Ask people what convicts the innocent and they will say corrupt police or bad lawyers. The actual leading contributor is more disturbing, because nobody involved is doing anything wrong.
According to the Innocence Project’s analysis of DNA exonerations in the United States, mistaken eyewitness identification played a role in roughly seven out of ten cases. Not lying witnesses. Sincere ones, frequently very confident, describing an experience they genuinely had.
Memory Is Not a Recording
The intuition almost everyone holds is that remembering is playback: the event went in, and recall retrieves the file. Decades of research say it is reconstruction. Each time you recall something you rebuild it from fragments, and the version you rebuild is what gets stored again.
Elizabeth Loftus spent a career demonstrating how easily that rebuild can be nudged. In her best known work, people who watched the same footage of a car accident gave higher speed estimates when asked how fast the cars were going when they smashed rather than when they hit, and were more likely, a week later, to remember broken glass that was never there.
One verb, days earlier, added glass to a memory.
The Conditions Are Terrible to Begin With
Before anything is contaminated, encoding is usually bad:
- Weapon focus. If there is a gun or a knife, attention goes to it. Witnesses can describe the weapon in fine detail and the face barely at all.
- Stress. Extreme fear does not sharpen memory for detail the way films suggest.
- Duration and light. Real incidents are short and often badly lit, and people consistently overestimate how long they lasted.
- Cross-race identification. Accuracy is measurably lower when the witness and the person are of different races, a robust, uncomfortable, repeatedly replicated finding.
Then the Procedure Makes It Worse
This is the part that turns a shaky memory into courtroom certainty.
- Unconscious transference. A face that is familiar for some unrelated reason (a bystander, someone from the neighbourhood, a face seen earlier in a photo array) gets recognised as familiar, and the mind supplies a reason. Familiarity is remembered accurately. Its source is not.
- Relative judgement. Shown six people at once, a witness tends to pick whoever looks most like their memory, rather than deciding whether anyone actually is. If the real person is absent, that produces an identification anyway.
- Confirming feedback. This is the one that does the damage. A witness who picks someone tentatively and hears “good, that’s our suspect” does not merely feel reassured. Their remembered confidence rises, along with their recollection of how good a view they had. By trial they testify, honestly, that they were sure all along.
A jury then does the reasonable thing and treats confidence as evidence of accuracy. Under clean conditions, at the moment of a first identification, confidence does carry real information. What it cannot survive is feedback, and by the time a witness reaches a courtroom, months of feedback have happened.
What Actually Fixes It
These are not exotic. A 2014 National Academy of Sciences report set most of them out, and many jurisdictions have adopted them:
- Double-blind administration. The officer running the lineup does not know which person is the suspect, so they cannot cue anything, even accidentally.
- Proper instructions. Tell the witness explicitly that the person may not be present, which removes the pressure to pick the closest match.
- Fair fillers. Everyone in the lineup should match the witness’s original description. If one person alone fits it, the lineup is a pointing exercise.
- An immediate confidence statement, recorded in the witness’s own words before anyone says anything at all.
- Record the whole procedure, so a jury can see it rather than hear it reconstructed a year later.
The Case That Made This Concrete
Jennifer Thompson was raped in 1984 and deliberately studied her attacker’s face during the assault so she could identify him. She picked Ronald Cotton, testified against him, and was certain. He served eleven years before DNA identified another man who had been in the same prison, and who Cotton had encountered there.
Thompson and Cotton later became friends and wrote Picking Cotton together. She had done everything a good witness is supposed to do. More essential reading here.
Why I Went Down This Hole
Because the novel I am writing depends on it. In The Last Bites, a chef becomes convinced that a man on death row could not have committed the murder the way the state described it, and the thread she pulls is not a witness at all, but a physical detail in a crime scene photograph that everybody looked at and nobody saw.
The problem she runs into is the one this whole field runs into: it is far easier to demonstrate that a story is wrong than to get anyone to look at it again. The first three chapters are free.
A Record That Does Not Rewrite Itself
Everything below is about the same uncomfortable fact: your memory is not a recording. It is a reconstruction, and it edits itself a little every time you run it.
The Last 5 Years is the boring corrective. It is a journaling app that never asks you to write anything. Once a week it opens your own photo library, step count, weather and music, and replays this exact week from one, two, three, four and five years ago.
It is not more meaningful than your memory. It is just less creative. And because it reads everything on your device and keeps it in your own iCloud, with no account and no server, the record stays yours.
The Book This Site Is Attached To
The Last Bites is a thriller I’m writing now. A chef named Ella, whose career came apart a year before the restaurant she built won its star, gets a call from a warden: a man on death row has requested a last meal, and the dish he named is hers. She has five days, a kitchen that can’t make it, and no idea how he seems to know so much about her while she knows nothing about him.
The first three chapters are free, and they open the moment you sign up.
Or read the premise first.