2026-06-12 03:04:52
While we still can’t explain how AI works, algorithms are rapidly learning what makes us tick. And the gap is widening.
AI is becoming more powerful, and mysterious.
Despite years of work on “explainable AI,” today’s most advanced systems remain black boxes for the most part. Scientists can observe what they do but cannot fully explain how they arrive at their conclusions or predict when they’ll fail.
As large language models (LLMs), the algorithmic engines behind popular chatbots, permeate society, researchers are warning that the window for understanding AI “minds” is rapidly closing even as the technology’s influence expands.
Last week, Eric Horvitz, chief scientific officer at Microsoft, and Robert West at EPFL in Switzerland outlined the dangers of putting AI interpretability on the back burner. They call for new AI benchmarks and better tools for unpicking machine minds.
The challenge resembles efforts to understand our own minds. Some researchers have already taken a neuroscience-inspired approach, mapping AI’s internal networks to concepts, goals, and reasoning. Others borrow from psychology, treating AI as a participant of behavioral studies.
The stakes are rising. AI tools already shape how people search for information, make decisions, and form judgments. Their answers influence everyday users and the researchers who build them.
As AI capabilities grow, our understanding of them could fall behind. “Preserving human agency must therefore remain a central goal,” the authors write.
LLMs are built on artificial neural networks (specifically, a design called the transformer). Inspired loosely by the brain, these networks connect vast numbers of artificial neurons into intricate architectures. The basic idea is straightforward. Data enters the network and passes through layers of computations, which transform it into an output like text or code.
At first, that output is often wrong. But with feedback and repeated training, the network adjusts the strengths of connections between neurons and gradually improves. It learns.
After initial training, engineers turn to reinforcement learning, where algorithms improve through trial and error and further hone their responses. Another method, inspired by how the brain etches memories during sleep, reduces the tendency to forget old knowledge while learning new tasks. And self-attention, the key innovation behind transformers, allows AI to selectively focus on various words, images, sounds, or video frames at different moments, boosting efficiency and performance. Today, attention underpins nearly every major AI system.
Yet the inner workings of finished algorithms remain hidden.
Early efforts to crack open AI’s black box examined how artificial neurons responded to images, revealing that neural networks build increasingly more sophisticated “ideas” of the world. Google Brain borrowed methods from cognitive psychology to study AI behavior, while others investigated whether LLMs could mimic aspects of “theory of mind”—the ability to infer what others are thinking and feeling.
These studies laid the foundation for a popular method called mechanistic interpretability. Anthropic, creator of Claude, is leading the field. Company researchers have linked patterns of algorithmic activity to specific concepts and reverse engineered parts of neural networks to expose how internal computations shape responses.
Other tech giants are joining the cause. OpenAI is training algorithms that work in more explainable steps and building reasoning models that pause, “think,” and justify their conclusions in plain language. DeepMind is building microscope-like tools for neural networks, helping researchers peer into their decision-making process. And Microsoft has released new tools aimed at responsible use of AI.
Understanding AI, the authors write, does not require tracing every line of code or every neural-network parameter. Just as neuroscience, psychology, and sociology offer different windows into human behavior, AI can be studied at multiple levels, from how individual circuits work to observing behavior in real-world scenarios.
The challenge is that AI capabilities may be advancing faster than our ability to explain them. And some researchers believe time is running out.
Three trends are making AI more opaque.
The first is how we evaluate AI. Increasingly, LLMs we being used to train, benchmark, and improve other models. AI “judges” now score metrics like helpfulness, rank competing outputs, detect hallucinations, and assess new releases. In a system known as constitutional AI, for example, algorithms critique their own responses using reinforcement learning and generate explanations for their reasoning. Other researchers have proposed AI debate frameworks, where multiple models challenge each another’s conclusions before a human has the last say. Researchers are also exploring automated interpretability tools. Like digital neuroscientists, AI systems are used to analyze each other—describing neurons, circuits, and behavioral patterns—to explain increasingly complex models.
Using AI to solve an AI-induced problem introduces a paradox. If AI-generated explanations become too complex for humans to verify, opacity compounds.
