Generative AI can help you explore a topic, understand difficult material, organize ideas, and revise your writing. It can also produce inaccurate references, incomplete summaries, or generate confident but incorrect claims (often called 'hallucinations'.
The value of an AI tool depends on the task, the information it can access, and how carefully you review the result. This guide introduces practical ways to use AI during study and research, AI features available through NCCU Libraries, academic research tools, and steps for checking and disclosing AI-assisted work.
AI products change quickly. Features, access conditions, quotas, and policies may have changed since the last review.
Chinese page is more comprehensive and up-to-date; you may consider using the Chinese version with a web translation plugin.
Where should I begin?
I am working on an assignment or thesis
Begin with Using AI in Research and Assignments. It explains where AI may be useful during planning, searching, reading, research design, writing, and translation—and where your own judgment remains essential.
I want to try AI features in library databases
Go to AI Features in NCCU Library Resources. The page explains what each feature searches, whether a personal account is required, and what you should check before relying on an answer.
I am comparing academic AI tools
Go to AI Tools for Academic Research. It covers Deep Research, literature-discovery tools, citation maps, evidence tables, and source-based tools such as Gemini Notebook.
I need to check an AI-generated reference or claim
Go to Verifying AI-Generated Content. You will learn how to confirm that a publication exists and whether it actually supports the statement attached to it.
I am unsure whether AI use is permitted
Go to Academic Integrity, Disclosure, and Citing AI. Course, department, research-ethics, journal, and professional requirements take priority over general advice.
A large language model produces text by predicting likely continuations. It can explain an idea clearly, reorganize notes, translate a paragraph, or suggest useful terminology. It can also produce a fluent answer containing a wrong date, an invented publication, an inaccurate quotation, or a conclusion that the cited source does not support.
Higher education is not only about producing a correct-looking answer. It is also where we learn to formulate questions, follow an argument, weigh evidence, and develop a position of our own.
Some AI concepts you may encounter
Reasoning models spend additional computation on multi-step problems before producing an answer. They may perform better on complex reasoning tasks, but this does not guarantee that the final answer is correct.
Deep Research systems can plan searches, retrieve multiple sources, and synthesize a longer report. Their choice and interpretation of sources still require review.
AI agents can use tools and take actions rather than simply generate an answer.
RAG (Retrieval-Augmented Generation) allows an AI system to retrieve external information before generating its response. Better sources can improve grounding, but poor retrieval can still produce poor answers.
MCP (Model Context Protocol) is an open standard that allows AI applications to connect with external tools and data sources. A connection does not by itself guarantee that the source is accurate, complete, authorized, or safe.
When AI May Not Be the Best Starting Point
AI can make it easier to begin a difficult task, especially when you are staring at a blank page or trying to make sense of an unfamiliar topic. It becomes less helpful when it removes the very practice that the task is meant to develop.
Before using an AI tool, ask yourself:
What am I trying to learn or decide?
Which part of this task needs my own judgment?
Do I know enough to recognize a weak answer?
What will I use to verify the result?
Am I sharing material that should remain private, or copyrighted materials (such as downloaded library journal articles)?
Consider using another source or method when:
an official or original source is readily available;
the task requires a complete and reproducible search;
the material is personal, confidential, unpublished, or licensed;
using AI would replace the skill being assessed;
you do not yet know enough to evaluate the output;
an exact legal, statistical, historical, or technical answer is required.
Developing a few key habits can make a big difference:
Check even when the input was correct. Errors can appear while AI is translating, summarizing, turning text into a table, or merging several documents.
Expect uneven coverage. AI models are trained predominantly on English-language content. They often miss licensed academic journals, Taiwanese local data, NCCU archives, and scholarship published in Chinese or other languages.
Try more than one route. Different models and search systems use different data, ranking methods, and content partnerships. The same question can produce quite different answers.
Use the direct source when you know where it is. If an official statistics site, legal database, or government agency already provides the answer, going through AI may only add another layer of checking.
Keep learning the subject. The best safeguard against a plausible error is enough background knowledge to notice that something does not fit.
Further Reading: AI cources on DeepLearning.AI, Coursera, and edx.
An AI search response often looks more complete than an ordinary results list. It may combine information from scholarly articles, news sites, Wikipedia, videos, commercial pages, discussion forums, and the search provider's own services. The reader sees one coherent paragraph, but the sources behind it may differ greatly in quality and purpose.
What to look at
Do not evaluate the prose alone. Check three things separately.
The answer: Is it specific enough? Does it distinguish fact from inference? Does it acknowledge uncertainty or disagreement?
The citation: Does the linked source actually support the sentence? Is the quotation exact? Is the DOI or URL valid, and does the publication actually exist?
The source set: Are there enough sources? Are important scholarly, official, local-language, or primary sources missing? Are several links repeating the same underlying material?
A useful personal test is simple: when an AI gives you a cited answer, open at least one source before you continue. Ask whether you would have chosen that source yourself.
