This is Part 2 of a four-part series from Tanya Thomas. Part 1 examined how we’re training a generation of lawyers who rarely engage with the raw materials of their profession, and are increasingly consuming only processed, pre-digested, AI-synthesized versions like the mechanically separated chicken parts that go into chicken nuggets.

Part 2

The Cognitive Architecture of Dependency

What makes this crisis particularly insidious is that some research indicate that it may create a reinforcing loop of low self-efficacy and high dependence on generative AI. This may show up in legal research where students who feel uncertain, which is most students initially because legal research is genuinely difficult, increasingly turn to generative AI (GAI) tools for quick answers. These tools respond with fluent, confident explanations that seem authoritative and comprehensive. The student’s uncertainty dissipates. They feel like they’ve found the answer. Their cognitive load decreases. The experience is satisfying.

But something crucial has been lost. The student hasn’t developed any research skills. They haven’t learned how to formulate effective search queries, how to evaluate results, how to read cases for holding versus dicta, how to trace precedent forward and backward, how to recognize when their research is incomplete. They’ve learned only that asking the GAI produces answers. Over time, this dependency deepens. The skills they didn’t develop would have built genuine confidence, the confidence that comes from knowing how to find and evaluate legal authority. Instead, they have false confidence in the GAI’s output and diminishing confidence in their own abilities. This makes them more likely to rely on GAI for the next research task, which further atrophies their underdeveloped skills, which makes them even more dependent. The confidence-competence death spiral tightens with each iteration.

Legal educators see this manifest in disturbing ways. Students submit briefs citing cases they’ve never read. When questioned about their reasoning, they can’t explain the connection between their cited authorities and their arguments because they don’t understand the cases, they only know what the AI told them about the cases. They become defensive when told their research is insufficient, genuinely confused about why the GAI’s answer isn’t good enough. They can’t perform well in oral arguments because they don’t know the cases deeply enough to answer questions about facts, procedural posture, or alternative interpretations. They struggle with the Socratic method because they haven’t read the assigned opinions, just GAI summaries of them.

The Amplification of Cognitive Bias

Generative AI doesn’t merely enable poor research habits; it actively amplifies the cognitive biases that undermine critical thinking. Consider automation bias, the tendency to trust computer-generated outputs over human judgment. In traditional legal research, automation bias might lead a student to trust Wexberg (shorthand for Westlaw, Lexis, and Bloomberg Law collectively) relevance ranking too heavily. But Wexberg’s non-GAI interface shows you the cases it ranked. You can see the list, evaluate the results, decide whether the algorithm got it right. GAI automation bias is far more dangerous because the AI’s selection process is completely opaque. When it tells you what the key case is, you have no way of knowing what other cases it considered and rejected, or more troublingly, what important cases it never encountered at all because of gaps in its training data.

Authority bias becomes particularly problematic when prestigious platforms add generative AI features. Students who have learned that Wexberg i[JC1] s the gold standard for legal research automatically extend that trust to their generative AI tools. But those generative AI tools have a completely different search mechanism. The century of editorial oversight, the careful citation verification, the comprehensive coverage may not apply to the GAI’s output. Yet students don’t make this distinction. They think “this came from Wexberg” and assume it carries Wexberg’s traditional reliability. The platform’s legitimate authority creates a halo effect around a fundamentally different tool.

Anchoring bias, always a problem in research, becomes acute with GAI because of linguistic fluency. GAI generates consistently fluent prose. Every answer is well-written, clearly organized, confident in tone. This fluency creates a false sense of accuracy. Students anchor to the first GAI response they receive because it seems so authoritative and accurate. They don’t recognize that fluent writing and accurate analysis are entirely separate qualities. A beautifully written wrong answer is more persuasive than an awkward but correct one.

Perhaps most dangerous is what we might call the illusion of comprehensiveness. When a GAI tool cites six cases in proper Bluebook format, students see what looks like thorough research. They don’t understand that those six cases might represent a tiny fraction of relevant authority, that the GAI might have missed landmark decisions, that some of the cited cases might not even exist. The phenomenon of GAI hallucinating plausible sounding, but entirely fictional case citations reveal the depth of this problem. Students have started citing cases that have never been decided by any court because the GAI invented them, complete with realistic case names, citation formats, and even fictional holdings. When caught, students are genuinely shocked. They thought they were citing real cases. The AI told them these cases existed. How were they supposed to know otherwise?

The Source Blindness That Preceded GAI

To understand why students are so vulnerable to GAI, we need to recognize that source blindness predated generative AI. Students had already lost the ability to distinguish between different types of information sources before GAI made sources disappear entirely. This happened because digital interfaces flattened the material differences that once made source hierarchies obvious.

