Your Phone Injects False Memories at 2x the Normal Rate. The Surgical Version Is 500x Away.

2.05. That is the multiplier by which AI-generated video increased false recollections in a pre-registered 2025 study of 200 participants run by Pattie Maes's group at the MIT Media Lab with Elizabeth Loftus as co-author. Participants saw AI-generated videos of AI-edited images, and they later "remembered" details that never happened, at more than twice the rate of the control group, in a study that was pre-registered, meaning the researchers locked their analysis plan before seeing the data, precisely because the result was going to sound unbelievable. Their confidence in those false memories ran 1.19x higher too. No surgery, no drugs, just pixels arranged by a model.

Meanwhile, the version of memory injection that looks like science fiction still runs on mice and blue light. In 2013, Susumu Tonegawa's lab at MIT did the famous experiment: they genetically engineered mice so that neurons active during memory formation would grow a light-sensitive protein, tagged the neurons firing while a mouse explored one chamber, then reactivated those neurons with an optical fiber while shocking the mouse in a different chamber. In the first chamber, the mouse froze, terrified of a shock it never received there. "Whether it's a false or genuine memory, the brain's neural mechanism underlying the recall of the memory is the same," Tonegawa said. It remains the cleanest demonstration that a memory can be written, not just read.

Thirteen years later, nobody has written an episodic memory into any animal larger than a mouse. Here is why, in one number: about 500,000; do the division.

The engram addressability gap

About 500,000: the number of neurons you would need to individually address to run the 2013 experiment in a human.

To replicate the 2013 experiment in a human, you would need to individually address roughly half a million neurons in the dentate gyrus. What follows is a Fermi estimate, and I will show the work. Your dentate gyrus holds on the order of 10 million granule cells. In the mouse experiment, something like 5 percent of those cells got labeled, the fraction active during context exploration. Five percent of 10 million is 500,000 neurons, each needing to be tagged by its activity history and then reactivated on command.

Against that, set the best human neural interface ever implanted: Neuralink's N1, with 1,024 electrodes. 500,000 divided by 1,024 is 488. So the channel count alone is roughly 500x short, and that understates the gap, because electrodes do not address single neurons at all, only the blurry volumes around them, which means the true addressability deficit runs deeper than the headline ratio. Each one stimulates a volume containing hundreds to thousands of cells. Worse, the activity-dependent tagging at the heart of the mouse experiment requires genetic engineering of the hippocampus, which exists in no approved human therapy. And the optogenetics the mice relied on needs a fiber optic cable in the skull. Human optogenetics has been tried exactly once, in the retina, flooding a million cells with light at once, a floodlight where the mouse experiment needed a laser pointer aimed at cells chosen by their own activity history. Nobody has put light-sensitive proteins in a living human hippocampus, and nobody is close.

So the honest status of surgical memory injection is: 500x away in channel count, plus one categorical gap, plus one missing technology. Call it three orders of magnitude and a canyon, decades minimum.

There is a second honesty anchor that rarely gets stated. Step back and count what the 2013 experiment actually wrote. One bit of information: fear, attached to a room. Not a birthday party. Not a face. A single association between a context and a shock, expressed as freezing. Every "we implanted a memory" headline since has been describing one bit. Nobody has written an episode into any brain. One binary digit, delivered thirteen years ago, to a mouse.

Four routes, one scoreboard

Route	Mechanism	Human evidence	What it actually writes
Optogenetic (Ramirez et al., 2013)	Light reactivation of activity-tagged hippocampal neurons	None. Mice only.	One association (fear ↔ context)
Pharmacological (propranolol reconsolidation)	Beta-blocker during memory reactivation disrupts re-storage	13 RCTs: 4 positive, 7 null, 2 negative	Weakens emotional charge; unreliable
Electrical (Berger/Hampson MIMO prosthesis)	Play back "correct-answer" CA3→CA1 codes during encoding	+35–37% recall improvement, 8 patients, DARPA-funded	Boosts memories you already formed; writes nothing new
Pixel (AI-edited media, VR, chatbots)	Suggestion + vivid imagery exploits reconstructive memory	2.05x false recollections (n=200); 36.4% false-memory rate via chatbot	Rich false autobiographical detail

The pattern is almost insulting. Here is the pattern, and it is almost insulting: the three routes that sound like engineering barely work in humans, while the route that works needs nothing but a screen.

Drugs: a coin flip

Propranolol, the blood-pressure drug, became famous as a potential memory editor because memories become briefly editable each time they are recalled, a window called reconsolidation. Dose the patient, trigger the traumatic memory, and the emotional charge allegedly fails to re-save. Alain Brunet's 2018 randomized trial at McGill found real effects: a 56 percent drop in self-reported PTSD scores versus 15 percent on placebo across six weekly sessions.

Then came the non-replication. A later double-blind trial found no significant difference between propranolol and placebo, with the placebo group improving almost as much, probably because six weeks of structured trauma recall is itself a treatment, and the placebo kept up. A 2024 systematic review of 13 randomized trials of drug-augmented trauma therapy, pre-registered on PROSPERO before the literature search began, tells the final score across every trial the reviewers could find: four positive, seven null, and two in which the patients who got the drug did worse than the patients who got the sugar pill. That is not a therapy. That is a coin flip wearing a lab coat. So the biochemistry route to editing human memory is real, occasionally effective, and currently unusable as engineering.

