Belief bias

The roses that are flowers

A logic worksheet asks whether this argument is valid: “All flowers have petals. All roses have petals. Therefore, all roses are flowers.” A student marks it valid: “Of course. Roses are flowers.”

The conclusion is true, but it doesn’t follow. The premises only say that flowers and roses both belong to the group of things with petals, which leaves room for roses to sit outside the flowers entirely. That’s an Undistributed middle. Swap in “All dogs have fur. All cats have fur. Therefore, all cats are dogs,” and the same form is plainly invalid. Believable, invalid arguments like this one are exactly where the bias is strongest.

The walking whales

“All mammals can walk. Whales are mammals. Therefore, whales can walk.” Asked whether the argument is valid, someone says: “No. Whales obviously can’t walk.”

This is the other direction: a valid argument rejected because its conclusion is false. The argument has the same form as “All mammals breathe air; whales are mammals; so whales breathe air,” and if its premises were true, its conclusion would have to be. What’s wrong is the first premise, not the logic. Saying “it’s unsound, because not all mammals can walk” is correct; saying “it’s invalid” is the error.

The proof with the right answer

Two students swap homework to check each other’s geometry proofs. One proof ends with the theorem the class already knows is true. The checker skims it, sees the familiar result on the last line, and marks it correct. A step in the middle assumes what it was supposed to show.

No syllogism is involved, but the move is the same: a known-true conclusion lowers the scrutiny given to the reasoning that led to it. A proof that ended with a false result would have been read line by line. The task was to check whether the steps work, and the truth of the final line can’t answer that. (The flawed step itself is Begging the question.)

• When the question is whether to believe the conclusion, not whether this argument proves it. Background knowledge is evidence, and ignoring it would be a mistake of its own.
• When you reject a valid argument because a premise is false. That’s judging soundness, which needs both good logic and true premises.
• When the argument isn’t meant to be deductive. For everyday arguments that claim only to make a conclusion likely, how plausible the conclusion was to begin with rightly affects how much the argument has to show.
• When a surprising conclusion prompts a closer look, as long as unsurprising conclusions get the same check. The bias lies in the difference in scrutiny, not in the checking.

The test: would you give the same verdict if the argument had the same form but a conclusion you disbelieve?

Rejecting a conclusion, not the logic

“All birds can fly. Penguins are birds. So penguins can fly.” A listener replies: “That follows, but I don’t buy it. The first premise is false, since penguins, ostriches and kiwis can’t fly.”

The listener rejects the conclusion because they know it’s false, but they don’t call the argument invalid; they grant the form and locate the false premise. That’s the soundness condition above, and it’s exactly the distinction belief bias fails to make.

Doubting an implausible study

A news story reports that a small study of twelve people found eating a spoonful of honey every morning made them twice as productive. A reader says: “I’d want to see that replicated before I believe it. It’s a huge effect from something that small.”

The reader lets the implausibility of the conclusion affect their judgment, but this is not a logic exercise. A small study is not a deductive proof, and when evidence can only make a claim more or less likely, how likely it was beforehand is part of weighing it. That’s the not deductive condition.

In everyday life, using what you know to judge a claim is usually the right thing to do. Judging validity asks for something unusual: setting aside whether statements are true and looking only at whether one follows from the others. Several accounts of how belief intrudes have been proposed. In the selective scrutiny account, people accept believable conclusions without much checking and examine unbelievable ones more critically. In mental-model accounts, an unbelievable conclusion prompts a search for a situation where the premises hold and the conclusion fails, while a believable one doesn’t. A third view, from signal detection models, holds that a believable conclusion mainly makes people more willing to say “valid” at all, without changing how well they tell valid from invalid (see Evidence).

The pattern is related to Confirmation bias: what you already believe shapes how hard you look for problems. Unlike most motivated reasoning, though, belief bias appears with conclusions no one cares about, such as whether roses are flowers.

These are standard methods from logic teaching. The research cited here didn’t test them as remedies for belief bias.

• Replace the content with letters. Rewrite “All flowers have petals” as “All A are B” and check the form on its own. If “All A are B; all C are B; so all C are A” doesn’t follow, neither does the version about roses.
• Look for a counterexample. Try to describe a situation where every premise is true and the conclusion is false. If you can, the argument is invalid, however true its conclusion is.
• Give believable conclusions the same check. The ones you agree with are the ones most likely to slip through.

Status: contested. That people accept more arguments as valid when the conclusion is believable has been found consistently since 1983, analyzed in two meta-analyses, and accepted by the researchers who most sharply criticized the older interpretation of it. What’s disputed is a second claim: that believable conclusions also make people worse at telling valid from invalid arguments. No preregistered or multi-lab replication was located, and neither meta-analysis formally corrected for publication bias. By this site’s criteria that falls short of robust, and the less confident label applies, even though the case for the basic pattern is strong: the effect is large (in the founding experiment, believability shifted acceptance by about 65 percentage points on average, validity by about 19), it appears across independent research groups, the 2018 meta-analysis included unpublished data sets, and researchers who set out to overturn the older interpretation found the effect in their own experiments.

• Evans, Barston and Pollard (1983) crossed validity with believability, so that every argument form appeared with both a believable and an unbelievable conclusion (for example, conclusions about whether some addictive things are not cigarettes, or some cigarettes are not addictive). In their first experiment, as tabulated in a later meta-analysis, valid arguments were accepted 92% of the time with a believable conclusion and 46% with an unbelievable one; invalid arguments were accepted 92% and 8%. Believability moved acceptance more than validity did, and logic made a difference mainly when the conclusion was unbelievable. That last pattern, a validity-by-belief interaction, became the standard index of how beliefs affect reasoning.
• Klauer, Musch and Naumer (2000) re-examined 22 studies with a formal model separating reasoning from response tendencies and found that the usual data were too sparse to distinguish the competing explanations. Across eight further experiments, none of the existing accounts fit, and they proposed a new one.
• Dube, Rotello and Heit (2010) argued that the interaction index assumes a particular, linear relationship between correct and incorrect acceptances. In three experiments with confidence ratings, that assumption failed. Analyzed with a model that fit their data, believability changed how readily people said “valid” but not how accurately they discriminated, which they summarized as “a response bias effect”. On this reading, the famous accuracy difference is a measurement artifact, but the bias itself is real.
• Trippas and colleagues (2018), including one of those critics, reanalyzed 22 confidence-rating data sets (993 participants), drawn mainly from two research groups. They found no overall effect of believability on discrimination, with a Bayes factor of about 7 in favor of none, in line with Dube and colleagues. But individual differences mattered: in one large study, people who showed a more analytic thinking style discriminated better with unbelievable conclusions than believable ones, while more intuitive reasoners showed no such difference. In that study, believability raised acceptance as much as validity did. The authors also cautioned that labeling the shift a pure “response bias” goes further than the models can strictly identify.

What remains uncertain. Whether, when and for whom beliefs change reasoning accuracy, rather than only the willingness to accept, is still debated, and the answer depends on modeling choices. Most of the evidence comes from categorical syllogisms in laboratory tasks, largely with student samples, and how well the size of the effect carries over to everyday arguments is less studied.