When everything appears as a PDF in a browser tab, a Supreme Court opinion looks identical to a law review article, which looks identical to a blog post, which looks identical to a think tank white paper. The same fonts, the same scrolling interface, the same ability to search and copy text. All the material cues that once distinguished these sources have vanished. You used to know you were reading a Supreme Court opinion because it appeared in a specific bound reporter with distinctive physical characteristics. You knew you were reading a law review article because it was published in a law journal with specific formatting conventions. You knew you were reading a magazine because it felt like a magazine with glossy pages, advertisements, different paper quality.

These material differences weren’t just aesthetic. They encoded information about the information source. The physical form communicated authority, purpose, and institutional backing. Students absorbed this information without conscious effort through repeated physical interaction with different types of sources. Digital interfaces eliminated this ambient learning. Everything became “content,” undifferentiated and equivalent. Information sources don’t just have different “levels” of quality on a single spectrum; they exist in different categories of knowledge production serving different functions. But students don’t see these categories. They see text on screens. When asked to evaluate sources, they apply general heuristics about credibility, i.e. “does this seem legitimate?” rather than understanding the specific role each type of source plays.

Source Erasure: When Search Engines Become Answer Engines

The transformation of Google from search engine to answer engine represents more than just an interface design shift, it’s a fundamental shift in how students encounter information. A recent study from Botify noted that 60% of Google searches now end without users clicking through to any website. This means that more than half the time, users are getting their information directly from Google’s AI-generated response and never look at an actual source. On mobile devices, which are the primary research tools for most students, organic search results have been pushed so far down the page that they effectively disappear beneath GAI overviews and sponsored content.

Legal research instructors have traditionally taught students to evaluate search results carefully, look at the URL, consider the source’s authority, check publication dates, read critically. However, these lessons become irrelevant when students never see search results, all they see are answers and answers don’t come with visible evaluation criteria. There’s no single URL to evaluate when the information is synthesized from multiple sources. There’s no clear publication date when the GAI draws from materials spanning years or decades. There’s no obvious source authority when the answer is an amalgamation.

This shift has turned research from an active process into passive consumption. Research used to mean finding sources, evaluating them, synthesizing insights across multiple authorities, and reaching conclusions based on that synthesis. Now it means asking questions and accepting answers. Students have become consumers of information rather than investigators of it. They don’t develop the iterative thinking that characterizes skilled research—trying a search, evaluating results, refining the query, following unexpected leads, discovering connections, recognizing gaps, circling back to fill them. They simply ask and receive.

The Particular Vulnerability of Legal Research

This crisis is particularly acute in legal research because legal authority is irreducibly source specific. Unlike many disciplines where synthesis is acceptable or even preferable, law requires engagement with exact textual sources. You cannot understand the holding of a case by reading a summary of it. You must read what the court said, because courts interpret and apply the specific language of statutes and regulations and other case. When a court cites a case, they cite specific passages, reasoning, limiting language, little to none of this survives summary.

Legal reasoning operates through a strict hierarchy of authority that GAI often flattens and obscures. GAI presents all sources as roughly equivalent, synthesizing across them without regard for their relative authority. A student using GAI might receive an answer that blends constitutional text, statutory interpretation, regulatory guidance, and law review commentary without any indication of which sources are binding authority, and which are merely potentially persuasive.

The precision of legal language creates another vulnerability. In law, exact wording matters. The difference between “reasonable expectation of privacy” and “legitimate expectation of privacy” isn’t stylistic variation, it reflects distinct doctrinal tests with different elements and applications. The difference between “shall” and “may” in a statute determines whether something is mandatory or discretionary. AI paraphrases, it summarizes in different words smoothing over linguistic precision in pursuit of clarity. This makes AI-generated legal analysis structurally unreliable even when it’s trying to be accurate, because accuracy in law requires textual fidelity that GAI’s paraphrasing nature undermines.

Temporal dynamics add another layer of vulnerability. Law changes, cases get overruled, statutes get amended, constitutional interpretations evolve. Understanding current law requires knowing not just what courts have said but when they said it and what’s happened since. GAI’s training data has a cutoff date that students don’t know and often don’t think to ask about. Even when the cutoff date is disclosed, students rarely understand what it means. If the GAI was trained on data through January 2024, does it know about cases decided in February 2024? Does it know that a key precedent was overruled in March 2024? Does it reflect statutory amendments that took effect in April 2024?

Part 3 will look at the professional stakes of this crisis and what our pedagogical response should be.

Editor’s Note – This article republished with permission of the author as well as the Editor of RIPS Law Librarian, first publication.