Which sources are cited?
An analysis of more than 36 million Google AI Overviews and 46 million citations between March and August 2025 reported that Wikipedia, YouTube, Google-owned properties, Reddit, and Amazon together accounted for 38 percent of citations. The analysis also found substantial differences among systems. Reddit represented a particularly large share of some Perplexity and Google AI Overview citations.
How experts evaluated cited answers
A qualitative study asked 21 experts from different academic and professional backgrounds to evaluate complex searches in their own areas of expertise. They identified problems at three levels.
The wording of the answer
too general or incomplete for the question;
shaped by assumptions in the user's prompt;
more confident than the evidence justified;
simplified, shallow, or eager to please.
The citations
a correct-looking statement attached to an unrelated source;
selective use of one part of a source while ignoring the full argument;
no clear reason why one source was selected over another.
The source set
too few sources;
sources listed but not actually used;
weak, old, or untrustworthy material;
duplicate content presented as several sources.
Does AI search change user behavior?
A 2025 Pew Research Center analysis of browsing behavior found that users who encountered a Google AI summary were less likely to click through to a source website and more likely to end the search session sooner. Reduced source traffic can affect news organizations, reference projects, creators, and other sites that provide the information being summarized.
Some publishers and creators have responded by blocking AI crawlers. At the same time, AI services may still surface content through other arrangements or copies, making it difficult for users to understand exactly how a source entered the answer.
Bias can enter through training data, labeling, model design, search ranking, safety rules, and the way a question is framed. It may appear as stronger performance in English than in Chinese, greater familiarity with some regions than others, or a tendency to reproduce a dominant political or cultural viewpoint.
For historical, legal, political, and cultural questions, it is worth searching in more than one language and comparing AI results with primary sources and specialized databases. A model's confident judgment is not neutral evidence.
Prompt injection is another concern. A webpage, paper, or résumé may contain hidden instructions intended for an AI reviewer rather than a human reader. This is especially relevant when using agents that can read documents or act across email and files. Give an agent only the permissions it needs, and require confirmation before it takes an external action.
Explore further:
A — Public material
Information you would be comfortable posting openly online is generally the lowest-risk input.
B — De-identified material
Research data, student work, interview material, internal documents, and work processes may be usable only after names, contact details, identifiers, precise locations, and combinations of identifying details have been removed. Consent, contracts, and research ethics approval still apply.
C — Confidential material (Do not share)
Do not paste highly confidential, legally protected, unpublished, security-sensitive, or personally identifiable material into a public AI service. Use an approved institutional or local environment instead.
The same caution applies to licensed articles. Downloading an article through NCCU Library does not automatically give permission to upload it to another commercial platform.
Learn More: TCAI Guide: How to stop your images and data from being used to train AI
The New York Times v. Microsoft and OpenAI: allegations include reproduction of closely similar Times content and hallucinated material wrongly attributed to the newspaper.
Bartz, Graeber, and Johnson v. Anthropic: the dispute concerns books used for model training, including material alleged to have been obtained from pirate sites.
Taiwan Central News Agency litigation: a dispute involving the collection of news content for a training dataset.
Thomson Reuters / Westlaw v. Ross Intelligence and Lawbank v. SevenLaw: examples of database providers challenging the use of their content.
Other developments move in the opposite direction. Harvard released the Institutional Books 1.0 collection of public-domain material, while libraries and technology companies have funded digitization projects. Publishers including Wiley and Taylor & Francis have entered licensing arrangements with AI companies. These agreements may create new revenue, but researchers have also questioned whether publishers should license scholarly work without consulting the authors.
For NCCU users, the practical rule is simple: lawful access to an article through the Library does not necessarily include permission to transfer the PDF to an unrelated AI service. Check whether the item is open access and whether the database or publisher license permits the use.
Library databases
Useful when
you need academic journals, books, theses, statistics, legal material, or other licensed resources;
you want precise field searching, filters, citation relationships, or reproducible searches;
you need to save a query or set an alert.
How it works
Database fields are designed around the material they contain. Understanding the structure of scholarly articles and the terminology of a discipline improves the search. Advanced interfaces allow keyword combinations, filters, sorting, saved searches, alerts, citation export, and record management.
Limitations
Users need some subject knowledge and must learn the interface. Literature is distributed across databases; a weak keyword or the wrong database can produce very little. Most databases do not provide effective figure-level searching, and NCCU cannot supply content it has not licensed.
Search engines
Useful when
you are exploring a topic;
you need official websites, news, organizational reports, or gray literature;
you are looking for an open version of an item.
How it works
Search engines cover a broad web environment. Government, institutional, and reference pages may rank highly, but results can also include advertising, social media, content farms, and optimized marketing pages. Operators such as site:.gov, site:.edu, filetype:pdf, and quoted phrases can narrow down results.
Google Scholar limits the search to scholarly-looking material and shows citations and versions, but its records and automatic citation formats can contain errors.
