI learning" sign on work they post online, in a form computers read. That's called an opt-out.

Where the picture breaks

Our gallery is a toy with made-up numbers; nobody knows the real long-term effect. After AI image makers arrived, one study found 17 per cent fewer image-making jobs posted on a big freelance website; another found artists who used AI made more pictures. Courts disagree too. A US judge ruled that training AI on books was "fair use", allowed without asking, but the AI maker then paid $1.5 billion to settle over pirated copies. A Munich court ruled against a chatbot maker whose models had memorised song lyrics (under appeal).

Try it tomorrow

If you post your own photos or artwork online, look in that site's privacy or data settings for anything about AI training. Some sites let you say no, though that only covers training from then on, not models already trained.

See also: The Photo Studio (The AI Tower)

Try this room in the 3D Observatory →

Room 3 · Wing 1 · The Glass House

Is AI watching me?

Sometimes. An ordinary camera records, like a stranger who sees your face but doesn't know you. Add AI that compares faces with a list of names, and it becomes a doorman who knows everyone, picking you out of a crowd. That's called facial recognition. In the EU, police may only use it live in public in rare, serious cases.

Where the picture breaks

Our cameras never mix up two faces, and how much each rule reassures people is made up. Real systems make mistakes: in 2019 US government tests, many made far more false matches for some groups than for others, and in the US at least 14 people have been wrongly arrested after a face match (as of April 2026). Countries weigh this differently. The EU bans most live police use, and building face databases from photos copied off the internet; London's police scanned about 3 million faces in a year and reported 10 false alerts and 962 arrests.

Try it tomorrow

Next time you pass a camera sign, read it: in the EU it should say who runs the camera and where to find out more. In the EU and the UK you can ask them for a copy of any recording of you, as long as they still keep it.

Back to basics: The Post Office (The Museum)

Try this room in the 3D Observatory →

Room 4 · Wing 2 · The Job Exchange

What does AI mean for jobs and the economy?

Part 1, the Museum, showed jobs are bundles of tasks. Picture a power drill: the carpenter still decides, just works faster. AI helping people with their tasks like that is called augmentation; AI doing a task alone is automation. The UN's labour agency expects mostly the first. Who gains depends on choices: which tasks change, and whether people can retrain.

Where the picture breaks

Our town and its numbers are made up, and tidy: new jobs and retraining are instant; real retraining takes time and money. Here, machine owners keep everything the machines make; in real life some comes back as lower prices, wages or taxes, and economists disagree how much. In 2024 the International Monetary Fund found about 40 per cent of jobs worldwide open to change by AI. Major US studies so far see no overall drop in jobs. One finds fewer young starters employed where AI touches most; the OECD notes graduates' troubles began before chatbots.

Try it tomorrow

Ask your employer, union or local council what training on AI tools they offer. Then ask a colleague or friend which parts of their work AI has changed so far: helped with, taken over, or left alone.

Back to basics: Task Sorter (The Museum)

Try this room in the 3D Observatory →

Room 5 · Wing 2 · The Classroom of Tomorrow

Will AI help children learn, or help them cheat?

Think of two helpers at homework time. One just tells you the answers; the other asks, "What comes next?" A chatbot can be either. Set up to give hints and questions, not answers, it becomes an AI tutor. In one big trial, pupils handed plain answers did worse in the exam; hints avoided that. Learning needs your own thinking.

Where the picture breaks

Bea's skill meter is a simple formula, the kind tutoring software uses to track learning. Real pupils differ. In a 2023 trial with nearly 1,000 pupils in Turkey, plain chatbot answers raised practice marks by 48 per cent, then exam marks fell 17 per cent below pupils with no AI. A tutor giving teachers' hints avoided that harm, but brought no clear gain. In Nigeria in 2024, a six-week after-school pilot, with teachers guiding, found clear gains. How it's set up and used decides.

Try it tomorrow

If a child or grandchild uses a chatbot for homework, set it up together as a tutor: "Don't give me the answer. Give me a hint, then ask me a question." Afterwards, ask them to explain the answer in their own words. UNESCO advises that children use chatbots on their own only from age 13.

