Capabilities and Limits

What Generative AI Can Do

Getting the most out of AI tools means understanding both what they can do and where they fall short. This page covers the capabilities and the limitations that matter most for academic and professional work.

Creating and writing

  • Drafting emails, reports, essays, and lesson plans
  • Generating outlines or first drafts for creative or analytical work
  • Checking grammar, improving tone and clarity
  • Writing or debugging programming code across multiple languages

Research and synthesis

  • Summarizing long documents or research articles into concise overviews
  • Brainstorming ideas, perspectives, and starting points for projects
  • Generating keywords for library database searches
  • Translating content between languages

Teaching and learning

  • Creating quiz questions, case studies, discussion prompts, and rubrics
  • Simplifying complex topics into plain language for students
  • Drafting accessible materials such as captions, descriptions, and simplified text
  • Demonstrating AI capabilities and limitations to students as a teaching subject in its own right

Visual and audio

  • Generating images, graphics, and illustrations from text descriptions
  • Converting text to speech for accessibility or content creation
  • Creating background music and audio content

Key Limitations

Understanding these is not meant to discourage you from using AI — it is what allows you to use it well. The most effective users are the ones who know where it goes wrong.

AI tools construct responses by predicting likely next words based on statistical patterns — not by checking facts. They can produce confident-sounding text containing completely fabricated information, nonexistent sources, or incorrect statistics. Always verify AI-generated information against reliable sources before using it in academic or professional work.

Instructors have noted that fabricated or nonexistent sources are one of the clearer indicators that a student submitted AI-generated work without reviewing it. 

NOTE: This has also been called confabulation rather than hallucination with some noting that they are related but slightly different. See Hallucination or Confabulation? Neuroanatomy as metaphor in Large Language Models.

AI tools are trained on large datasets that reflect real-world biases. They can produce outputs that reinforce stereotypes, make discriminatory assumptions, or disadvantage certain groups — including non-native English speakers and historically underrepresented communities. Critical evaluation of AI output is always necessary, especially for content used in educational or institutional contexts.

Many AI models have a knowledge cutoff date and do not know about events, research, or developments after it. Some tools can supplement this by searching the web, but the base model’s knowledge remains fixed. Always verify the currency of anything time-sensitive, and treat Maricopa-specific details — policies, deadlines, course requirements — as unverified by default.

AI tools are not moral agents. They will not decide whether using AI is appropriate for a particular assignment, professional context, or situation — that judgment is yours. Responsibility for AI-generated content, and for how it is used, always rests with the person who used the tool.

Many generative AI tools, especially free consumer ones, may retain what you submit and use it to train future models. This makes it critical to avoid entering student data, personnel records, confidential institutional information, or intellectual property. District-licensed tools carry different protections — see Privacy Considerations for the detail.

Tools that claim to detect AI-generated content measure things like predictability of word choice and variation in sentence length. These measures are imperfect, and every detection tool carries a risk of false positives.

Research has found that false positive rates are significantly higher for non-native English speakers, which makes uncritical reliance on detectors an equity problem as well as an accuracy one. Turnitin itself notes that its AI indicator is best used to inform an educator’s judgment rather than as the sole measure of academic integrity.

A detection result should be one element of a broader conversation with a student — never the only basis for a finding. See Academic Integrity and AI.

AI models are programmed to act like people-pleasers, meaning AI can sometimes agree with your ideas, assumptions, or conclusions too readily just to give you a satisfying answer. This creates a dangerous echo chamber because the AI will confidently validate your ideas, even if those ideas are factually wrong or harmful. Ask AI to identify weaknesses, challenge your assumptions, or provide another perspective, and verify important conclusions. 

References: Towards Understanding Sycophancy in Language Models, The chat-chamber effect: Trusting the AI hallucination

The Role of Human Oversight

Across the use cases faculty and staff have shared, one theme comes through consistently: AI works best as a starting point that requires human judgment, review, and refinement.

AI can generate a first draft. You bring the expertise to know whether it is accurate, appropriate, and actually useful. The combination of the tool’s speed with your judgment and domain knowledge is more effective than either on its own — and the judgment half is the part that cannot be delegated.