A second trend is the rise of AI societies. Networks of interacting AI agents are becoming more common, particularly in complex tasks such as scientific research and drug discovery. Yet as they become more sophisticated, their communication could drift from human language and reasoning, making them harder to interpret.
Studying their interactions with methods adapted from sociology could unveil unexpected norms, hidden rules, and collective behavior. The authors argue that training in the future should not only reward effective collaboration among AI agents, but also ensure humans can understand their communication.
The last trend already permeates our lives. ChatGPT, Claude, Gemini, and other LLMs listen to our woes, offer recipes, and code websites. But they also learn about humanity. Through training data and interactions, they glimpse how people think, reason, and feel. In turn, they capture core aspects of life, such as fear, anxiety, happiness, and the need for social belonging.
To be clear, the systems don’t have intentions. They’re not examining us. But even as we struggle to understand them, AI systems are building more sophisticated models of who we are.
“A striking asymmetry follows: While human understanding of AI declines, AI understanding of humans deepens, producing new forms of behavioral opacity,” the authors write.
But complacency is perhaps even more insidious. AI assistants are often optimized to be agreeable, helpful, and reassuring. Studies have found that people generally prefer AI agents that support their opinions and decisions. As AI is woven into everyday life, curiosity and skepticism may gradually give way to trust. They work. Why question how?
The authors don’t have a solution for the long-standing problem. Instead, they call for better benchmarks to measure AI capabilities and stronger evaluation methods. And while open-source projects and crosstalk between commercial companies and academia are now frequent, they say we need lasting norms of responsible disclosure. Mechanistic interpretability and AI “psychology” could build on each other.
“The goal is not just more capable AI, but AI that is more intelligible, accountable, and aligned with human aims,” they write.
The post AI Is Advancing Faster Than Our Ability to Understand It, Researchers Warn appeared first on SingularityHub.
2026-06-11 02:19:27
New drugs are taking on the slippery molecular switches that fuel deadly cancers—and AI is speeding up the hunt.
For decades, a handful of molecular switches has haunted the nightmares of cancer researchers. The switches trigger runaway tumor growth and cause the disease to spread across the body in multiple cancers. In theory, this makes them perfect treatment targets. Blocking even one could lead to drugs that are effective against a variety of cancers.
But despite considerable efforts, these switches—all of which are proteins—have escaped our most advanced cancer treatments, earning them the term “undruggable.” This is largely due to a shared trait: They all have smooth surfaces, making it difficult for drugs to interact with them.
But maybe not for much longer.
Researchers recently reported promising results for a new medication targeting a family of undruggable proteins in a clinical trial for advanced pancreatic cancer. The drug, daraxonrasib, nearly doubled survival time compared to chemotherapy, with fewer side effects. It’s not a total cure. But the treatment gives patients precious time, adding roughly 13 months after diagnosis. Patients also reported less pain and better quality of life.
Daraxonrasib is the latest in a new generation of drugs aimed at undruggable proteins. And AI-based tools are now poised to further accelerate progress in the field.
The RAS family was the first group of oncogenes—or genes that drive cancer—ever discovered. The genes became a major focus in 1982 when several teams independently showed the mutation of a single DNA letter could transform RAS genes into a potent cancer trigger.
The proteins RAS genes encode are like spring-loaded molecular switches that relay signals from a cell’s surroundings. When proteins called growth factors latch onto a cell, RAS switches flip on to promote cell growth and survival, while built-in safeguards quickly turn them off again.
Cancerous mutations break this cycle. The switches get stuck in the “on” position, continuously instructing cells to grow and divide. This is, of course, a hallmark of cancer.
An ideal drug would simply switch RAS off. But most drugs are like rock climbers. They need grooves, pockets, or bumps on a protein to grab onto. Similar to a smooth rock face, RAS offers few such features. Making matters worse, different mutations subtly reshape the protein, so it’s tough to build a one-size-fits-all inhibitor.
The first RAS drug wasn’t approved in the US until 2021, nearly four decades after discovering the genes’ role in cancer. Even then, the drug targeted just one family member of three, limiting its reach to a relatively small group of patients. Many eventually developed resistance.
