Human-First AI

Five habits for staying in charge of your thinking, judgment, and learning (by Tamara Tate & Mark Warschauer)

Generative AI has changed the economics of intellectual work. Tasks that once took hours—drafting a lesson plan, outlining an article, writing code, or summarizing a report—can now be completed in seconds. What has not become easier is deciding what is worth saying, whether the output is accurate, and how it should be used.

AI has made production easier, but it has not replaced judgment.

Over the past three years, our research team has worked with teachers and students as generative AI has become part of everyday learning. One finding has become increasingly clear: the educational value of AI depends less on what the technology produces than on how people use it. When learners remain responsible for framing problems, evaluating ideas, and making decisions, AI can support learning. When those responsibilities are delegated to the technology, opportunities for learning are often reduced.

The Human-First AI framework grew out of these observations. It identifies five habits that help keep people—not AI systems—in charge of thinking, judgment, and learning.

Although we developed the framework in educational settings, we have found that the same habits apply well beyond the classroom. They provide guidance for anyone who wants to use AI intentionally while retaining ownership of their work.

A Guiding Principle: Treat AI as an Assistant, Not a Teammate

One common metaphor describes AI as a teammate or collaborator. The analogy is understandable—these systems can generate ideas, revise drafts, and respond conversationally. But the metaphor also has limitations.

A teammate typically shares responsibility for decisions and outcomes. A generative AI system does not. It has no understanding of the task, no commitment to the outcome, and no accountability for mistakes. Current language models generate responses by predicting likely continuations of text rather than by reasoning about problems in the way people do.

For that reason, we have found it more useful to think of AI as an assistant. An assistant may contribute ideas, generate alternatives, or complete routine tasks, but the direction of the work remains with the supervising human user. People define the problem, evaluate the suggestions, and make the final decisions.

The distinction is more than semantic. The language we use shapes expectations about responsibility. If AI is viewed as an assistant rather than a teammate, it becomes clearer that responsibility for thinking and judgment remains with the person using the tool.

---

1. Think First: Your Best Ideas Shouldn’t Start as a Prompt^*

The first decision is not what to ask AI. It is whether AI belongs in the task at all.

Some activities are primarily about efficiency, and AI may provide substantial benefits. Others depend on human relationships, creativity, or deep learning, where relying too heavily on AI may undermine the very purpose of the activity. Developing AI literacy therefore begins with deciding when AI is an appropriate tool rather than assuming it should always be used.

When AI is appropriate, we encourage people to spend a few minutes thinking independently before writing their first prompt. What is the goal? What do you already know about the topic? What constraints matter?

Research on anchoring suggests that the first solution we encounter can disproportionately influence later thinking (Tversky & Kahneman, 1974). If the first substantive idea comes from AI, subsequent revisions may remain anchored to that initial response. Beginning with one’s own ideas—even if they are incomplete—helps preserve independent judgment.

This is particularly important in educational settings. Research on the generation effect demonstrates that producing ideas oneself contributes to stronger understanding and retention than simply receiving them (Slamecka & Graf, 1978).

One practical implication is to ask students to generate a brief “seed” before consulting AI—a thesis statement, an outline, a problem specification, or an initial solution. The goal is not completeness, but to establish an initial direction that AI-generated ideas can later be evaluated against.

---

2. Ask and Iterate: Prompting Is a Conversation You Lead

Effective AI use rarely happens in a single prompt. People begin with an initial request, evaluate the response, recognize what is missing, add new constraints, and refine the conversation. The process is iterative because the user—not the model—holds most of the important context.

No prompt can fully capture the realities of a classroom, a workplace, or a research project. You know your goals, your audience, your constraints, and the tradeoffs that matter. As the conversation develops, you supply that missing context by clarifying expectations, requesting alternatives, and rejecting responses that do not meet your needs.

AI can generate possibilities; people decide which ones matter.

Iteration also provides opportunities to challenge rather than simply accept AI output. Current language models tend to be highly agreeable, often reinforcing a user’s assumptions instead of questioning them. Some of the most productive prompts therefore ask AI to critique rather than simply elaborate. Questions such as “What assumptions underlie this recommendation?” or “What important perspective is missing?” encourage a more critical dialogue.

For educators, this also means helping students develop the language needed to question AI responses. Sentence stems such as “Can you explain why…?”, “What evidence supports this claim?”, or “What assumptions are you making?” help normalize productive skepticism and encourage deeper engagement with AI-generated ideas.

