Cabo da Roca, the westernmost point of continental Europe, Portugal. (Photo by Ana-Paula Correia)
Artificial intelligence is rapidly becoming part of the infrastructure of research and education. Researchers can use AI tools to search and synthesize literature, analyze data, write code, generate visualizations, refine manuscripts, and support administrative work. Educators can use them to develop instructional materials, generate examples, provide feedback, and personalize learning activities. Increasingly, AI agents promise to go further by carrying out sequences of tasks with less direct human intervention.
There is much to welcome in these developments. There is also reason to ask a more difficult question:
At what point does assistance become dependence?
This question matters because the most consequential effect of AI may not be what these systems can do. It may be what humans gradually stop doing when AI can do it for us. This essay reflects on this question.
From Prompt Design to Human Agency
In a previous blog, The Case for Prompt Design when Interacting with Generative AI, Sean Hickey and I (2025) argued that the quality of human-AI interaction depends substantially on what the human brings to the exchange. A thoughtfully designed prompt introduces context, purpose, assumptions, subject matter expertise, audience awareness, and human creativity.
Prompt design, in this sense, is not simply a technique for obtaining better responses. It is an intellectual activity.
The human determines what question is worth asking, what information matters, what assumptions need to be made explicit, and what constitutes a useful response. The AI system generates an output within the parameters established through that interaction.
This distinction becomes even more important as we move from generative AI toward agentic systems. With a chatbot, we typically ask a question and receive a response. With an AI agent, we may increasingly delegate a goal and allow the system to determine some of the intermediate steps required to achieve it.
The shift is subtle but consequential.
We are moving from asking AI to assist with a task toward asking AI to perform parts of the intellectual workflow itself.
That transition requires us to think carefully about what should be delegated and what should remain fundamentally human.
The Productivity Promise Is Real
The case for using AI in research is compelling.
Madanchian and Taherdoost (2025) examine the impact of AI on research efficiency across activities including literature discovery, data analysis, experimental design, manuscript writing, and collaboration. Their review highlights the potential for AI to automate time-intensive activities, summarize information, identify relevant publications, and support researchers in making decisions.
For research centers and universities facing increasing demands to produce high-quality scholarship, compete for funding, collaborate internationally, and communicate findings broadly, these efficiencies matter.
AI can help a researcher explore an unfamiliar literature more quickly. It can assist with coding or debugging. It can suggest alternative ways of organizing information. It can help a research team generate an initial summary of a large collection of documents. It can support brainstorming before a grant proposal is written.
Zack Batsaikhan's (2026, September 14) AI Tools for Research: A Practical Inventory offers a useful companion to this discussion by illustrating the expanding range of AI-enabled tools available across the research workflow.
The question, therefore, is not whether researchers should use AI. They already are.
The more important question is:
How they should use it without surrendering the intellectual processes that make research rigorous, original, and trustworthy.
The Risk of Cognitive Offloading
There is a useful concept from cognitive psychology for understanding this problem: cognitive offloading.
Humans have always used external tools to reduce cognitive demands. We write notes so that we do not have to remember everything. We use calculators to perform math calculations. Search engines allow us to retrieve information rather than memorize it.
Cognitive offloading is not inherently problematic. In many circumstances, it is precisely what allows us to devote our limited cognitive resources to more sophisticated problems.
AI, however, introduces a different scale of offloading.
A calculator can perform an equation, but it does not ordinarily formulate the research question, summarize the literature, interpret the results, draft the discussion section, anticipate reviewer comments, and rewrite the manuscript. Generative AI potentially participates in all of these activities.
This raises an important distinction between offloading a task and offloading thinking.
For example, Lee et al. (2025) surveyed 319 knowledge workers who provided 936 examples of how they used generative AI in their professional activities. The researchers found that people who placed more trust in generative AI tended to engage in less critical thinking, while those who were more confident in their own abilities engaged in more critical thinking. However, AI did not eliminate critical thinking. Instead, it changed how people used it, shifting their attention toward evaluating the AI-generated outputs, combining AI responses with their own knowledge, and overseeing the overall task.
Another example comes from Barcaui’s (2025) randomized controlled trial examining university students’ use of ChatGPT in a learning task. The study involved 120 undergraduate students randomly assigned to either a ChatGPT-assisted learning condition or a traditional learning condition. When students later completed a knowledge-retention assessment without access to ChatGPT, those who had learned with ChatGPT demonstrated lower retention than students in the traditional condition.
The finding should not be interpreted as evidence that using ChatGPT inevitably reduces learning. It does, however, raise an important question about the relationship between cognitive effort and durable learning. When an AI system makes it easier to obtain explanations, organize information, or formulate answers, learners may complete a task successfully while engaging less deeply with some of the cognitive processes involved in constructing and retaining knowledge.
This distinction has important implications for education. Successful task completion is not necessarily evidence of successful learning.
Together, these studies suggest that the central question is not whether AI makes us think or not think. It is what kind of thinking remains with the human, and what happens to learning when some of that thinking is transferred to the machine.
From Smartphone Dependence to AI Dependence?
This question reminds me of another technological phenomenon I have studied: nomophobia. In my previous discussion of smartphone dependence, The Ever-Present Companion: Understanding Nomophobia and Smartphone Dependence, Batsaikhan and I (2026) examined the anxiety associated with losing access to a device that has become deeply embedded in everyday life. Smartphones began as tools. Over time, they became constant companions, gateways to communication, information, navigation, entertainment, and social connection.