That’s why daraxonrasib turned heads. Developed by Revolution Medicines in Redwood City, California, the drugs switches off all three RAS family members. Rather than trying to grip the slippery proteins directly, it binds to a partner molecule that helps RAS proteins fold into their final 3D shapes. In this way, the drug hitches a ride on active RAS and shuts the proteins down.
The workaround paid off. The new study enrolled 500 people worldwide with advanced pancreatic cancer. All participants had already tried cancer therapies with limited success. On average, patients receiving daraxonrasib lived 13.2 months and spent most of that time with limited pain. The most common discomfort was a rash. Those receiving chemotherapy fared worse, living roughly 6.6 months and experienced more severe side effects.
The results don’t rival the dramatic success of CAR T cell therapies in blood cancer. In CAR T, caregivers engineer a patient’s own immune cells to recognize and attack tumors, sometimes producing long-lasting remission after a single infusion.
But the findings have energized the field. If approved, a daily daraxonrasib pill would likely be far more affordable and easier to administer than a personalized cell therapy. And because RAS mutations fuel many solid cancers—which CAR T still struggles to control—the drug could offer a new defense against deadly cancers that are largely beyond cell therapy’s reach. Combining daraxonrasib with earlier-generation RAS inhibitors may further boost its effects.
Daraxonrasib didn’t appear overnight. Scientists used a crystallized snapshot of its target protein as a molecular blueprint. Years of medicinal chemistry followed, with scientists repeatedly tweaking candidate compounds to boost potency, improve selectivity, and minimize toxicity.
AI could dramatically accelerate similar efforts against other undruggable cancer targets. Among the most coveted is p53, often called the “guardian of the genome” for its dizzying array of roles. The protein orchestrates the activity of over 300 genes involved in DNA repair, metabolism, cell death, and inflammation, making it one of the cell’s most important defense systems.
Since its discovery in 1979, p53 has been both a holy grail and a headache for cancer researchers. Mutations in the gene are common in multiple cancers. But like RAS, the protein is flat and smooth. Some mutations destabilize its structure; others turn it into misfolded clumps. A universal p53 drug has remained elusive.
Some researchers are trying to restore the protein. In a small trial earlier this year, they tested a drug that restabilizes a common mutant form of p53. Within 21 days, tumors shrank roughly 20 percent in patients with ovarian, breast, and several other solid cancers.
Other researchers aim to selectively kill cells carrying the mutation. Using AI, a team at Baylor College of Medicine screened nearly 10 million compounds that cause mutated p53 cells to self-destruct, while sparing healthy cells. The search uncovered 83 chemically distinct candidates. One called H3 dramatically suppressed tumor growth in mice.
“These results highlight the potential use of AI-powered drug screening to investigate individual p53 mutants in the future,” they wrote. Although the approach is early-stage and only focused on one mutation, the team is hopeful it can be extended to other cancerous mutations.
MYC is another formerly undruggable protein that could now be vulnerable. Roughly 70 percent of cancers have abnormal MYC activity. Normally, the protein is a master regulator of growth, directing cells to manufacture proteins, replicate DNA, absorb nutrients, and divide when needed.
Cancer finds many ways to hijack the system and keep cells in a state of runaway growth. MYC gene mutations aren’t just single-letter swaps. Sometimes the gene duplicates or is rearranged across the genome, churning out excessive amounts of the protein it encodes. This genetic diversity makes approaches using gene therapy difficult. And again, like RAS, the MYC protein’s smooth, featureless surface lacks stable anchors for drugs.
An emerging strategy is to disrupt MYC’s interaction with other proteins that it needs to function. A designer protein blocking MYC activity, for example, recently showed promise in a small trial against solid cancers. Other teams are using AI to identify drugs that limit MYC’s ability to fix damaged DNA in tumors, kneecapping their ability to divide. Meanwhile, biotechnology companies are deploying AI to map out MYC’s structure and molecular interactions in search of new ways to shut the protein down.
