Recently, scientists unlocked memories we thought were gone forever, AI proved weirdly resistant to manipulation (in the wrong way), and we discovered forests might be sabotaging our climate plans.
This week´s 10 Stories in AI Research:
🧠 Memory loss in Alzheimer's might be reversible: those memories aren't deleted, just locked away.
🤖 AI labels won't save us from manipulation campaigns; people know it's AI and still get persuaded.
🏥 Medical AI falls for formal hospital notes but spots Reddit-style misinformation, the opposite of what we expected.
🫀 AI stethoscopes now predict heart disease at 83% accuracy from a one-minute recording, outperforming human doctors.
🧬 The missing gene for the most common Ehlers-Danlos syndrome was never missing, but actually dozens of genes working together.
🎣 Random Forest and Cubic SVM beat 8 other AI models at catching phishing scams in the largest comparison study yet.
💊 AI-designed drugs now reach clinical trials in 12-18 months instead of 4.5 years, with 20 candidates already nominated.
🏗️ Scientists can now 3D print realistic brains with varying stiffness to plan surgeries on your exact brain structure.
🌲 Forests are leaking carbon from deep soil faster than trees can absorb it, undermining net-zero strategies worldwide.
🚶 The first fully-automatic smart brain implant learns when you're about to walk and adjusts stimulation in real-time.
#1
🧠 We can now unlock lost memories
Your memories aren't actually gone, they're just locked away.
Scientists at EPFL discovered that in aging and Alzheimer's brains, memory cells don't die, they just get jammed and can't be accessed anymore. Think of it like having a file on your computer that won't open because the software is corrupted.
What they did: They used gene therapy to rejuvenate specific memory neurons in mice. Using three special genes (called OSK), they essentially hit the reset button on old, worn-out memory cells but only the ones that held specific memories, not the whole brain.
The results were incredible:
Old mice with Alzheimer's suddenly remembered things as well as young, healthy mice
Their brain cells looked younger under the microscope
The cells actually became better versions of themselves instead of reverting to stem cells (which was a major fear)
A single treatment provided lasting benefits for weeks
Why it matters: This suggests Alzheimer's patients haven't permanently lost their memories, those memories are just inaccessible. One day, we might be able to unlock them again.
#2
🤖 AI Labels Don't Stop AI Persuasion
If we label AI-generated content, will people trust it less?
Many governments think the answer is yes. The EU even passed laws requiring AI labels. But researchers at Stanford and Northeastern decided to test this assumption.
The Experiment: They showed 1,601 people AI-written messages about policies like paying college athletes. Some were told it was written by AI, some were told it was a human expert, and some got no label at all.
The Surprising Results:
The AI messages changed people's opinions by nearly 10 percentage points
Knowing it was AI didn't make it less persuasive
92% of people believed the label, they just didn't care
The messages were logical and evidence-based, so the source didn't matter
The Exception: Older people were more skeptical of AI-labeled content than younger folks.
Why It Matters: Simply labeling AI content probably won't protect us from AI-driven manipulation campaigns. We need better solutions.
#3
🏥 Medical AI Falls for Lies (But Not the Ones You'd Expect)
Researchers bombarded 20 different AI health models with 3.4 million prompts containing medical lies to find their breaking point.
The Twist: Wrapping lies in obvious logical fallacies actually made the AI less likely to believe them. For example, saying everyone knows rectal garlic boosts immunity made the AI suspicious.
The Real Vulnerabilities:
AIs believed medical lies 31.7% of the time when presented neutrally
They were most easily fooled (46.1%) by formal, clinical-sounding hospital notes
They were much more skeptical of casual, Reddit-style conversations
Appeal to authority and slippery slope arguments actually worked to trick them
The Medical AI Paradox: AI models specifically trained for medicine actually performed worse than general-purpose models like GPT-4. Their strict training may have made them less flexible at spotting rhetorical tricks.
Why It Matters: Making AI bigger or wrapping it in fancy prompts won't make it safe for healthcare. We need to ground it in actual medical facts.