Electricity: write-assist, not write

Ted Berger's team at USC and Robert Hampson's at Wake Forest built the closest thing to a memory prosthesis in humans: a multi-input multi-output model that reads the neural code passing from CA3 to CA1 during correct memory formation, then plays that code back through implanted electrodes to reinforce it. In eight epilepsy patients, episodic recall improved 37 percent on short delays and 35 percent at 75 minutes. DARPA funded it under the RAM Replay program. Later work tuned the stimulation to specific content categories, with improvements outnumbering impairments about two to one, though the base rate of useful effect stayed low.

Impressive. But read the fine print: the device only reinforces the encoding of things the patient actually experienced. It is write-assist, not write. You cannot feed it a memory the patient never had. Nobody has demonstrated content injection through electrodes in a human, and the 22.4 percent hit rate on content-specific stimulation means most attempts do nothing or hurt recall.

Pixels: already working

Now the route nobody needed a brain implant for. That MIT study is the cleanest result, but it sits on a pile. Stanford's Virtual Human Interaction Lab showed that elementary-school children shown an avatar of themselves performing a novel action in VR later remembered doing it. A 2024 witness-interview study found that a short interaction with a generative chatbot produced false memories in 36.4 percent of participants, roughly triple the control rate. A PLOS One study of 436 people found deepfake videos fooled about half the participants into false familiarity, though text descriptions worked about as well, which tells you the active ingredient is suggestion plus vividness, not the medium.

This is memory injection, operating today, at consumer scale, with zero regulatory oversight, because nobody regulates a photo editor as a neurotechnology device. Your camera roll is now a plausible attack surface for your own autobiography. Maes and Loftus themselves flagged the legitimate flip side: therapeutic memory reframing, deliberately rewriting the emotional framing of real memories, which is arguably what good trauma therapy already does with words.

Steel-manning the skeptics: the strongest case against

Loftus's own literature is the best check on the hype, and it cuts hard, because the researcher who spent fifty years proving memories can be fabricated is also the person who knows exactly where the fabrication stops. Implanted false memories in humans are almost always peripheral details grafted onto real events, the famous "lost in the mall" paradigm, not whole episodes fabricated from nothing. Central details resist. Lab demand characteristics inflate every rate: participants are clever people guessing what the experimenter wants. Confidence is not accuracy, and the 1.19x confidence bump in the MIT study is modest. A mouse "memory" is a freeze response, an association, not a recollection with a narrator. And the propranolol scoreboard above suggests the pharma route may be mostly noise. Stated at full strength, the honest version of this article runs like this: nobody can write an episode into a human brain, the pixel route writes confabulated details rather than lived experience, and the mechanism is suggestion, not upload. A real threat, not Inception: advertising.

What this analysis did not prove

The 500,000-neuron figure is a Fermi estimate built on two order-of-magnitude inputs, the 5 percent labeling fraction and the 10-million-cell human dentate gyrus, so treat the gap as "hundreds-fold," not exactly 488x. That 36.4 percent chatbot figure comes from secondary reporting of the study, not the primary paper. The MIT study measured false recollections of image content over short timescales, not durable autobiographical memories persisting for years, which is a meaningful difference when the claim is about your life story rather than a photo caption. And no human optogenetic memory write has ever been attempted, so the addressability gap is an extrapolation from mouse work and electrode counts, not a measurement. Solid direction; shaky decimals.

What you can do

If you keep a camera roll that your future self will treat as evidence, start treating provenance as hygiene. Turn on C2PA content credentials where your camera app supports them, so AI-edited images carry their edit history. When a photo triggers a vivid "memory" of something you are not sure happened, check the original file before you trust the feeling; the MIT result says your confidence will be working against you, so check the file. Parents should know the Stanford result cuts both ways: VR is a powerful learning tool precisely because it writes so effectively, which means unlabeled synthetic childhood "memories" deserve the same skepticism as unlabeled synthetic news. And if you work in therapy or clinical research, the legitimate version of this technology, VR-based memory reframing under clinical supervision, is the application worth building, because the illegitimate version is already shipping in every app store.

Watch for two signals that the surgical route is moving. First, closed-loop optogenetic memory experiments in non-human primates, which would shrink the species gap that currently ends at the mouse. Second, any activity-dependent gene therapy targeting the human hippocampus, which is the missing technology the whole addressability calculation assumes. Until both exist, the needle loses and the screen wins.

The Bottom Line

Memory injection is not a future technology. It is a present one with the wrong branding. Needle version: 500x away, writes one bit. Screen version: in your pocket, writing at twice the baseline rate. We spent thirteen years marveling at a mouse that froze in the wrong room, while the actual memory-writing device got a billion installs. Turns out the engram was never the bottleneck. Suggestibility was.

Sources: Ramirez et al., Science (2013); full text, PMC; Pataranutaporn et al., CHI 2025; Segovia & Bailenson, Stanford VHIL; Hampson et al., Wake Forest (2018); Hampson et al., Frontiers (2024); Brunet et al., Am J Psychiatry (2018); non-replication, PMC; Meister et al. systematic review, Critical Care (2024); chatbot witness study (2024).