There are lots of tasks LLMs can do well for instructors and researchers — and plenty they are poorly suited for. This guide is designed to help you navigate those strengths and weaknesses and save time and frustration in your teaching or research work.

A few things worth knowing upfront:

• LLMs are not truth-telling machines. Tools like Gemini, ChatGPT, or Copilot are designed to produce text the user is likely to find helpful and plausible — not necessarily text that is accurate. Sometimes those things align. Often they don’t.

• Web search has changed the picture — but not completely. Most major LLMs (ChatGPT, Gemini, Copilot, Grok) now integrate web search by default. This improves their performance on current events and some factual questions, but it does not eliminate hallucination or fabrication. The traffic-light guidance below reflects this: web-enabled models are somewhat more reliable in the yellow and red zones, but still require verification.

• Not all models are equal. Newer “reasoning models” — like OpenAI’s o3, Google’s Gemini 2.5 Pro, or Claude’s extended thinking mode — perform meaningfully better on logic, math, and multi-step analysis than standard LLMs. If you’re working on tasks that require careful reasoning, it’s worth trying one of these.

• Data privacy matters. When you upload a syllabus, student work, a draft paper, or research data to a commercial LLM, that content may be stored, used to improve the model, or subject to the company’s data policies. Avoid uploading sensitive, confidential, or unpublished research data to any tool whose privacy policy you haven’t reviewed. See the privacy note at the end of this guide.

Whether you’re using ChatGPT, Gemini, Grok, DeepSeek, Microsoft Copilot, or another tool, the principles below apply broadly.

Quick Reference

Use this table as an at-a-glance guide. Details and examples follow below.

🟢  Green Light — Go For It

Summarizing

LLMs are good at summarizing and paraphrasing text that you provide to them. This is useful when you’re working through papers, articles, or longer texts and want a quick overview before diving in, or a condensed version to review afterward. Paste the text directly into the prompt or upload a PDF — most major tools now support file uploads.

Keep in mind that summaries can flatten nuance or subtly misrepresent an author’s argument. Treat them as a starting point, not a substitute for reading.

Rewriting and Language Work

LLMs don’t “know” things, but they play with language very well. Some things to try:

• Paste a paragraph and ask them to change its tone, rephrase it in different terms, or simplify it for a different audience

• Turn a list of bullet points or keywords into a complete paragraph

• Turn a dense paragraph into a list of questions

• Upload a syllabus or assignment and ask for a more concise version

Explaining Concepts

LLMs are genuinely useful for breaking down difficult concepts, explaining cultural or historical references, or giving you a plain-language overview of an unfamiliar topic. They are also good at interpreting vague or half-formed questions — useful when you have a rough sense of what you’re looking for but aren’t sure how to ask it.

One caution: LLM explanations can sound authoritative even when they are wrong. This is especially risky for readers who don’t yet have enough background knowledge to catch errors. Cross-check important explanations against reliable sources, particularly in specialized or technical areas.

Design Feedback

LLMs can give useful suggestions when you provide them with a specific text to respond to. Try uploading a syllabus, assignment prompt, or reading list and asking:

• Is this reading appropriate for [level] students?

• What might students find confusing or be missing background on?

• How might I redistribute these readings across a shorter or longer time frame?

• What kinds of essay prompts or discussion questions would work well with this material?

These are tasks where the LLM is working from material you’ve provided, which reduces (though doesn’t eliminate) the risk of fabrication.

Language Practice

LLMs are competent in a wide range of languages and can produce grammatically correct text even in languages with smaller online footprints. They can understand you reasonably well even if you make significant grammar or vocabulary errors, which makes them useful for language learners practicing writing or conversation. See our guide on using AI to practice conversation skills.

Note: for high-stakes or published translation, see the yellow light section below.

Coding Assistance

This is one of the strongest green-light use cases and one of the most underused by researchers. LLMs can:

• Write, debug, and explain code in Python, R, JavaScript, and other languages

• Help you understand what a piece of unfamiliar code does

• Suggest approaches to data cleaning, analysis, or visualization

• Translate code from one language to another (e.g., R to Python)

Even researchers without programming backgrounds can use LLMs to get started with basic data analysis tasks by describing what they want in plain language. Always test code before relying on it — LLMs can produce plausible-looking code with subtle bugs.

🟡  Yellow Light — Proceed with Caution

Historical Overviews

LLMs can provide accurate, informative summaries of well-known historical events and broadly documented topics. However, be more cautious with:

• Minor historical figures or niche topics with limited source coverage

• Politically contested events, where the training data may reflect particular interpretive traditions

• Specific dates, statistics, or quoted passages, which are prone to fabrication

Think of LLMs as useful for orienting yourself in a topic, not for settling factual questions about it.

Translation

LLMs are capable translators but not always faithful ones. They are useful for:

• Checking how idiomatic an expression is

• Comparing the nuance between similar terms or phrases

• Getting a rough sense of a passage in an unfamiliar language

Be more cautious with longer texts or high-stakes translation. LLMs may change tone without signaling it, skip words or phrases, or fail to maintain consistency in how they render a specific term across a long document — and you may not have enough knowledge of the target language to catch it. For any translation that will be published or formally used, have a fluent speaker review the output.

