How is AI used in different fields?

When someone says AI is changing your field, ask which job it does and who checks the result. A mistake means something different in a shop, a court, or a laboratory. These 9 documented uses mostly involve narrow tools inside larger processes.

Follow 3 questions through the examples: what goes in and comes out, what gets checked, and what a mistake costs. A scan goes in and a flag comes out. The check might be a regulator's review, a professional reading every answer, or a later comparison with the outcome. A wrong forecast leaves a shelf empty; a wrong citation misleads a court. Each field needs its own rules.

In medicine, the US Food and Drug Administration keeps a list of AI-enabled medical devices it has authorized for marketing. Radiology accounts for most entries, mainly software that analyzes medical images. The check is the FDA's review before the device can be marketed. The FDA says the list is not comprehensive, and it does not yet mark which devices use the large models behind chat assistants. The World Health Organization's guidance on large multi-modal models covers those models in health care, and notes it is not yet proven that they can do the wide range of tasks predicted for them. A fair question for any such device is whether it was tested on patients like the ones it will see.

In finance, much AI use is out of sight. In a 2024 survey by the Bank of England and the Financial Conduct Authority, 75% of responding UK firms said they already use AI, and 33% said they use it for fraud detection. The firms reported these unaudited counts themselves. A third of uses relied on outside providers. The US Treasury's report on AI in financial services names data privacy, bias, and dependence on those providers as risks, and recommends that firms check each use against existing law before it goes live. An unexplained model can leave a person refused credit or flagged as a fraud suspect.

Government services handle enormous payment volumes. The US Treasury reported that machine learning helped it recover $1 billion in check fraud in the year to September 2024, part of more than $4 billion in fraud and improper payments it prevented or recovered. The tool spots likely fraud faster so staff can act. The department reported its own dollar figure without publishing how many legitimate payments it flagged by mistake. A false flag can delay money someone depends on.

In education, a randomized trial in a Harvard physics course, with 194 students, found that students using a custom AI tutor learned more in less time than in an active-learning class covering the same material. The authors limit the result to introductory material, carefully designed prompts, high-quality videos, and a capable model. They also cite earlier studies where unguided chatbot use did not help. UNESCO's guidance for generative AI in education raises questions about student data privacy, and the age at which a child should talk to these tools alone.

In law, skipped checks can reach court. In a 2025 High Court judgment in England, judges dealt with 2 cases in which lawyers put false citations before the court, with actual or suspected use of AI tools. The court did not find that AI was used in every instance. It did say that freely available chat tools cannot do reliable legal research, and that lawyers must check AI output against authoritative sources such as official databases of laws and judgments. A court could rely on a case that does not exist.

Media has automated writing for years. In 2015, the Associated Press described how it generated more than 3,000 earnings stories each quarter from company financial data. That system predates today's chat assistants. AP said errors became rare once testing was done, and mostly came from a wrong number entered at the start. Its later standards for generative AI take a stricter line: output from a chat model is unvetted source material, not publishable copy. A false story would carry a trusted name.

In manufacturing, Siemens describes a visual inspection system for an unnamed car maker that checks stamped metal parts for splits and other defects, with cameras feeding a model that supports the human operator. Siemens reports up to 80% less manual classification effort. As in many industrial case studies, the vendor reports its own figure for an unnamed customer. A defective part could pass if the model never saw that kind of flaw.

Retail runs on forecasts. Amazon's science team describes how its demand forecasting changed over more than a decade, from single-number predictions to deep learning models that give a range of likely demand. A range matters because stocking depends on how uncertain the number is. Every forecast can be compared later with what customers bought, and the article names the hard cases, such as demand shifts during the COVID-19 pandemic. A wrong forecast can leave empty shelves or a warehouse full of unwanted stock.

In science, the 2024 Nobel Prize in Chemistry went in part to Demis Hassabis and John Jumper for AlphaFold2, a model that predicts a protein's 3D shape from its sequence. It has predicted structures for almost all of the roughly 200 million proteins researchers have identified. Each prediction carries its own confidence score, and EMBL-EBI's guide to that score explains why some regions get low scores. Experiments remain the final check. Years of lab work could depend on a low-confidence region.

Name the task, its input, and its output in one sentence. "Helps with documents" is not a task. "Drafts a summary of each support ticket for a person to approve" is.

Find out who checks the output, how often, and against what. A forecast can be checked against sales. A legal citation can be checked against a court database. If nobody can name the check, the tool is being trusted on its tone.

Most fields already had accuracy, privacy, and fairness rules. Ask which apply. The court and the Treasury report above both start from existing rules.

Keep the source beside the result: a regulator's list, peer-reviewed trial, court judgment, and vendor's case study offer different evidence. A vendor's percentage for an unnamed customer is a reason to ask questions, not a measurement to repeat.

Ask what a mistake costs and who pays it. The buyer of a tool and the person harmed by its error are often different people. For more, see reading model claims and AI rules and dates.

This is an authored illustration. The community hall that hosts the neighborhood book swap is offered a free AI tool. It reads sign-up forms and decides which volunteers get which shifts, including the shift supervising the children's corner. The vendor says it "saves organizers 5 hours a month."

Answer 3 questions: who is accountable for a bad assignment, what rules already apply, and what does a mistake cost?

Answer: it is reasonable to treat this as a scheduling tool, where a mistake means someone arrives at the wrong time. That holds for most shifts. The children's corner changes it. The hall stays accountable, not the vendor, and the hall probably already has rules about who may supervise children, such as background checks. The real check is whether the tool can see those rules at all. A wrong assignment there costs much more than 5 hours. The time saving is the vendor's claim, with no test behind it, so ask how it was measured before counting on it.

• FDA list of AI-enabled medical devices: the public record of authorized devices, updated periodically.
• Bank of England and FCA survey of AI in UK financial services: how firms say they use AI, including outside providers.
• UNESCO guidance for generative AI in education and research: a policy view on privacy, age limits, and teaching design.
• High Court judgment on false citations: the court's full reasoning on checking AI research.
• Stanford HAI AI Index 2026: a yearly data report with chapters on science, medicine, and the economy.
• On this site, will AI take my job, or change it? looks at the research on tasks, productivity, and jobs.