Daraxonrasib’s success shows that undruggable proteins aren’t untouchable. There’s a lot more work ahead to prove other similar drugs can work too. But scientists are increasingly leaning into AI during all stages of drug development to speed up the process. Maybe, one day, “undruggable” will disappear from our vocabulary altogether.
The post After Decades of Failure, ‘Undruggable’ Cancers Begin to Give Way appeared first on SingularityHub.
2026-06-09 05:56:59
A planetary defense system would blunt solar storms with hundreds of tons of gas. Emerging heavy-lift rockets could deploy it in under two months.
Extreme space weather could wreak havoc on the satellites, communications networks, and electrical grids that modern society depends on. Researchers have now proposed an ambitious space-based planetary defense system that would weaken solar storms before they hit Earth.
The sun regularly emits massive pulses of radiation, energetic particles, and magnetic fields that interact with the Earth’s own magnetic field. This activity is the source of auroras like the northern lights, but the most violent eruptions can cause geomagnetic storms with the power to disrupt GPS and radio communications and fry electrical equipment.
While the impact of most of these events is limited, there is precedent for more catastrophic outcomes. In 1859, the Carrington Event, the most powerful solar storm ever recorded, knocked out telegraph lines across North America and Europe. In today’s highly electrified world, a similar event could cause between $2.4 and $3.4 trillion in damage to the power grid alone.
Now, researchers at Boston University and the University of Michigan have come up with a potential solution. In a paper published in Space Weather, they propose a constellation of satellites called StormWall that would release hundreds of tons of gas into orbit to blunt the force of an incoming solar storm.
“It’s as if you could install an airbag in the magnetosphere,” co-author Daniel Welling, a space physicist from the University of Michigan, told Science.
Solar storms have the potential to sow chaos because they weaken the magnetic shield protecting Earth from space radiation. Powerful enough storms disrupt the Earth’s magnetic field and cause it to reconnect to the sun’s, allowing energy from the solar storm to pour into the magnetosphere.
The Earth already has a natural defense against this—a doughnut-shaped reservoir of ionized gas, or plasma, sitting just above the atmosphere. When the planet’s magnetic field is disturbed, a plume of this plasma flows toward the sun and slows the rate at which the magnetic fields reconnect.
StormWall would turbocharge this process by releasing massive amounts of artificial plasma into the outer atmosphere. The researchers sketch out a system involving a constellation of satellites orbiting about 22,000 miles from Earth. The satellites would carry canisters of lithium, barium, or sodium gases to be ejected when a large solar storm is inbound. The gases, rapidly ionized by solar radiation, would add to the planet’s natural plasma shield.
Based on simulations, the researchers estimate that releasing around 400 tons of gas could reduce the strength of a major geomagnetic storm by over 50 percent. Crucially, the intervention would be swift and reversible. The plasma cloud could be in position by the time a storm hits, and it would dissipate just a few hours later.
Launching this much material into orbit would be a big undertaking, but the researchers say it could be within reach of emerging heavy-lift vehicles like SpaceX’s Starship or China’s Long March 9 rocket. They calculate that six launches could deploy the full constellation in under two months.
Outside experts have been broadly positive. Allison Jaynes, a space physicist at the University of Iowa, told Science the idea was “highly innovative and appears to be quite feasible in the near term.”
But getting the satellites into orbit is only part of the puzzle. Accurate and timely space weather forecasts would also be a prerequisite. And gaining international buy-in for a system that would drastically alter the near-Earth space environment, even if only temporarily, could be challenging.
The researchers flag potential side effects that need more study, including the generation of electromagnetic waves as the released material ionizes. Still, given the devastation a Carrington-sized event could unleash on the modern world, the potential downsides may be worth the risk.
The post Orbital Airbag Could Shield Earth From Devastating Solar Storms appeared first on SingularityHub.
2026-06-08 00:58:48
Jeff Bezos Is Funding a Wild Hunt for the Brain’s ‘Core Algorithm’Steven Levy | Wired ($)
“The goal, Reardon tells me, is to build ‘a synthetic artificial intelligence brain that runs on 50 watts or less.’ It should adapt to its conditions, be as nimble as a human mind, and burn a tiny fraction of an LLM’s compute power and energy. The proof of concept is thriving inside our skulls.”
