Eyewitness memory is reliable - people accurately remember what they see, especially under stress.

In American courtrooms through most of the twentieth century, eyewitness testimony occupied a position of special authority. A witness who had been at the scene, who had seen the defendant's face, observed the weapon, watched the crime unfold, provided evidence of a kind that seemed immune to the procedural challenges available to juries. Jurors rated eyewitness testimony more persuasive than physical evidence. Judges instructed juries to weigh it seriously. Defense attorneys had little effective counterargument because the scientific understanding of memory endorsed the courtroom's assumptions: memory was understood as a recording, imperfect perhaps, but fundamentally stable once formed.

Elizabeth Loftus, a graduate student at Stanford in the late 1960s, arrived at memory research through mathematics. She moved into cognitive psychology and began investigating whether the act of questioning a witness altered the memory being retrieved.

In 1974, Loftus and John Palmer published "Reconstruction of Automobile Destruction" in the Journal of Verbal Learning and Verbal Behavior. They showed participants film clips of traffic accidents, then asked them to estimate the speed of the vehicles using different verbs in the critical question. "About how fast were the cars going when they smashed into each other?" produced higher speed estimates than "when they hit each other." In a follow-up experiment, participants who heard "smashed" were significantly more likely to later report seeing broken glass in the footage, broken glass that was not there.

The result was striking because the manipulation was so small. A single word, not prolonged disinformation, not a suggestive lineup, not hostile interrogation, was enough to implant a detail that hadn't existed in the original footage. Memory, the data indicated, was not a recording played back. It was a reconstruction assembled anew each time from traces that could be quietly overwritten by subsequent experience, including the language used by the person asking the questions.

Loftus spent the following decades mapping the architecture of this malleability. Post-event information, a leading question, a co-witness's account, a news report, a photo spread administered in a particular way, could alter what witnesses reported seeing, with substantial confidence. Stress, contrary to the courtroom assumption that traumatic events were seared into memory with unusual clarity, could impair accuracy while increasing certainty. Witnesses who had seen a weapon in the perpetrator's hand were reliably worse at identifying the perpetrator's face, they had spent their attention elsewhere.

The criminal justice implications accumulated slowly. DNA exonerations through the Innocence Project, which began in 1992, revealed that mistaken eyewitness identification was the leading contributing factor in wrongful convictions, present in more than 70 percent of cases later overturned by DNA evidence. These were not cases where witnesses lied. They were cases where witnesses were certain, and wrong.

Reforms arrived unevenly. Some jurisdictions adopted sequential rather than simultaneous lineups, which reduce the tendency to select the person who looks most like the perpetrator relative to others. Some required that lineups be administered by officers who didn't know which participant was the suspect. Expert testimony on eyewitness reliability is now admitted in most states. But the legal system's relationship with eyewitness evidence changed far more slowly than the scientific understanding of why it so often fails.