Questions and Concerns About AI

You can find AI useful and still have serious objections to how it is built or used. Trying a writing exercise does not require endorsing a company's data practices. These questions separate what you can check yourself from decisions that need employers, communities, researchers, and governments.

Sources checked October 29, 2026. Evidence and policies can change.

Choose the question closest to yours:

• Trust and understanding: wrong answers, search and calculators, intelligence, evidence of value.
• People and livelihoods: jobs, bias, creators' rights, who benefits.
• Your information and judgment: privacy, learning, professional help, scams.
• Wider choices: energy and water, whether you need to use AI.

Yes. An assistant can produce an invented fact, quotation, or source in perfectly ordinary prose. NIST calls this confabulation, often called hallucination. Fluent writing is not a reliability test.

Start with material you can inspect. Ask the assistant to separate what the source says from its suggestions, then compare dates, numbers, names, and quotations with the original. Open any cited source and check that it supports the claim. A real link can still accompany a wrong answer.

If you cannot check a consequential answer, don't act on it alone. How AI answers work explains why supplying good information helps without guaranteeing a good answer.

Often, you should. Search helps you locate original information. A calculator gives a defined answer to arithmetic. An AI assistant can help reorganize material, draft alternatives, or explain a passage through a conversation.

Suppose you have an event notice. Search can find the notice. AI can turn it into a draft invitation or a checklist. You still check the venue and time against the notice. If you only need the venue, reading the notice is probably quicker.

Choose the tool for the finished result, including the time spent checking it. Getting help from AI walks through that choice.

A language model learns patterns from large amounts of material and generates an answer from the information available to it. Predicting text during training can produce capabilities such as translation, explanation, and code generation. That does not establish human understanding, feelings, or a reliable sense of what is true. The QJE research paper's introduction to language models explains the mechanism.

For a practical decision, ask whether this system can do this task reliably enough, and how you would know. A convincing answer to a philosophy question would not settle either part. What is AI? distinguishes the broader field from chatbots.

There is evidence for specific uses. A 2023 randomized writing study assigned tasks to 453 college-educated professionals. ChatGPT access reduced average completion time by 40% and increased evaluator-rated quality by 18%. Those were bounded writing assignments, not a measurement of every kind of work.

In a published workplace study, researchers used data from 5,172 customer service agents to study an assistant rollout. Access to the assistant raised resolved issues per hour by 15% on average. Results varied: the most skilled agents saw small declines in quality.

Ask which people, task, tool, and outcome were studied. Your own trial may save less time, or none. AI's value in society examines uses beyond writing and the limits of the evidence.

Some tasks can be automated, jobs can change, and people can lose work. No guide can promise your job is safe. An estimate of what AI could do is also different from a count of jobs already lost.

The ILO's 2025 task-based index estimates that 1 in 4 workers worldwide is in an occupation with some generative AI exposure. It expects transformation to be the most likely overall effect because most jobs include tasks that need people. That is an assessment of potential change, not a guarantee about employers' choices.

For your own work, separate tasks that are easy to check from work involving trust, physical activity, responsibility, and judgment. Ask how your employer will measure errors, train staff, and share gains. Individual training cannot settle questions about wages, hiring, or worker protections. AI and work compares exposure, experiments, and employment evidence.

Only after checking the actual service and your permission to share the material. Chat storage, model training, human review, connected services, and deletion are separate questions. An employer-approved account may have different terms from a personal account. NIST identifies data privacy and security risks across the AI lifecycle.

Begin with fictional or public material. Keep passwords, identity documents, confidential work, and other people's sensitive information out of an unapproved service. Removing a name may leave enough detail to identify someone.

A privacy setting is useful only for the behavior it actually controls. AI privacy and actions explains what to check before sharing data or connecting an assistant to other tools.

It can. Training material, model design, and the way a system is used can produce unfair outcomes. NIST treats harmful bias as a risk to manage, rather than something erased by a polite instruction.

For a draft, check whose circumstances are missing and whether the wording makes assumptions about people. For hiring, lending, medical care, or access to services, informal spot checks are insufficient. Those uses need appropriate evaluation, accountable decision-makers, and a way for affected people to challenge mistakes.

Asking for an unbiased answer does not prove fairness. A helpful personal use cannot establish that a different system is fair to everyone it affects.

The costs are real. The IEA's 2026 analysis estimates that all data centers used 485 terawatt-hours of electricity in 2025. Its central projection reaches about 950 by 2030. These totals include services other than AI, and the future figure is a projection.

The same report says simple text and more intensive uses, including video generation and multi-step agents, can have very different energy demands. There is no single footprint that describes every prompt.

Use the simpler tool when it meets the need. At a larger scale, electricity sources, cooling, hardware, and local infrastructure matter. Personal restraint does not replace company transparency or community decisions about new facilities. AI energy and compute separates electricity, water, and the claims attached to them.

Training permission, compensation, output copying, and ownership of generated work are different questions. They deserve separate answers. The US Copyright Office's AI reports address digital replicas, copyrightability, and training; their conclusions are specific to US law and the questions examined.

For your own work, use material you own or have permission to share. Check licenses before publication. Don't assume that generated text or images are free of others' rights, or that a prompt gives you exclusive ownership. Avoid copying a person's work or impersonating them.

Law does not settle every question about consent or fair compensation. Choosing a useful AI task does not require treating creators' objections as solved.

It can become a shortcut around the practice you need. A 2025 randomized high-school mathematics study found that unrestricted AI assistance improved supported practice but harmed later unaided performance. A tutor designed to protect learning largely mitigated that harm in the study. This does not establish that every tutor works or every AI use damages learning.

Try first. Ask for a hint or a question, explain your reasoning, and check yourself later without assistance. At work, form your own view before asking for a critique. Keep enough knowledge to recognize a bad suggestion.

A finished answer and a learned skill are different results. Decide which one you need before asking the tool to do the work.

A useful role is helping you prepare questions, organize your own notes, or explain unfamiliar words for later checking. That is different from making a consequential decision about your health, rights, money, or safety.

Even a purpose-built accessibility service has boundaries. Be My Eyes explicitly says not to use its image descriptions for reading medicines, dosages, or other safety issues. A general chatbot should not be given more authority because it sounds confident.

Use a qualified person for consequential advice and the appropriate local service for an urgent problem. AI can help prepare for a conversation; it cannot supply the professional duty or relationship behind it.

AI can make deceptive material easier to produce. The FTC warns that scammers can clone a familiar voice to support a fake emergency.

If someone requests money or sensitive information, verify through a separate route, such as calling a number you already know. For a striking claim, look for the original source, date, and independent reporting. A convincing voice, realistic picture, or polished paragraph is weak evidence of authenticity.

Don't depend on visual oddities or a detector score to settle whether something is real. Verification matters even when you cannot tell how the material was made.

The person saving time may be different from the worker supplying training data, the creator whose work is used, or the community hosting the infrastructure. Useful output alone cannot establish a fair arrangement. The ILO finds uneven exposure across occupations, incomes, and genders; the IEA describes infrastructure pressures and growing community concerns.

Ask who can access the tool, who has a choice, who can challenge an error, and who receives the gains. Those are design and policy questions as well as personal purchasing decisions.

AI's value in society puts those questions beside the promising uses.

You can understand a technology without using it for everything. A sensible trial might end with deciding that your current method works better.

Choose a small task with public or fictional information. Compare the finished, checked result with your usual method. Keep the tool if it helps enough to justify the effort, cost, and tradeoffs. Explore everyday examples or start with what AI is.