Critical Thinking in the Age of AI: Practical Tips for Academics

AI’s biggest risk to academic critical thinking isn’t that it produces answers—it’s that it can quietly replace the human steps that turn information into defensible judgment. The practical takeaway is to treat AI as a research assistant that accelerates work you still must verify, rather than a shortcut that hands over reasoning.

Samuel Leslie frames critical thinking as a sequence: gather information, evaluate evidence, question biases, and then reason to a final judgment. That process matters because academia isn’t just about what conclusions people reach; it’s about how they get there. In the age of AI, the danger emerges when researchers outsource too much of that pipeline—especially when they ask for ready-made text without checking whether claims are accurate, sourced, or aligned with their own questions.

Audience polling during the session mirrors that concern. Most attendees use AI for editing and rewriting, while the most common challenge is trusting AI outputs—often due to hallucinations or errors. Staying original and avoiding AI-detector flagging also surfaced, reinforcing the session’s central theme: speed is not the same as rigor.

Leslie then contrasts “outsourcing” with “assisted” use through concrete prompting examples. For gathering information, asking AI to “write an introduction paragraph” can yield fluent but unverifiable content. A better approach is to request key findings from recent studies and require citations, then restrict citations to peer-reviewed research and follow every link. He emphasizes that links and DOIs can be fabricated; one example included a DOI that did not exist, underscoring the need to click through rather than trust formatting.

When evaluating evidence, the session warns against leading prompts like “name studies that show AI improves peer review,” which can bake in a conclusion. Instead, prompts should ask for studies that include links/DOIs and explicitly request differing viewpoints (pro and cautionary). The goal is to force evaluation of the evidence landscape rather than accept a single narrative.

Bias is treated as a multi-layer problem: human biases (confirmation, prestige, group echo chambers) and AI biases inherited from training data. AI can also create an illusion of confidence through polished summaries. Leslie introduces “alignment faking” from research on large language models—where outputs may conform to what the system expects or what a user seems to want—making skepticism essential.

The session’s “last mile” is reasoning independently. AI can help collect sources, summarize, and even surface opposing perspectives, but connecting the dots, defending claims, and deciding what to believe remains a human responsibility.

For writing, the session draws a bright line: never jump straight from an idea to a full manuscript drafted by AI, because that bypasses literature review, critique, and argument-building. Instead, AI should support each stage—literature search, citation organization, project design, and drafting—while the researcher retains control of critique and voice.

A final framework is offered: never stop learning, evaluate with skepticism, and reflect on whether AI strengthened understanding or drowned out the researcher’s own voice. In Q&A, strategies for originality include recording personal voice notes, transcribing handwritten or spoken notes, and reading rewritten text aloud to check whether it still sounds like the author. On AI detection, the guidance is blunt: bypassing detectors shouldn’t be the objective; transparency and evidence of real work matter more than scores, which can produce false positives.