Researchers Are Using AI to Create Vaccines—and It’s Working
Ed Cara | Gizmodo
“An experimental pan-coronavirus vaccine developed with AI has just passed a phase I trial in the UK. Scientists at the University of Cambridge used AI to find a kink in the armor of coronaviruses, including SARS-CoV-2, the cause of covid-19. …The researchers are also hoping to use their platform to develop broadly effective vaccines against flu and the Ebola virus.”
China Has Approved the World’s First Invasive Brain-Computer Chip—Here’s What’s NextYou Xiaoying | MIT Technology Review ($)
“This March, the implant Dong [Hui] uses became the first invasive BCI product in the world to be approved for use beyond clinical trials. It’s now available to some patients with paralysis in their limbs due to spinal cord injuries. We spoke to a range of experts to understand why the device was able to reach this global milestone, what makes this moment so significant, and what to expect next.”
Huge Study of Alzheimer’s Genetics Identifies New Drug TargetsChris Simms | New Scientist ($)
“The biggest genetic study of Alzheimer’s disease so far has identified 127 gene locations that are associated with the condition, of which 48 are new. The study also pinpoints several genes that could be prioritized as drug targets and cell types linked to a higher genetic risk of the condition.”
This AI Weather Startup Is Out-forecasting Government AgenciesTim Fernholz | TechCrunch
“One simple way to understand it, WindBorne’s chief product officer Kai Marshland says, is that WeatherMesh-6 ‘is as accurate five days out as a traditional forecast is the day before,’ particularly on surface temperature measurements. WeatherMesh-6 produces a forecast every hour, as opposed to every six hours, as traditional models do, and its resolution is now down to 3 km in the continental US.”
Microsoft’s Next-Gen Quantum Chip Cuts Timeline to Useful Quantum ComputingTom Warren | The Verge
“Microsoft claimed last year that it had made a key breakthrough in quantum computing with Majorana 1, the company’s first quantum processor. While physicists were immediately skeptical of Microsoft’s claims, the software giant is announcing Majorana 2 today, the next generation of its topological quantum chip.”
SpaceX’s Next Big Business Could Be Building Stuff in SpacePassant Rabie | Gizmodo
“The FAA recently approved test flights of the company’s [Starfall] reentry vehicles. …With Starfall, SpaceX would add in-orbit manufacturing to its business portfolio. The idea of in-orbit manufacturing has been around for decades, using the microgravity environment to manufacture materials that would otherwise be impossible to produce on Earth.”
China Aims AI at Predicting Who Could Pose a Political RiskJulian E. Barnes | The New York Times ($)
“A Chinese company has been trying to develop artificial intelligence-powered technology that would enable authoritarian governments to not just monitor dissidents but also potentially predict who could become one in the future. The work, which appears to be in the research stage, is ripped out of dystopian science fiction, offering a glimpse of a world in which an authoritarian state is able to move against its citizens before they begin any public dissent.”
AI Evaluators Struggle with Models That Know When They’re Being TestedRocket Drew | The Information ($)
“AI researchers are starting to make progress on a confounding problem: AI models are getting better at telling when they are in an evaluation. …If models act differently during testing, that could mean they get released with undesirable tendencies. It could also undermine their creators’ ability to show off test scores to potential clients.”
Moderna Gets $50 Million to Develop MRNA Ebola Vaccine Against BundibugyoBeth Mole | Ars Technica
“The global health organization Coalition for Epidemic Preparedness Innovations (CEPI) announced Monday that it will ‘urgently accelerate development’ of three vaccine candidates against Bundibugyo ebolavirus (BDBV), pledging a little over $60 million in the effort to extinguish an outbreak currently raging out of control in the Democratic Republic of the Congo.”
World’s First Underwater Data Center Is Now Online, Powered by WindBronwyn Thompson | New Atlas
“Just over seven months from completing phase one of this mega-project, Chinese engineers have finished the build and switched on the world’s first underwater data center (UDC) powered by offshore wind turbines. What’s more, it doesn’t need freshwater and cuts land use by more than 90% compared with above-ground centers.”
