The Good, the Bad and the Ugly of AI

Healthcare Diagnostics

The good. AI brings powerful advantages to healthcare by spotting patterns that human eyes often miss. Algorithms can scan X-rays, MRIs, or lab results in seconds, flagging potential problems long before they become critical. This not only helps doctors save time but also expands access to expert-level diagnostics in areas with limited healthcare resources. The result is earlier treatment, fewer errors, and better outcomes for patients.

The bad. AI in healthcare raises concerns around accuracy, trust, and bias. A model trained on limited or skewed data might miss diagnoses for certain populations, worsening existing health disparities. There are also ethical and legal questions: who is responsible if an AI-driven diagnosis is wrong — the doctor, the developer, or the hospital? Privacy is another issue, since training these systems requires enormous amounts of sensitive personal data.

The ugly. Nobody has settled how to validate a system that keeps changing. Regulators know how to approve a device that stays the same; an AI model changes with every update, and one cleared on last year's data may behave differently today. Meanwhile, a system that performs beautifully on average can fail consistently for an unlucky minority — harm that is real yet nearly invisible in the statistics, with no agreement on whose job it is to keep looking after deployment.

Energy Use

The good. AI offers huge potential for more sustainable energy systems. By analyzing real-time data, it can optimize how electricity flows through smart grids, balance renewable sources like wind and solar, and predict demand to reduce waste. For industries and households alike, AI can recommend efficiency measures, lowering costs while cutting carbon emissions.

The bad. The equation has hardened since the AI boom began: training and running large models now drives one of the fastest infrastructure build-outs in history. Data centres compete with households and industry for grid capacity and water, electricity prices climb in hosting regions, and utilities revive retired plants — including fossil ones — to meet demand. The efficiency gains AI enables are real, but so is its own appetite, and the two are rarely counted on the same ledger.

The ugly. The real numbers are secret. Companies report energy and water use selectively, definitions vary, and independent auditing barely exists — so societies are approving grid expansions on projections supplied by the parties who profit from them. AI is simultaneously a climate tool and a climate load, and almost no one is doing the accounting that would tell us the net effect.

AI Companionship

The good. AI companions can provide comfort, conversation, and emotional support to those who might otherwise feel isolated. For the elderly, people with disabilities, or anyone who wants company at any hour, AI can offer presence that learns personal preferences and adapts to moods. These systems may help reduce loneliness and, used well, contribute to better mental health.

The bad. Relationships with machines, however realistic, are not the same as human bonds and may discourage people from seeking real social connection. There is also the danger of commercialization — "companions" subtly designed to influence users' decisions, from shopping to politics. Regulators have begun to act, with China's rules on companion apps forcing the biggest platforms to switch features off in 2026, but most markets still rely on the companies' own restraint.

The ugly. Millions of people are already in continuing emotional relationships with systems owned by companies — which can change a companion's personality, move it behind a higher paywall, or retire it overnight. Real grief follows deprecated models. Because these products are tuned for engagement, the line between comforting a lonely user and cultivating a dependency is drawn, in practice, by a growth team. The experiment is running at population scale, without a control group, and its subjects include children.

Autonomous Systems

The good. The promise has begun to cash out: robotaxis carry paying passengers in a growing list of cities with safety records that in places surpass human drivers, delivery drones and warehouse automation are routine, and autonomous machines work in mines, ports, and disaster zones people cannot safely enter. The gains in around-the-clock productivity and avoided accidents are no longer hypothetical.

The bad. Accountability remains genuinely hard: when an autonomous vehicle harms someone, blame diffuses across the manufacturer, the software supplier, the operator, and the regulator that approved it. Job displacement lands unevenly — on drivers long before designers — and machines must still be programmed, implicitly or explicitly, for situations that offer only bad outcomes.

The ugly. An uncomfortable pattern has emerged: the human is often kept in the loop to absorb blame rather than to exert control. The safety operator supervising a system they cannot meaningfully oversee becomes what researchers call a moral crumple zone — the person facing charges after a failure. Incident data stays proprietary and settlements are sealed, so each company learns from its own crashes while society learns little from any of them.

Education and AI

The good. AI has the potential to transform education by creating personalized learning paths that adapt to each student's strengths and weaknesses. Automated tutors provide instant feedback, track progress, and free teachers to focus on human interaction. Administrative tasks, grading, and lesson planning can be streamlined, allowing educators to devote more time to creativity and mentorship.

The bad. The same tools may create over-reliance on algorithms at the expense of genuine human connection. If education becomes too standardized through AI, it risks flattening the diversity of thought that comes from teacher-student interaction. Privacy is a concern, since these systems collect sensitive learning data. And unequal access to technology can deepen the digital divide, leaving some students far behind.

The ugly. Assessment is the mess no institution has solved. The take-home essay — for a century the backbone of humanistic education — no longer proves anything; detection tools are unreliable and accuse the innocent; schools oscillate between bans they cannot enforce and adoption they cannot fund equitably. Underneath sits a harder question: when a capable assistant is always available, what should a student still learn by heart? No curriculum has a settled answer yet.