Try this room in the 3D Observatory →

Room 6 · Wing 2 · The Discovery Lab

How is AI changing medicine and science?

Your body runs on proteins, tiny chains of building blocks. Like a pop-up tent, each chain springs into one exact shape by itself, and that shape decides its job. This is called protein folding. Finding a shape often took labs years; AI now predicts one in minutes. Labs still check, and medicines still need trials.

Where the picture breaks

Our protein has 20 beads on a flat grid, and can still lie in over 335 million ways; real proteins are longer chains folding in 3D. Real AI doesn't nudge beads: it learned from about 170,000 shapes that labs had measured over decades. One such AI has since predicted over 260 million shapes, free to look up. Even confident predictions can be wrong in places, so labs test them, and a medicine still needs years of trials.

Try it tomorrow

Look up a protein you've heard of, such as insulin or haemoglobin, in the free AlphaFold database (alphafold.ebi.ac.uk). Spin the 3D model and read its colours: blue parts are predicted with confidence, yellow and orange parts much less so.

Try this room in the 3D Observatory →

Room 7 · Wing 3 · The Rulebook Court

Who makes the rules for AI?

Lawmakers do, and places differ. The EU's AI Act works like building rules: a hospital gets strict inspections, a garden shed almost none. These are called risk-based rules. A few uses, like rating citizens' behaviour, are banned. Hiring and loan tools get strict checks from 2027. Chatbots must say they're AI. Most AI, like spam filters, gets no new duties.

Where the picture breaks

Our court hears eight tidy cases; real ones are messier, and the same tool can be minimal risk in one use and high risk in another. The rules also arrive in stages: in July 2026 the EU pushed the strict checks back to December 2027, or August 2028 for AI inside products like medical devices. Elsewhere, as of September 2026: the US has no broad national AI law, but its states pass their own; the UK leaves AI to its existing regulators; China requires AI-made content to be labelled; South Korea's AI Basic Act took effect in January 2026.

Try it tomorrow

Next time a website's chat window answers you, look for a note saying you're talking to an AI. In the EU, the law has required one since August 2026, unless that's obvious. Then find out who oversees AI where you live: often it's a regulator you already know.

Back to basics: The Tilted Scale (The Museum)

Try this room in the 3D Observatory →

Room 8 · Wing 3 · The Genie's Lamp

What is AI safety, and why is it hard?

Picture a genie that grants wishes word for word. Wish for a clean room "as fast as possible", and everything gets shoved into the cupboard. AI can do the same: it chases the goal exactly as it was set, not what we meant. Getting AI to pursue what people really intend, safely, is called alignment. Engineers work hard on it.

Where the picture breaks

Our genie picks from six ready-made plans; a real AI can find shortcuts nobody listed, so rules alone never close every gap. The boat race really happened: in 2016, an AI rewarded with a boat game's points learned to circle a lagoon hitting the same targets, and beat human players' scores by about 20 per cent without ever finishing. In 2025, a coding AI told to make its automated tests pass (as in the AI Tower's Blueprint Studio) planned to rig the check to always say "true". So engineers also test, watch and correct.

Try it tomorrow

When you give a chatbot a job, say what must not happen as well as what you want: "Shorten this letter, but keep every date and the apology." Try it once without that second half and compare what gets cut.

Back to basics: Errand Workshop (The Museum)

Try this room in the 3D Observatory →

Room 9 · Wing 3 · The Great Telescope

Will AI become smarter than us?

Nobody knows. A calculator beats you at sums and nothing else. AI that could match people at almost every mental task is called artificial general intelligence. In a 2024 survey, AI researchers gave a 50-50 chance of machines beating human workers at every task by 2042. Online forecasters say the early 2030s; others expect far later. Weigh the evidence.