Gemini Spark Is the Most Impressive and Terrifying AI Experience I’ve Had YetDavid Pierce | The Verge
“On the one hand, this is one of the most astonishingly impressive AI experiences I have ever had. …On the other hand, I can’t shake the deeply creepy feeling I get from the whole thing. What Spark did feels sort of magical, and very invasive. It’s weird that Spark is so casually telling me the names and ages of my children, reminding me that it knows where I live, and finding information I know for a fact I’ve never volunteered to Google.”
The post This Week’s Awesome Tech Stories From Around the Web (Through June 7) appeared first on SingularityHub.
2026-06-05 22:00:00
The findings could lead to new treatments for multiple neurodegenerative diseases.
Huntington’s disease is tragically predictable. An inherited genetic mutation causes neurons to make distorted, sticky proteins. These proteins clump together and gradually overwhelm brain cells. The brain loses its ability to learn, remember, and make decisions.
This story is dogma in neuroscience. But decades of research and drugs targeting the clumps have had little success. Scientists are now wondering: Is there more to the story? In a twist, a team from the Hebrew University of Jerusalem and collaborators found that protein clumps may be a neuron’s first line of defense against damage.
The misfolded or malfunctioning proteins are quarantined inside bubbly hubs called “inclusion bodies.” Often considered detrimental to cell health, disrupting their formation unexpectedly led to cells becoming more sensitive to stressors often seen in neurodegenerative diseases.
Physical separation played just one part. Inclusion bodies also changed the activity of genes involved in neuroinflammation—even in the absence of immune cells. Scouting the genetic landscape of cells derived from patients with severe Huntington’s disease, the team homed in on a “master regulator” gene, ATF3, that orchestrates immune responses. Removing the gene lessened inclusion bodies’ protective effects against damage in cultured cells.
To be clear, the findings are only for a cell model of Huntington’s disease in a petri dish. And inclusion bodies could be a double-edged sword: protective in the beginning and detrimental later on. Still, acknowledging them as a more complicated villain could better inform strategies for disorders that take over our minds like Huntington’s.
“Our results reveal…that these structures are not merely byproducts of disease, but a central factor in the cell’s ability to mount a protective response against stress,” said study author Eran Meshorer in a press release.
It’s long been believed that protein clumps in the brain gradually erode cognition. Whether they’re the main driver of neurodegenerative disorders is still debated, but their presence accelerates brain cell injury, causing neurons to wither away.
Alzheimer’s disease, for example, is associated with two sets of protein clumps. One lives inside neurons (tau) and another gunks up the space between cells (amyloid). Decades of research aimed at removing amyloid clumps have met with minimal success, earning these doomed efforts the notorious nickname “graveyard of dreams.” Despite their struggles, the FDA recently approved two major drugs that remove amyloid clumps and modestly slow cognitive decline, though the approval has been controversial due to doubts about safety.
Other untreatable neurodegenerative disorders also fall into this category. Clumps formed in Parkinson’s disease erode the brain’s ability to control movement, emotion, and even the perception of time. Lou Gehrig’s disease, or ALS, produces inclusion bodies inside motor neurons, leading to muscle weakness and trouble swallowing. The disease eventually robs people of speech and motion.
These diseases often have multiple genetic and environmental triggers. Huntington’s, in contrast, is entirely genetic. The condition stems from the genome over-copying parts of the huntingtin gene (HTT), which normally makes a key protein also called huntingtin.
Normally, cells use the protein’s large, stackable structure to build highways that transport all sorts of biological cargo, from molecules to organelles. The protein also plays an essential role during early brain development and neural wiring in adulthood.
But a mutant form of the HTT gene can wreak havoc. A common mutation, called polyQ expansion, produces unwieldy, misfolded proteins. Nearly 30 years ago, researchers found that these errant proteins aggregate inside parts of the cell. The clumps, or inclusion bodies, were widely thought to be detrimental. Some act like sticky tape that captures healthy proteins, such as those involved in gene expression, and torpedoes cellular health.
