The short version
Ad hominem means attacking the person making an argument instead of engaging with the argument itself. Literally, it's Latin for "to the person." You'll see it happen anytime a claim gets met with something about the claimant, their job, their funding, their past, their character, none of it a response to what they actually said.
Here's the part most explainers skip: that's not automatically a fallacy. Whether it is depends on one question. Is the personal fact being raised actually relevant to whether you should trust the claim? Sometimes it is. A financial conflict of interest can be a legitimate reason to weigh a claim more skeptically. An insult about someone's appearance never is. Both get called "ad hominem" in casual use, but only one of them is doing anything to the argument's actual truth.
The word meant something almost backwards in its earliest recorded use
John Locke's 1690 An Essay Concerning Human Understanding is the earliest surviving source for the term "argumentum ad hominem," and Locke was careful not to claim he'd invented it. Writing about the technique of pressing an opponent with the consequences of their own stated principles or prior concessions, an argument aimed at what the other person had already committed to, not at their character, Locke says doing this "is already known under the name of argumentum ad hominem" (Book IV, Chapter XVII, Section 21). The term was already circulating; Locke's Essay is simply the oldest place anyone has found it written down. As philosopher David Hitchcock lays out in his analysis of the term's history, that non-personal-attack meaning held for roughly two centuries of logic textbooks after Locke, and there is, in his words, "not a hint" in that whole tradition that an argumentum ad hominem was a personal attack.
The shift to today's meaning happened gradually as writers on argumentation started using "ad hominem" for a family of related but distinct moves: attacking someone's character directly, questioning their motives, or pointing out their own inconsistency. By the time Canadian philosopher Douglas Walton wrote the modern academic treatment of the subject, Ad Hominem Arguments, the term covered ground Locke never intended. Knowing that history doesn't make the modern usage wrong. It's just a reminder that "ad hominem" was originally a name for a legitimate logical move, and only picked up its fallacy reputation later.
The five shapes it actually takes
Walton's taxonomy breaks the modern ad hominem into five recurring patterns, and knowing them makes it easier to spot which one you're actually looking at in a given argument.
Abusive is the bluntest version: attacking someone's character directly, calling them stupid, dishonest, or unqualified, with no attempt to connect that trait to the claim at hand. "You're wrong because you're a bad person" is the whole structure.
Circumstantial points at a person's situation rather than their character: their job, their affiliations, who benefits if their claim is accepted. "Of course the factory owner says the chemical is safe" is circumstantial.
Bias is a close cousin of circumstantial, framed specifically around an undisclosed or unacknowledged stake the person has in the outcome, rather than just their general situation. Markets run into the same structural gap under a different name: adverse selection happens when whichever side holds information the other side can't verify keeps the advantage, disclosed or not.
Poisoning the well happens before the person even speaks: discrediting them in advance so that whatever they say afterward is already suspect to the audience, regardless of its content.
Tu quoque, Latin for "you too," deflects a criticism by pointing out the critic is guilty of the same thing. It doesn't defend the original claim. It just tries to make the accusation feel hypocritical enough to ignore.
A real case where the 'attack' wasn't actually a fallacy
In May 2003, BMJ published a study by James Enstrom and Geoffrey Kabat concluding that secondhand smoke exposure showed a weaker link to tobacco-related mortality than widely believed, based on 39 years of data from the American Cancer Society's original Cancer Prevention Study cohort. The paper's own disclosure section noted the authors had received funding connected to the tobacco industry, including money that traced back to the Center for Indoor Air Research, a body largely funded by tobacco companies.
More than 140 readers wrote in with rapid-response letters, according to BMJ's own summary published that August, and a 2005 academic analysis by Sheldon Ungar and Dennis Bray in Public Understanding of Science found many of those responses moved quickly past the study's methodology and onto the authors' funding and motives. Whether that response counts as an unfair ad hominem attack or a legitimate scientific objection turns on exactly the question this piece keeps coming back to: is the personal fact relevant? A funding source that could bias a researcher toward a specific conclusion is relevant information for judging how much to trust that conclusion. That's a circumstantial or bias-type ad hominem in Walton's terms, and by the relevance test, it isn't a fallacy at all. It's evidence, delivered in the shape of a personal attack.