Researching Sources

This depends heavily on how you’re using the tool.

Standard LLMs (free versions of Gemini, ChatGPT, Copilot, DeepSeek): do not trust them for specific source information. They will often produce plausible but fabricated citations.

Tools with uploaded documents (NotebookLM, ChatGPT Plus with uploaded files): much more reliable. When you explicitly provide the texts you want the LLM to work with, it restricts its answers to those sources. Citations and quotes from uploaded documents can be trusted at a higher level — though still worth spot-checking.

Web-search-enabled LLMs: better than standard models for current information, but still prone to errors. Treat results as a starting point and verify against the original source.

Brainstorming

LLMs can be useful conversation starters when you’re exploring a new research area or trying to map out a problem. They can surface angles you hadn’t considered, generate lists of possible questions, or help you articulate a half-formed idea.

Their limitation: they may not know enough about your specific domain to ask the right questions, identify what’s missing, or tell you what’s actually significant. A useful framing is to treat LLM brainstorming as a warm-up — something to react to and push against — rather than as a reliable map of the intellectual terrain. Follow up with colleagues, subject librarians, or domain-specific literature.

Image and Multimodal Analysis

Most major LLMs can now analyze images, charts, tables, and scanned documents — a capability that is particularly useful for researchers. You can upload a figure from a paper and ask the model to describe what it shows, or paste a table and ask it to summarize the key patterns.

Use this with caution: LLMs can misread charts, misinterpret axes, or describe visual patterns with more confidence than is warranted. Always check the model’s interpretation against your own reading of the image, especially for quantitative data.

🔴  Red Light — Find a Different Approach

Verifying Facts, Dates, and Quotes

LLMs are not reliable fact-checkers. They are designed to produce plausible, helpful-sounding text — and that text is sometimes accurate, but that’s not the primary goal. For specific facts, dates, statistics, or quoted passages, use Google or a trusted reference source.

A particular dynamic to watch for: LLMs tend to be more accurate in their first few responses on a topic, but as you prompt them further, they increasingly tell you what they think you want to hear rather than what is true. They will also often change their answers in response to pushback — not because they have updated their knowledge, but because they are responsive to social cues. A preferred truth? A preferred version of history? These are not things an LLM can reliably give you.

Citations and Quotes

A well-documented and persistent problem: LLMs fabricate citations. They may invent the author, the title, the journal, the page number, or the content of a quote — often in response to requests about specialized or niche research areas. The fabrications are frequently convincing enough to pass a quick read.

Do not use LLMs to generate citations, verify whether a quote is accurate, or confirm that a source exists. Use Google Scholar, your library’s databases, or a subject librarian. If you’re using a tool like NotebookLM with uploaded sources, citation reliability improves significantly — but still spot-check.

Domain-Specific Critical Analysis

LLMs are good writers but not good thinkers. They can produce text that looks like analysis but cannot reliably:

• Evaluate the quality of an argument in a specialized field

• Identify what is missing or overlooked in a piece of research

• Make well-reasoned estimates or hypotheses in a domain they don’t deeply understand

• Distinguish what is significant from what is superficial in a body of literature

LLMs default to listing pros and cons and presenting “both sides” of issues — a habit that can be helpful in some contexts but produces misleading results when applied to domain-specific questions where some positions are better supported than others. For critical analysis, there is no substitute for expert engagement.

Information About People

Use Google. This is especially true for anyone who is not a major public figure — researchers, minor historical figures, people with limited online presence. LLMs may produce correct information, may produce plausible-sounding wrong information, or may confuse one person with another. The risk is that a few accurate interactions can create misplaced confidence, leading you to stop double-checking. Just use a search engine.

A Note on Data Privacy

Before uploading any content to a commercial LLM, consider what you’re sharing and with whom:

• Student work: Uploading student assignments or identifying information to a commercial platform may implicate FERPA and your institution’s data policies. Check before you share.

• Unpublished research: Draft papers, datasets, and research notes uploaded to commercial tools may be stored and used to train future models, depending on the platform’s terms of service. Review privacy policies before uploading.

• Sensitive or confidential data: Do not upload data subject to IRB restrictions, data use agreements, or confidentiality obligations to commercial LLMs without explicit clearance.

• DeepSeek specifically: DeepSeek is a Chinese company subject to Chinese data law. Researchers working with sensitive materials should treat it with additional caution and consult their institution’s data governance office before use.

Tools that allow you to work with locally hosted models or institution-managed deployments offer stronger privacy guarantees for sensitive work. Your institution’s library or IT office may be able to advise on options.

Final Thought

LLMs are genuinely useful tools — for the right tasks. The goal of this guide is not to make you suspicious of them but to help you deploy them where they actually add value and avoid the frustration (and risk) of relying on them where they don’t. The landscape is changing quickly, and some of today’s limitations will improve. But the fundamental principle is likely to stay true for a while: always check and verify outputs.