Limitations
Most users read only the first pages of results, where ranking, advertising, and SEO have the greatest influence. Search engines do not provide access to every subscription database.
Generative AI
Useful when
the research question and terminology are still unfamiliar;
you need a customized explanation;
you want help combining text, images, public information, and user-provided material;
you need assistance with translation, organization, analysis, or visualization.
How it works
The source environment varies by product. Even when a system can search the web or accept uploaded documents, it usually selects a relatively small set of material for an answer. Academic AI can improve source quality, but may still rely on open-access documents or abstracts for paywalled publications.
The wording of the prompt affects what is retrieved. Repeating the same question in another language, in another product, or at another time can change both the sources and the synthesis.
Limitations
AI can invent references, distort real sources, expose personal data, add verification work, and make the search difficult to reproduce.
What Counts as Enough Evidence?
The answer depends on the task, discipline, and claim.
A short course assignment may require a focused set of reliable sources. A thesis needs broader engagement with the relevant research conversation. A systematic or scoping review requires a documented and reproducible effort to locate all eligible studies.
The appropriate source types also vary:
Humanities: Books, chapters, archives, primary sources, language variants, and interpretive depth may matter more than a large count of recent journal articles.
Social sciences: Reviews often combine qualitative and quantitative studies and must consider population, cultural setting, measurement, and representation.
Natural sciences and engineering: Recent journal and conference literature, precise terminology, scientific notation, standards, and citation networks may be central.
Medicine and health: Search protocols may use PICO, MeSH, trial registries, Cochrane guidance, predefined eligibility criteria, and documented stopping rules.
Taiwan-focused policy, law, and social questions: Government documents, local journals, news archives, statistics, theses, and Chinese-language material may be essential.
Do not judge completeness only by the number of results. Ask whether the search represents the source types, languages, periods, communities, and perspectives needed to answer the question.
Every long-running chatbot may behave differently because of memory, settings, or accumulated context. Rather than memorizing one “perfect” prompt, learn how to clarify the task and evaluate the response. You can also ask AI to help design or revise a prompt.
Useful official and educational resources include OpenAI, Anthropic, and Google prompt guides; the Prompt Engineering Guide; the University of Sydney's discipline-based examples.
Basic principles
Align knowledge first. Test whether the AI understands the subject and your information before asking it to perform a complex task.
Provide enough context. State the purpose, audience, known information, format, and constraints.
Use roles when they add real context. A role should clarify expertise and audience, not merely decorate the prompt.
Give examples. Show the type of output you need or ask the system to produce an example for review.
Follow up. Break a large task into smaller questions, ask for reasons and evidence, challenge the answer, or request competing viewpoints.
Leave room for exploration. A narrowly framed prompt can trap the answer inside your assumptions.
Avoid leading preferences. A prompt that announces the desired conclusion can distort the result.
Keep a task-specific conversation. Repeated examples and corrections can make a translation or editing thread more consistent.
Verify through other channels. Bring evidence from databases and original sources back into the conversation.
You can ask the AI to:
list each major claim with evidence and a source link;
identify where in the source the evidence appears;
separate contested viewpoints;
state when evidence is insufficient;
distinguish verified facts from inference;
give a confidence level and explain limitations;
point out bias or misleading assumptions in your prompt.
Other useful approaches include contextual anchoring, layered verification questions, progressively adding detail, asking the AI to review its own answer, and providing specific feedback.
A helpful way to structure your thinking is the CLEAR framework
Context: Explain the research question and goal.
Limits: Set the period, language, source type, or exclusions.
Expectations: State the desired output and level of detail.
Assessment: Ask the AI to evaluate source quality and representativeness.
Refinement: Invite questions and suggestions for improving the task.
Concise: Remove irrelevant wording.
Logical: State the order of steps.
Explicit: Specify the information, format, role, and tone.
Adaptive: Revise the prompt in response to the output.
Reflective: Continue evaluating and questioning the answer.
Search engine optimization, or SEO, is intended to make a webpage visible in conventional search. Generative engine optimization (GEO) considers how content is retrieved, interpreted, and cited by AI systems.
Consider these five practical approaches:
Semantic precision: Give each page a clear topic and support claims with facts.
Structured data: Use headings, metadata, tables, and machine-readable markup.
Signals of authority: Connect authors, institutions, publications, and citations consistently.
Freshness: Update content and display a clear revision date.
Factual accuracy: Verified content is easier for readers and retrieval systems to trust.
Google's quality guidance is often summarized as E-E-A-T: experience, expertise, authoritativeness, and trust. In the AI-search environment, first-hand experience and expert explanation can distinguish a useful page from generic generated material.
For researchers, useful additions include: ORCID and DOI links; structured and plain-language abstracts; FAQs; multilingual keywords and synonyms; tables connecting claims with evidence; explicit data sources and limitations; accessible HTML or Markdown; clear update dates.
AI systems often favor recent material when several pages cover a similar topic. Important research pages should therefore be reviewed rather than left unchanged for years.