Eventually, for many users, the device became difficult to separate from the experience of daily life.
AI may be following an analogous trajectory, although the nature of the dependence is different.
Imagine beginning every research task by asking an AI system what to do.
Which articles should I read? What research question should I pursue? What theoretical framework should I use? How should I interpret these findings? What should I write next?
None of these questions is inappropriate on its own. AI can provide useful perspectives on each. The concern emerges when consultation becomes default behavior and, eventually, dependency.
If researchers repeatedly turn to AI before engaging their own disciplinary knowledge, curiosity, judgment, or creativity, we may begin to outsource something much more important than productivity: the formation of intellectual judgment itself.
AI Does Not Know What Matters
This concern becomes particularly important when we remember what current AI systems actually are.
In my reflections on the AI, Science, and Society: Insights from Global Experts 2025 conference in Paris, I discussed Eric Xing's observation that large language models (LLMs) fundamentally operate through prediction rather than genuine human-like understanding. Michael Jordan similarly challenged visions of an all-knowing artificial general intelligence by emphasizing the importance of individual context, needs, and preferences.
This distinction is easy to forget because LLM outputs can sound remarkably authoritative. Fluency can create an illusion of understanding.
But producing a persuasive explanation is not the same as understanding why a research question matters to a community. Generating a methodological recommendation is not the same as accepting responsibility for the consequences of that methodological choice. Summarizing a body of scholarship is not the same as participating in the disciplinary history and debates that give that scholarship meaning.
AI systems can generate plausible possibilities. Researchers must determine which possibilities are meaningful. That responsibility cannot be delegated.
Preserving the Human Experience
This also connects with a question I raised previously about the user-first mindset in AI systems in my 2025 essay, Human Experience: Are We Losing the User-First Mindset as AI Systems Assume Greater Roles in Our Daily Lives?
Efficiency is valuable, but efficiency cannot be the sole measure of successful technology.
In education, the fastest route to an answer is not necessarily the best route to learning. Struggle, uncertainty, revision, discussion, and error are often part of how knowledge develops and is constructed.
The same is true of research.
Reading a difficult article slowly may reveal a conceptual distinction that a summary misses. Wrestling with contradictory findings may generate a new research question. Writing an imperfect first draft may help a scholar discover what they actually think.
If AI removes too much of this space for discovery, it may also diminish the serendipitous encounters, unexpected insights, and moments of wonder through which expertise develops.
The goal, therefore, should not be scholarship stripped of opportunities for discovery, reflection, and wonder. It should be meaningful human-AI interaction, where AI supports the scholarly process while human expertise, judgment, and curiosity remain at its center.
Assistance Without Dependence
The future of research will almost certainly involve more AI, not less. Agentic systems will become more capable. Research workflows will become increasingly automated. Activities that currently require several disconnected tools may eventually be coordinated by a single AI agent.
As AI makes more aspects of scholarly work easier and faster, how do we preserve our appreciation for what humans have collectively discovered, created, and learned?
Perhaps the way forward is to remain curious about what AI can offer while keeping our inquisitive and critical eyes wide open. AI outputs should invite further questioning, not mark the end of it. Our curiosity, judgment, and capacity for wonder remain essential to how knowledge advances.
The measure of a successful human-AI research ecosystem should not simply be how much work the machine can perform. It should also be whether the human researcher becomes more capable of asking important questions, evaluating evidence, making connections, exercising judgment, and generating new knowledge.
Assistance becomes dependence when we can no longer perform, explain, evaluate, or take responsibility for the intellectual work we have delegated.
The challenge ahead is therefore not to keep AI out of research and education. It is to build practices, cultures, and technologies in which AI expands human capability without quietly diminishing human agency.
A final note to my readers
I write these essays as a form of independent public scholarship, developed outside my university responsibilities and grounded in my professional experience, research, and ongoing reflection.
If you find these essays useful or thought-provoking, you can support this work by subscribing to my LinkedIn newsletter, The Nautilus Navigator, sharing it with others, or Buying Me a Coffee.
Support is always optional. Reading, sharing, and engaging with these ideas are already meaningful ways of sustaining the conversation.
References
Barcaui, A. (2025). ChatGPT as a cognitive crutch: Evidence from a randomized controlled trial on knowledge retention. Social Sciences & Humanities Open, 12, 102287. https://doi.org/10.1016/j.ssaho.2025.102287
Lee, H.-P. (Hank), Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R., & Wilson, N. (2025). The impact of generative AI on critical thinking: Self-reported reductions in cognitive effort and confidence effects from a survey of knowledge workers. Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Article 1121, 1–22. https://doi.org/10.1145/3706598.3713778
Madanchian, M., & Taherdoost, H. (2025). The impact of artificial intelligence on research efficiency. Results in Engineering, 26, 104743. https://doi.org/10.1016/j.rineng.2025.104743
Please cite the content of this blog
Correia, A.-P. (2026, September 18). When Assistance Becomes Dependence: The Risks of Over-Reliance on AI in Research and Education [Blog post]. Ana-Paula Correia’s Blog. https://www.ana-paulacorreia.com/blog/when-assistance-becomes-dependence-the-risks-of-over-reliance-on-ai-in-research-and-education