Generative AI arrived in academia faster than the norms to govern it. Large language models (LLMs) like ChatGPT, Claude, Gemini, and Perplexity moved from novelty to daily instrument in under three years, and the institutions meant to guide their use are still catching up.

Here's the tension in one sentence: students and researchers have near-universal access to AI, and almost no shared literacy about how to use it well.

The numbers make the gap concrete. In the Digital Education Council's 2024 Global AI Student Survey (which polled roughly 3,800 students across 16 countries), 86% of students reported using AI in their studies, and 54% use it weekly, with ChatGPT the dominant tool at 66% adoption. Yet in that same survey, only a small minority felt their institution had prepared them: a majority (58%) said they did not feel they had sufficient AI knowledge and skills, and just under half (48%) said they did not feel adequately prepared for an AI-enabled workplace. Most strikingly, around 80% of students said their university's integration of AI did not fully meet their expectations.

The gap isn't access. It's literacy.

So what is AI literacy for research?

Strip away the hype and AI literacy in an academic context comes down to four competencies:

The literature review is where AI can save the most time and cause the most damage. Done well, it compresses weeks of searching and screening. Done blindly, it produces a review built on papers that were summarized inaccurately, or don't exist.

Break the workflow into four stages and match the tool to each.

Stage 1. Discovery: finding the right papers

This is where you're searching for sources, and where general chatbots are most dangerous, because a model asked to “list key papers on X” will happily invent citations that look perfectly formatted. Use tools that search a real index instead: Semantic Scholar, OpenAlex, arXiv, and discovery tools built on them. The right question to ask a discovery tool isn't “what are the important papers” but “search these databases and return what actually exists, with links.”

Two habits make discovery work when you don't know the right keyword. First, describe the question, not the term. Research-grade tools search on meaning, so “studies that measured whether students can tell AI-written text from human writing” finds work that “AI detection” misses. Second, follow the chain. Every source you find has two trails: backward, through its reference list, to the work it built on; forward, through “cited by” in Semantic Scholar, OpenAlex, or Scite, to the work that built on it. Citation counts tell you what the field considers significant. The forward trail tells you where the conversation went next. Chase both before you conclude you have found the literature.

Stage 2. Screening: is this paper relevant?

Once you have candidates, AI can triage, scanning abstracts against your inclusion criteria far faster than you can. Tools like Rayyan are built specifically for systematic-review screening. The literacy skill here is writing explicit inclusion criteria, because a vague screen produces a vague corpus.

Quietly harmful

The integrity line, stated plainly

Use AI to help you express your thinking. Don't use it to replace your thinking. Editing the words is usually fine; manufacturing the substance is not. When in doubt, ask whether you could defend every claim and every source in front of your committee. If yes, you're using it as a tool. If no, you're using it as a ghostwriter, and ghostwriting fails the integrity test regardless of whether the ghost is human or machine.

We'll get to formal disclosure rules in Chapter 8. For now, the working rule: the ideas, the argument, and the sources must be genuinely yours and genuinely verified.

The verification workflow (do this every time)

For every AI-suggested citation:

The tools that eliminate most of this pain are the ones that start from the real source record rather than generating references from memory. When a tool verifies each citation against CrossRef or a live index (the way CoChat's citation verification does), steps 1 and 2 are handled by design, and you're left only with the irreducibly human step 3: does this source actually support my point?

Match the tool to the stage

The takeaway isn't "use tool X." It's match the tool to the stage: discovery tools for finding, summarizers for orienting, and a verification-capable workspace for the citation and synthesis stages where accuracy is non-negotiable.

Here's the discipline that separates AI-assisted analysis from AI-corrupted analysis. That means:

Used this way, AI is a powerful analysis accelerator. Used carelessly, it's a fast path to results you can't stand behind.

A decision framework for any AI use

Before using AI for a research task, ask:

Pass all five and you're on solid ground. Fail any one and stop.

AI literacy isn't a certificate; it's a practice. Here's how to build it.

The 30-day plan

The three habits that matter most