The Age of Vibes: Vibe Coding, Lawyering, and Vibe Everything

In February 2025 the AI researcher Andrej Karpathy named something programmers had already started doing: describing what they wanted in plain English and letting an AI model write the code, rather than reading and understanding it themselves. He called it "vibe coding." A year later, a version of the same habit turned up somewhere far less forgiving of error. A recent Economist report on "vibe lawyering" describes how ordinary people, guided by AI chatbots rather than legal training, are drafting complaints, contesting disputes, and pursuing litigation they would once have needed a lawyer for. Research cited in that report, by Anand Shah at MIT and Joshua Levy at USC, examined 4.5 million federal civil cases and found that the share of self-represented litigants, flat at around 11% for two decades, climbed to 16.8% by fiscal year 2025, while the number of self-filed suits roughly doubled. The chatbots involved don’t just help people write; they tend to invent case law outright, encourage litigation, discourage settlement, and inflate people’s sense of how likely they are to win. Courts have started responding in kind: nearly a thousand reprimands have gone out over improper AI use in filings, and a federal appeals court recently suspended two lawyers over fabricated citations.

Law makes an unusually good case study, because courts keep a public, searchable record of what happens when confidence outpaces competence. But it is a case study, not the whole story. The same dynamic, a fluent AI answer standing in for judgment someone doesn’t actually have, is turning up wherever people now use AI to make decisions they used to need real expertise for. That is what this piece is actually about: not litigation specifically, but what AI does to a person’s sense of their own competence once it is quietly doing part of the thinking for them, and what that could mean as the pattern spreads well past the courtroom.

This isn’t about legal knowledge, medical knowledge, or any specific expertise. It’s about a much older feature of how people judge quality: fluency. Psychological research on the “fluency heuristic” has long shown that people use how easily and confidently information is delivered as a proxy for how true it is. An AI chatbot rarely hedges the way a careful professional does. It doesn’t volunteer that the answer depends on facts it doesn’t have, unless asked; it produces a clean, grammatically confident response, because confidence is a stylistic property of the text, not a measure of whether the underlying reasoning holds up. For someone with no independent way to check that reasoning, fluency is very hard to distinguish from real expertise. This sits close to automation bias, the well-documented tendency to defer to an automated system’s judgment even against a person’s own instincts, simply because the system feels authoritative and disinterested.

The result is a specific, asymmetric kind of overconfidence. People do not necessarily believe they know more, in the abstract, than they did before. Rather, the model’s tone transfers directly onto their sense of their own case, their own diagnosis, their own investment thesis, their own code. When the software says a plan is sound, a symptom is minor, or a trade is smart, that judgment is reached with none of the situational reasoning a real expert would apply, yet delivered with all of an expert’s certainty. The gap between how sure someone feels and how sound their actual position is doesn’t announce itself. It just widens, quietly, every time the tool is used. This is also why the effect is so hard to correct from the inside: spotting a fluency illusion normally requires exactly the expertise the illusion is standing in for. The people most exposed are rarely the ones who trust AI blindly, but the much larger group with no independent yardstick to check it against.

Law is simply the domain where the gap is easiest to catch, because a fabricated citation is a very findable kind of wrong. Elsewhere it is harder to see but no less real. Programmers doing “vibe coding” now regularly ship software whose logic they could not fully explain if asked, treating a confident AI output as good enough without necessarily being able to verify it. Researchers have started naming the broader version of this “vibe inference”: AI-assisted analysis, whether a regression, a legal argument, or a strategic recommendation, that computes cleanly and produces a confident-looking result without anything in that result indicating whether the underlying assumptions were actually sound. Doctors describe patients arriving at appointments having already reached a confident, AI-generated self-diagnosis, sometimes resistant to a contrary reading once they’ve seen a clean, authoritative-sounding answer online. Amateur investors are using AI tools to justify trades and tax positions they could not have defended unprompted a year earlier.

The same pattern is creeping into how people present themselves professionally. A well-prompted AI can produce a business plan, grant application, or job pitch that reads as fluent and expert regardless of whether the person behind it understands the substance, which makes the fluent surface a weaker signal than it used to be for anyone evaluating it, an investor, a hiring manager, a grant committee, at the exact moment those evaluators are relying on it more, not less, to manage their own workload. None of this requires anyone to be careless or unintelligent. It only requires the tool to sound sure, and the person using it to have no independent way of checking whether it’s right.