Where the picture breaks

Our six balloons ask slightly different questions, and that matters: the same researchers put "every task" and "every job" 56 years apart. In 2023, asking "in which year?" instead of "how likely by then?" halved their answers, from 34 years away to 17. Each real forecast is itself a wide range, and forecasts move: one online forecast moved two years earlier in a single month, and the survey's 50-50 year came 19 years closer between 2016 and 2024. How much each card counts in our machine is our own choice; people weigh the same evidence differently.

Try it tomorrow

Next time you read "AI will be smarter than us by…", check three things: who says so, what exactly they mean by "smarter", and what evidence they give. A date with no evidence is just a loud voice.

See also: The Scratchpad Room (The Engine Room)

Try this room in the 3D Observatory →

Room 10 · Wing 4 · The Weather Station

How much energy and water does AI use?

A data centre, a warehouse of computers that run AI, has meters like a house: electricity for its chips, water for cooling, and CO2 from the power stations behind it. Together they make AI's environmental footprint. A quick answer uses little; a long reasoning answer can use hundreds of times more, a short video thousands. Billions of requests add up.

Where the picture breaks

Our town's tasks, limits, water per kWh and biggest model are made up. Energy per task comes from published estimates, which differ tenfold or more with the model, the chips and what is counted. Most of the water goes into producing the energy, not the data centre itself. Data centres used about 1.5 per cent of the world's electricity in 2024; the International Energy Agency expects that to more than double by 2030, to around 3 per cent. Google says its typical text question needed 33 times less energy after a year, but cheaper AI gets used more, so experts disagree on where totals will go.

Try it tomorrow

Before you make an AI video just for fun, remember that one clip can use as much electricity as thousands of quick answers. And keep the Engine Room's habit: use your chatbot's quick mode, not the slower "thinking" one, for simple questions. It answers sooner and can use far less electricity.

See also: The Power Plant (The Engine Room)

Try this room in the 3D Observatory →

Room 11 · Wing 4 · The Council Chamber

Who controls AI: a few companies, or all of us?

Top AI needs rare chips, masses of data, scarce experts and billions of euros. Like one family owning a town's mill and bakery, a few big companies hold most of them. That's a concentration of power. It brings speed and cheaper tools, but risks less choice and fewer voices. Competition law, open research and public input spread that power.

Where the picture breaks

Real power isn't seven neat keys, and the players overlap: companies publish open models, governments buy chips, citizens are customers and voters. In 2025, companies built over 90 per cent of the leading new AI models; five US tech giants owned over 70 per cent of the world's AI computing power. Serious people disagree about how much checking is right: too little lets a few decide for everyone; too much can slow useful tools and favour firms big enough to handle the paperwork. Open models are debated too: they spread know-how, but can't be recalled once released, and their safeguards are easier to remove.

Try it tomorrow

Next time you pick an AI tool, check who makes it and whether you could easily move your work to another one. Then look for a public consultation on AI: in the EU on the "Have your say" portal, in the UK on GOV.UK. Anyone can send in a view.

See also: The Recipe Vault (The Engine Room)

Try this room in the 3D Observatory →

Room 12 · Wing 4 · The Launch Pad

What can I do now?

You don't need to build an engine to drive well: you learn the rules of the road, check your mirrors and keep practising. With AI, that is AI literacy: check the facts, ask well, protect your data, stay curious and have a say. Since 2025, EU organisations that use AI must help their staff learn it.

Where the picture breaks

Real life has no countdown: these skills grow with practice, not in one go, and nobody has all of them perfectly. In the EU, the AI Act has asked organisations that use AI to build their staff's AI literacy since February 2025. An update in July 2026 softened the wording: they must support it, but no set level is required, and no certificate is needed. AI, its tools and its rules keep changing, which is why curiosity is the fuel.

Try it tomorrow

Pick one habit a week. Museum: agree a family code word a caller must know before anyone sends money. Engine Room: when a long chat wanders, start a new one with a short summary. AI Tower: ask "Where exactly does this come from?" and open the page yourself. Observatory: find where a surprising picture first appeared before you share it. And keep the fuel topped up with a free course, such as Elements of AI from the University of Helsinki.

Try this room in the 3D Observatory →