---

3. Take Charge: If Your Name Is on It, You Own It

One aspect of AI use that receives less attention than it deserves is responsibility.

Regardless of how much assistance AI provides, the person who chooses to use the output remains accountable for the final product. AI systems cannot assume responsibility for factual errors, fabricated citations, biased interpretations, or inappropriate recommendations. Those decisions ultimately belong to the human user.

For that reason, using AI well requires more than generating content. It requires making deliberate decisions about what to keep, what to revise, and what to discard. The transition from generated to authored occurs through those decisions.

This principle also helps guard against automation bias—the tendency to place excessive trust in recommendations simply because they come from a machine (Parasuraman & Riley, 1997). Rather than assuming an AI suggestion is correct, Human-First AI encourages users to treat every recommendation as provisional until they have evaluated it themselves.

Ownership extends beyond factual accuracy. It also includes style, voice, and purpose. Especially for multilingual writers, AI can easily produce fluent but standardized prose that obscures an individual’s own way of communicating. Human-First AI is not about resisting assistance; it is about ensuring that the final work still reflects the author’s own intentions.

One simple classroom practice is to ask students to briefly explain what they kept, what they changed, and why. Making those decisions visible often provides richer evidence of thinking than the finished product alone.

---

4. Check Your Facts: Fluency Is Not Accuracy

One of generative AI’s greatest strengths is also one of its greatest risks. It produces responses that are coherent, confident, and linguistically polished. Those qualities make information feel credible, even when it is incomplete or incorrect.

Psychologists refer to this tendency as fluency bias: information that is easier to process often feels more believable, regardless of its accuracy (Alter & Oppenheimer, 2009).

For that reason, AI output should be treated as a starting point for evaluation rather than as a final answer.

Verification takes different forms across disciplines. Historians corroborate sources. Scientists examine methods and evidence. Engineers test designs under a range of conditions. Journalists confirm facts through multiple sources. AI does not replace these disciplinary practices; if anything, it makes them more important.

Verification also involves looking beyond factual accuracy. Every AI response reflects choices about what information is included, what perspectives are emphasized, and what assumptions remain implicit. Asking whose voices are represented—and whose are absent—is part of evaluating the quality of an AI-generated response.

In this sense, verification includes both corroboration and interrogation. We check whether claims are correct, but we also consider whether the framing is complete, balanced, and appropriate for the context.

A simple classroom routine is to ask students to identify one factual claim, verify it using an independent source, and identify one perspective that is underrepresented or missing. Repeated over time, these routines help develop habits of critical evaluation that extend well beyond AI.

---

5. Reflect and Learn: Debrief the Process, Not Just the Product

The fifth habit emphasizes learning from AI use rather than simply completing the task.

After finishing an AI-assisted activity, it is worth asking a few simple questions. What aspects of the task did AI genuinely improve? Where did it introduce confusion, oversimplification, or error? What decisions did you ultimately make yourself? What did the experience reveal about your own understanding?

These questions encourage what educational researchers describe as self-regulated learning: planning, monitoring, and reflecting on one’s own thinking (Pintrich, 2000). Reflection helps transform AI from a productivity tool into a learning tool.

Without reflection, AI use can become increasingly efficient without necessarily leading to greater expertise. With reflection, each interaction becomes an opportunity to calibrate one’s judgment, refine one’s strategies, and identify areas for further learning.

Our experience working with teachers suggests that these conversations are often where the deepest learning occurs. The AI interaction itself may be brief, but discussing why certain outputs were useful—and why others were not—helps learners develop the judgment needed to use AI effectively in future situations.

Reflection need not be time-consuming. Even a few minutes at the end of an AI-assisted activity—asking what AI contributed, what students changed, and what questions remain—can promote metacognition without significantly adding to instructional time.

---

A Human-First Approach

These five habits are not intended as a checklist for every AI interaction. Rather, they describe a way of approaching AI that keeps human judgment at the center. As generative AI becomes increasingly integrated into education and professional life, the question is no longer whether people will use these tools, but how.

Our hope is that the Human-First AI framework provides a practical foundation for using AI in ways that strengthen, rather than replace, human thinking and learning. The goal is not to avoid AI or to compete with it. It is to use these technologies intentionally—to frame the problem, guide the process, evaluate the results, and ultimately remain responsible for the work.

*For the full case of why your brain already knows more than the AI, and why the "messy middle" is where the learning lives, see our dedicated post Think First.