#4
🫀 AI Stethoscopes Hear What Doctors Miss
Over half of heart valve disease cases go undiagnosed because doctors mistake symptoms like fatigue for normal aging. Regular stethoscopes only catch about 45% of dangerous heart murmurs.
The Solution: Cambridge researchers built an AI that doesn't just listen for murmurs but predicts what an expensive echocardiogram (ultrasound of the heart) would show, just from audio.
The Performance:
83% accuracy overall (way better than human doctors)
98% sensitivity for severe aortic stenosis
94% sensitivity for severe mitral regurgitation
Only needs one minute of recording
The Surprise: The tricuspid auscultation site (a spot doctors don't usually focus on) turned out to be the most important single location.
Why It Matters: This could become a cheap, one-minute screening tool that catches deadly heart problems before they become emergencies—no specialist required.
#5
🧬 The Missing Gene for the Most Common EDS Finally Found (Sort Of)
Hypermobile Ehlers-Danlos syndrome (hEDS) affects 90% of EDS patients, but unlike other types, no one could find the gene causing it.
The Investigation: Researchers used AI to analyze nearly 36,000 genetic variants in 116 people from 43 families with hEDS.
The Revelation: There is no single hEDS gene. Instead, they found it's caused by multiple genetic hits across three major systems:
Structural proteins (the scaffolding that holds cells together)
Immune system genes (found in 74% of patients)
Mitochondrial genes that control cell energy (found in 34% of patients)
The Shocking Discovery: Kids with hEDS who had mitochondrial variants were 4.2 times more likely to have a history of bone fractures, suggesting a link between energy metabolism and bone strength.
Why It Matters: This explains why finding a single gene has been impossible; hEDS is more like a complex puzzle than a simple genetic typo.
#6
🎣 We Finally Know Which AI Best Catches Phishing Scams
Phishing incidents jumped 220% in 2022, with 1.4 million malicious websites created every month.
The Showdown: Researchers tested 10 different AI models against three different datasets of phishing attempts.
The Winners: Random Forest and Cubic SVM consistently performed best across all scenarios.
The Failure: The Ensemble Subspace KNN model looked perfect during training but fell apart during real-world testing. It was memorizing examples instead of learning patterns.
The Red Flags: The AIs learned that URL length and weird redirection symbols (//) are the biggest warning signs, along with pop-ups and external website icons.
Why It Matters: Companies now have clear evidence about which AI models to use in their email filters and security systems.
#7
💊 AI Drug Discovery Just Got 3× Faster
Traditional drug discovery takes an average of 4.5 years just to get to the first preclinical candidate.
The New Way: Insilico Medicine's AI platform can do it in 12-18 months.
The Efficiency Gains: From 2021-2024, Insilico nominated 20 drug candidates while only needing to test 60-200 molecules per program, a tiny fraction of the traditional trial-and-error approach.
The Partnership: They've teamed up with China Medical System Holdings to bring these AI-designed drugs through clinical trials and to patients.
Why It Matters: Faster drug discovery means life-saving treatments reach patients years earlier and potentially at lower costs.
#8
🏗️ Scientists Can Now 3D Print Realistic Brains
Computer models and standard 3D prints of brains are too uniform because they don't capture the variation in texture and stiffness that makes real brains unique.
The Innovation: University of Missouri researchers developed a technique that prints in a jelly-like support bath, allowing them to create soft folds and varying levels of stiffness simultaneously.
The Achievement: They can now print regions that mimic the mechanical, thermal, and electrical properties of real gray and white matter. They've successfully printed a 15% scale model and are working on a full-sized brain.
The Future Applications:
Surgeons could print a custom model of a specific patient's brain from their MRI to plan complex surgeries
Engineers can test how medical implants interact with brain tissue without human subjects
Researchers can study concussions and brain injuries in realistic models
Why It Matters: Personalized brain surgery planning and safer medical device testing without risking human lives.
