How to use ChatGPT effectively despite its limits

ChatGPT can search the web, analyze files and data, work with images, write and debug code, and conduct multi-source research. It can save time, simplify repetitive tasks, and help people work through difficult problems.

But greater capability can make its limitations easier to forget. ChatGPT can produce confident errors, miss context, reflect bias, and make poor decisions about work it does not fully understand. The solution is not to avoid AI, but to use it for the parts of a job it handles well while keeping human judgment where it matters.

This article focuses on ChatGPT, but many of the same principles apply to other large language model (LLM) tools, including Claude, Gemini, and Perplexity.

ChatGPT can sound convincing when it is wrong

ChatGPT can produce incorrect facts, invented details, and unsupported conclusions in clear, confident language. Use it to find leads, summarize material, and accelerate research, but treat factual claims as unverified until you have checked them against credible sources.

AI errors, often called “hallucinations,” can include:

• Incorrect facts or dates
• Nonexistent or misidentified sources
• Invented quotations
• Details that weren’t present in material you provided
• Conclusions that go further than the available evidence

Web access does not eliminate the problem. An AI research tool can select weak or outdated sources, misunderstand what a source says, miss important information, or be unable to access material behind paywalls and other technical restrictions.

For important work, open the sources and confirm that they exist, support the claims being made, and are credible enough for your purpose. Ask ChatGPT to distinguish facts from assumptions and identify uncertainty rather than fill gaps with plausible details.

AI can make research much faster. It can’t decide what evidence you should trust.

Give ChatGPT the context it lacks

ChatGPT does not automatically know your goal, audience, standards, or which details matter most. Vague requests such as “research this,” “improve this,” or “write something about this” leave it to make those decisions for you. Define the job clearly, provide the relevant material, and tell it what a useful result must accomplish.

For better results:

• Define the goal. Explain what you are trying to accomplish, who the audience is, and what the result should do.
• Provide the necessary material. Share the sources, notes, data, examples, or background it needs instead of expecting it to reconstruct the project.
• Protect important details. Identify facts, quotations, terminology, dates, and other material that should not be changed or overlooked.
• Break complex work into stages. Handle research, analysis, organization, drafting, and editing separately when one large request produces shallow or confused results.
• Work iteratively. Treat the first response as a starting point. Tell ChatGPT what worked, what it misunderstood, and what should change, then refine the result through focused follow-up requests.
• Specify the kind of help you want. Ask for criticism, alternatives, missing evidence, or weaknesses if you want analysis rather than agreement.
• Tell it not to guess. Ask it to flag missing information and clearly label assumptions.

Clear instructions do more than improve the first answer. They give you standards for evaluating whether the answer is actually useful.

Look for bias and missing complexity

Generative AI can reflect biases in its training material, user prompts, and online sources. It may overlook perspectives, accept a question’s assumptions, flatten a complicated disagreement, or make an uncertain interpretation sound definitive. 

Ask what evidence supports the answer, which perspectives may be missing, and whether competing positions are equally well supported. 

You can also ask ChatGPT to challenge your premise directly: What evidence contradicts it? What assumptions are built into the question? What would a knowledgeable critic say? 

ChatGPT can be overly agreeable, particularly when a prompt suggests the expected answer. Inviting disagreement makes it more useful as a thinking tool.

Protect confidential information

ChatGPT cannot decide whether you have the right to share the information you enter. Before pasting text or uploading a file, check it for sensitive material and confirm that your use complies with your organization’s policies and the service’s privacy settings.

Sensitive material may include:

• Proprietary or confidential business information
• Customer or employee data
• Personal information
• Unpublished or sensitive work
• Material you don’t own or have permission to share

The same basic rule applies to AI as to other cloud services: know what you are sharing, where you are sharing it, and whether you have permission to do so.

Keep human judgment at the center

ChatGPT can generate options and perform tasks, but it cannot decide what your project should accomplish or take responsibility for the result. The most effective workflow keeps the goal, important judgment calls, and final approval in human hands while using AI for research assistance, organization, comparison, troubleshooting, and repetitive work.

Start with your own idea, thesis, or objective. Then use ChatGPT to help develop it or carry it out. Give it a defined role instead of handing over the entire project.

Useful tasks to delegate include:

• Summarize or organize material
• Compare possible approaches
• Identify gaps or contradictions
• Critique an argument or draft
• Troubleshoot a problem or perform repetitive work

Review the output according to the risks involved. Check facts and citations, test code and calculations, and read edited prose closely enough to catch changes in meaning or voice. Do not assume ChatGPT’s version is better because it sounds confident. Restore details it has smoothed away, reject weak suggestions, and keep your original wording when it works better.

Be especially careful when using AI to edit prose. It can recommend inappropriate grammar changes, homogenize your voice, or alter your meaning. 

Work with AI, not for it

ChatGPT’s features will keep changing, but the basic division of labor is likely to remain useful. AI can eliminate busywork, accelerate research, suggest alternatives, and make your expertise more productive. Humans still need to decide what matters, judge whether the evidence is sound, and take responsibility for the finished work.

The goal is not to have AI do everything it can. It is to spend less time on the parts of the work that get in your way and more time on the parts that actually need you.