If the pattern keeps generalizing, the costs look different depending on where you’re standing. For individuals, the more people lean on a tool that never runs out of confident answers, the less practice they get building the kind of judgment that comes from getting things wrong and correcting course; skill, on this view, atrophies exactly where AI feels most useful, because the discomfort that normally forces learning gets smoothed over. For professions built around scarce, credentialed judgment, law, medicine, finance, engineering, the pressure runs the other way: when non-experts can produce expert-sounding output, institutions either verify more, which is slow and expensive, or verify less and accept a noisier, riskier system. Most are still deciding which cost they would rather absorb, and the answer may differ by field.

At a wider level, the effect compounds because these decisions interact with each other. Settlement in law, second opinions in medicine, and due diligence in finance are all quiet mechanisms that work only when both sides of a disagreement can form a roughly realistic, shared estimate of who is actually right. A tool that systematically inflates one side’s confidence erodes exactly that shared sense of realistic risk. The cost shows up gradually, as slower resolutions, disputes escalated further than they needed to go, and decisions made with unearned certainty on more than one side of the table at once. Multiply that across enough domains at once and the effect stops being local to any one profession; it becomes a general rise in unverified assertion relative to the system’s capacity to check it, in courts, in newsrooms, in markets, in ordinary workplace decisions.

There is also a generational dimension worth watching. People who build real expertise before leaning on AI have a yardstick the tool can’t easily fool. People who learn a field with AI assistance built in from the start may never develop one, and the fluency heuristic works on them just as well, with nothing internal to catch it. That argues for treating AI literacy, in the specific sense of learning to tell a confident answer apart from a correct one, as closer to a basic educational skill than an optional add-on, taught well before anyone relies on these tools for something that matters.

The responses starting to appear in law, mandatory AI-use disclosures, sanctions for fabricated citations, judges quietly recalibrating how much weight to give a suspiciously polished filing, are early examples of the kind of correction every affected field will eventually need: some way to separate fluency from accuracy before trusting either. None of this argues for reversing course. A tool that helps someone write a clear legal complaint, get a faster read on a symptom, or ship a working prototype is not nothing, and for people who were previously locked out of expert help entirely, it can be a genuine gain. But the specific failure mode is worth naming precisely, because the fixes that would actually help, AI systems that flag their own uncertainty instead of hiding it, deliberate friction at high-stakes decision points, and teaching people to treat fluency and accuracy as different things, are not the ones that get built by default. Left alone, an interface that always sounds certain will keep being trusted more than it deserves, in one field after another, right up until someone finally checks the work.

Ashton, Lydia. Vibe Econometrics and the Analysis Contract. Working paper, University of Wisconsin–Madison, May 2026. https://arxiv.org/abs/2605.08071.

The Economist. “The Rise of Vibe Lawyering.” June 29, 2026. https://www.economist.com/business/2026/06/29/the-rise-of-vibe-lawyering

Flank. “This Week in Legal AI.” Insights newsletter, June 5, 2026. https://insights.flank.ai/briefings/legal-ai-weekly-june-5-2026.html.

Hertwig, Ralph, Stefan M. Herzog, Lael J. Schooler, and Torsten Reimer. “Fluency Heuristic: A Model of How the Mind Exploits a By-Product of Information Retrieval.” Journal of Experimental Psychology: Learning, Memory, and Cognition 34, no. 5 (2008): 1191–1206.

Jean-François, Moses. “AI Is Entering the Courtroom. Could It Help You Win?” Inc.com, May 15, 2026. https://www.inc.com/moses-jeanfrancois/ai-is-entering-the-courtroom-could-it-help-you-win/91345558.

Karpathy, Andrej (@karpathy). “There’s a new kind of coding I call ‘vibe coding.’” X, February 2, 2025. https://twitter.com/karpathy/status/1886192184808149383.

Parasuraman, Raja, and Dietrich H. Manzey. “Complacency and Bias in Human Use of Automation: An Attentional Integration.” Human Factors 52, no. 3 (2010): 381–410.

Shah, Anand V., and Joshua Y. Levy. “Access to Justice in the Age of AI: Evidence from U.S. Federal Courts.” Working paper, Massachusetts Institute of Technology and University of Southern California, 2026. https://avshah1.github.io/assets/pdf/papers/pro-se/Pro_Se_Automation.pdf.