The research: which ad hominem attacks actually change minds
Whether an attack is technically fallacious is a separate question from whether it works, and a 2018 PLOS ONE study by Ralph Barnes and colleagues tested the second question directly. Across two experiments, 439 college students and, in a replication, 199 adults recruited through a survey panel, read a series of science claims attributed to a researcher ("Dr. [Name]"), each followed by one of six kinds of pushback: a direct attack on the empirical data, an ad hominem attack on the researcher's past misconduct, an ad hominem attack alleging a conflict of interest, an ad hominem attack on the researcher's weak educational credentials, an ad hominem attack calling the researcher a sloppy scientist, or both a data attack and an ad hominem attack combined. Participants rated their agreement with each claim afterward, and researchers measured how far that rating shifted from a baseline reading of the same claim with no pushback attached.
The results didn't sort neatly into "data attacks work, ad hominem attacks don't." They sorted by relevance instead. Attacks on competence, weak credentials or a sloppy reputation, produced no statistically significant shift in belief in either experiment. People apparently didn't treat "this person went to a low-ranked school" as a real reason to distrust a specific claim. But attacks alleging past misconduct or a conflict of interest shifted attitudes by an amount statistically indistinguishable from attacking the underlying data directly, a pattern that held in the first experiment, F(5,2190)=25.4, p<.001, η²=.06, and replicated in the second with a more representative adult sample, F(5,990)=14.62, p<.001, η²=.07. An ad hominem attack aimed at a researcher's integrity moved people exactly as much as showing them the data was flawed.
That maps cleanly onto the relevance test from Walton's taxonomy, even though the study wasn't designed around Walton's categories. Competence attacks, sloppy work, a weak degree, are only loosely connected to whether a specific claim is true, and people's attitudes barely moved. Integrity attacks, hidden funding, a history of faking results, speak directly to whether the claim can be trusted at all, and people updated their beliefs as much as they would from seeing the actual evidence challenged. The attacks that worked were, by this measure, the ones that were actually relevant.
What this means when someone attacks the person, not the argument
The practical habit worth building isn't "ignore any argument that gets personal." It's asking one narrower question: does the personal fact being raised bear on whether this specific claim can be trusted, or is it noise? A disclosed funding source, a documented history of fabricating results, a direct financial stake in the outcome, these are the kind of facts the ethos, pathos, logos framework already treats as part of a speaker's credibility, which is exactly why attacking them lands. An insult about someone's education, their appearance, or an unrelated flaw in their life is a different animal entirely: technically about the person, but disconnected from the claim, which is the abusive and poisoning-the-well end of Walton's list.
Ad hominem shares a family resemblance with the equivocation fallacy, even though they attack different targets. Equivocation lets a word quietly swap meanings mid-argument so a conclusion that only follows under one definition gets presented as if it followed under all of them. Ad hominem substitutes a fact about the arguer for engagement with the argument itself. Both work by getting a listener to accept a substitution they'd reject if it were pointed out directly.
Online arguments blur this on purpose. A pile-on that starts as "here's their undisclosed conflict of interest" and drifts into mockery of their credentials or character has quietly swapped a relevant objection for an irrelevant one, and it's easy to keep nodding along without noticing the swap happened, the same slide that shows up in how one person's low-stakes trolling scales into rage bait once it starts working. Once you've decided to distrust someone, the same self-justifying pull described in why small commitments harden into strong beliefs makes it easy to keep finding more reasons after the fact, whether or not those later reasons are actually relevant to the original claim. The question worth asking before your own opinion shifts isn't whether an attack feels personal. It's whether the personal fact it's pointing at would actually change your confidence in the claim if you set the emotional charge of the attack aside.