Coding with AI

The good. The change here has been dramatic: AI has gone from suggesting snippets to working as an autonomous colleague. Modern coding agents take on entire tasks — reading a codebase, writing features, running tests, fixing what fails — often working for hours while the developer only reviews the result. "Vibe coding," describing what you want in plain language and letting AI build it, has opened software creation to people who never learned to program. This museum practices what it observes: much of this very website, including the daily posts on the home page, is written, built, and published by AI agents.

The bad. The risks scaled with the capability. Code that no human has read line by line now ships to production, and plausible-looking output can hide subtle bugs or security flaws; "the AI wrote it" blurs accountability just when it matters most. The economics are reshaping careers — the routine work that once trained junior developers is vanishing, raising the question of where the next generation of senior engineers will come from. And when everyone builds with the same assistants, software risks converging on the same patterns, blind spots included.

The ugly. Nobody — including the companies that build these models — can fully explain why an AI writes the code it writes. As agents increasingly maintain code that other agents wrote, the share of software no human deeply understands grows year by year. The industry's answer so far is more AI reviewing AI. Whether that is a solution or an escalation is one of the field's most uncomfortable open questions.

Intellectual Property Rights

The good. AI trained on humanity's creative record puts craft skills within anyone's reach, and working creators use these tools daily as collaborators — for drafts, variations, translation, and tedium. A licensing market is actually emerging: publishers, archives, and image libraries now strike training-data deals, a revenue stream that did not exist three years ago. And provenance technology — content credentials that travel with an image and say how it was made — gives honest labeling a fighting chance. This museum follows that norm: its AI-generated daily posts are labeled as exactly that.

The bad. Most of today's models were trained on copyrighted work without consent or payment, and artists watch systems reproduce their recognizable manner on demand. Courts are splitting across jurisdictions, so the same act is lawful in one country and infringing in the next. The burden of opting out falls on individual creators facing automated crawlers. Meanwhile AI output floods the very markets that trained it — illustration, stock imagery, translation, session work — undercutting the people whose work made the tools possible.

The ugly. The deepest problems have no legal mechanism at all. A model cannot be un-trained: even where courts find the training unlawful, no remedy exists short of destroying the model, so the law's usual tools simply do not fit. Style is not copyrightable — and style is precisely what creators feel is being taken. And authorship itself is wobbling: purely AI-generated work receives no copyright, while human-AI collaboration sits in a gray zone no one can draw precisely — meaning a growing share of new culture may effectively belong to no one.

Robotics

The good. Robotics can take on dull, dirty, and dangerous work, improving safety and productivity. In factories and warehouses, robots handle precise, repetitive tasks around the clock; in hospitals they assist with surgery and logistics; in agriculture and disaster response they extend human reach and reduce risk. The long-promised humanoid robot has now reached the factory floor, with the first commercial deployments taking on repetitive industrial work.

The bad. Risks include job displacement, new safety hazards, and opaque decision-making when perception models fail. Over-automation can reduce resilience, create brittle supply chains, and concentrate power, while the mining, manufacturing, and disposal of hardware carry environmental costs. The labour conflicts have begun: 2026 saw the first major strike in the car industry with humanoid robots explicitly on the list of grievances.

The ugly. Robotics is where AI's mistakes gain mass and momentum — a hallucination in a chatbot is an embarrassment; in a machine beside a human worker it is an injury. Capability is routinely marketed years ahead of reality, with staged demos standing in for autonomy. And a robot that works alongside people necessarily watches them, all shift long — normalizing workplace surveillance as a by-product of automation, with the footage owned by the employer.

Laws, Regulations and Treaties

The good. For the first time, AI is governed by real law rather than voluntary pledges. The EU's AI Act phases in obligations by risk class; safety institutes on several continents test frontier models before release; China enforces its own strict rules; and the UN has begun structured global talks, informed by an independent scientific panel. Regulation arriving this early in a technology's life is historically rare — and it means transparency requirements, recourse for harm, and someone to hold accountable are slowly becoming enforceable.

The bad. The rulebooks are fragmenting into blocs — a rights-based European model, a lighter-touch American patchwork, a state-led Chinese system — and anyone building globally must satisfy all three. Compliance costs favour giants over startups, rules written for last year's AI age quickly, and enforcement budgets are thin against trillion-dollar incentives. Both failure modes are live at once: over-regulation that pushes research elsewhere, and under-regulation that leaves harms unanswered.

The ugly. There is no binding international treaty on the risks that matter most — frontier systems and military AI — and racing powers each fear constraining themselves while rivals sprint. Even the institutions are splitting: new international AI bodies are forming along geopolitical lines rather than across them. The UN's own scientific panel warns that the window for effective global governance "may not stay open for long." And the regulators writing the rules depend, for their expertise, largely on the companies they regulate.