#9
🌲 Bad News: Forests Are Leaking Carbon
Plant trees → Trees absorb CO2 → Carbon safely stored in soil. Many net-zero policies rely on this.
The Reality Check: Researchers studied European beech forests for 38 years (1984-2022) and found something disturbing.
What They Discovered:
While the trees grew and stored carbon, the soil underneath was losing carbon
Deep soil layers (up to 90 cm down) leaked so much carbon that it offset 17-20% of what the trees absorbed
The culprit: A 2°C temperature increase boosted microbial activity by 15%, causing soil to exhale carbon faster than it could store it
Why It Matters: Current carbon credit programs don't account for these deep-soil losses. The climate benefit of tree-planting may be significantly overestimated. We can't just plant trees and assume the problem is solved.
Read more here
#10
🚶 The First Smart Brain Implant for Parkinson's
Current Parkinson's brain implants provide constant stimulation, which is like leaving your house lights on 24/7, which is inefficient and not ideal for different situations.
The Innovation: Researchers created the first fully implantable device that knows when you're walking and adjusts automatically.
The Study: Four patients wore the device at home for 80 hours. The device learned personalized brain signals that indicated when they were about to walk.
The Breakthrough: The implant can run a low-power AI classifier on its own chip, meaning it can automatically switch to walking mode the second you stand up, with no external computer needed.
Why It Matters: This could dramatically reduce freezing of gait episodes (when Parkinson's patients suddenly can't move) and prevent dangerous falls. It's personalized medicine at the neural level.
Extras
By the Numbers
83%: The accuracy rate (AUROC) of a new AI algorithm in detecting valvular heart disease, which significantly outperforms the 45% sensitivity common among human general practitioners.
2°C: The average annual temperature increase in European beech forests that triggered substantial deep-soil carbon losses, offsetting 17% of the trees' aboveground carbon gains.
12–18 months: The lightning-fast timeline for AI-driven drug platforms to nominate a preclinical drug candidate, compared to the traditional industry average of 4.5 years.
Deep Dives
The Readable Memory Bug: Why your brain might still store memories it can no longer access due to epigenetic erosion—and how targeted gene reprogramming can polish cellular machinery to restore youthful learning.
The Hidden Signal: Why the tricuspid site—a location often given less priority in routine heart checks—emerged as the most vital single spot for AI to catch deadly murmurs.
The Best Shield: With 1.4 million new phishing sites appearing every month, researchers tested ten different machine learning architectures to find the ultimate winner: Random Forest.
Pique Your Interest
Predicting Fragility: New genomic analysis shows that specific mitochondrial variants in children with hypermobile Ehlers–Danlos syndrome are linked to a 4.2-fold higher risk of pediatric fractures.
The Fine-Tuning Paradox: Counter-intuitively, AI models specifically trained on medical data were often more susceptible to medical misinformation than general-purpose models like GPT-4o.
The Influence Trap: Telling a user that a message is AI-generated has almost zero effect on its persuasiveness, even though 92% of people believe the label is true.
Interesting Concepts
Engrams: Sparse groups of neurons that serve as the physical trace or storage site of a specific memory in the brain.
Partial Reprogramming: A technique using Yamanaka factors to reset a cell's biological age without causing it to lose its specialized identity or turn into a stem cell.
Logical Fallacies: Flawed reasoning patterns (like Appeals to Popularity) that contemporary AI models actually use as coarse heuristics to help them identify and reject medical lies.
Phonocardiogram: A high-fidelity digital recording of the sounds and murmurs produced by the heart, which AI can analyze for patterns inaudible to the human ear.
Today´s Wildcard
The Jelly-Bath Brain: Engineers are using a jelly-like support bath to 3D-print artificial brains that mimic the soft folds and varying stiffness of real gray and white matter. Within the next year, surgeons could potentially practice complex procedures on a life-sized, physically accurate replica of a patient’s specific brain.
Thanks for reading, and until next time,
Maryam