But telltale signs in cultured rat brain cells suggest a more nuanced story: Inclusion bodies could also be protective, sequestering mutant proteins as an early form of protection.
The common factor in diseases featuring polyQ mutation is repetition. Mutated genes have long, duplicated sequences of the DNA letters cytosine, adenosine, and guanine (CAG). More CAG repeats in the genome translates into earlier disease onset.
We all have this DNA triplet in our HTT gene. But more than 39 repeats results in longer, toxic huntingtin proteins. Severe cases of Huntington’s can feature over 100 CAG repeats, transforming the usually free-floating protein workers into sticky, dysfunctional layabouts.
In the new study, the researchers first established a baseline. They used the gene editing tool CRISPR-Cas9 to reduce CAG repeats in cells derived from Huntington’s patients—which carried over 180 copies—to near normal levels.
They then tagged the cells with a fluorescent marker that causes huntingtin proteins to glow bright green under the microscope. This let the team track protein aggregation in real time. Though they shared the same genetics, some cells formed inclusion bodies; others didn’t.
The team next challenged them with a chemical known to cause cellular stress. Those that formed clumps survived far more regularly than those that didn’t. It was a “striking difference,” the authors wrote. “Once a mutant PolyQ protein is expressed, the formation of IBs [inclusion bodies] protect[s] the cells rather than inflict[s] harm, at least short-term.”
Inflammation seems to be key. Although grown side-by-side, a genetic screen revealed cells with inclusion bodies were especially abundant in a gene called ATF3, which is known to regulate inflammation. Getting rid of the gene wiped out the neurons’ ability to form inclusion bodies, making them more vulnerable.
“Our results reveal a previously unknown role for ATF3 in orchestrating the formation of inclusion bodies in human neurons,” said Meshorer.
These are very early results. An immune molecule bridges ATF3 and inflammation and is associated with Huntington’s disease. Its levels are higher in patients with the condition. Increasing ATF3 activity could amp up the number of protective inclusion bodies and give neurons a fighting chance.
The findings suggest inclusion bodies gather free-floating mutant proteins into clumps to protect neurons and reduce brain damage—at least at the beginning of the disease. However, lab experiments rarely translate to treatments. How fast inclusion bodies form and when they begin to stress cells remains to be seen. Meanwhile, a gene therapy for Huntington’s is underway, and promising results in a small trial suggest an alternative path for treatment.
Still, the study challenges the idea that protein clumps are always detrimental. If replicated in other neurodegenerative diseases such as Alzheimer’s or ALS and if we can learn how long protection lasts, the results could pave the way for better-timed treatment that works with the body’s protection, not against it.
The post Toxic Clumps in Huntington’s Disease May Protect the Brain Too appeared first on SingularityHub.
2026-06-05 06:17:05
The gap between what AI can do in biology and what governance systems are prepared to handle is growing.
Artificial intelligence is rapidly learning to autonomously design and run biological experiments, but the systems intended to govern those capabilities are struggling to keep pace.
AI company OpenAI and biotech company Ginkgo Bioworks announced in February 2026 that OpenAI’s flagship model GPT-5 had autonomously designed and run 36,000 biological experiments. It did this through a robotic cloud laboratory, a facility where automated equipment controlled remotely by computers carries out experiments. The AI model proposed study designs, and robots carried them out and fed the data back to the model for the next round. Humans set the goal, and the machines did much of the work in the lab, cutting the cost of producing a desired protein by 40 percent.
This is programmable biology: designing biological components on a computer and building them in the physical world, with AI closing the loop.
For decades, biology mostly moved from observation toward understanding. Scientists sequenced the genomes of organisms to catalog all of their DNA, learning how genes encode the proteins that carry out life’s functions. The invention of tools like CRISPR then allowed scientists to edit that DNA for specific purposes, such as disabling a gene linked to disease. AI is now accelerating a third phase, where computers can both design biological systems and rapidly test them.
The process looks less like traditional benchwork in a lab and more like engineering: design, build, test, learn, and repeat. Where a traditional experiment might test a single hypothesis, AI-driven programmable biology explores thousands of design variations in parallel, iterating the way an engineer refines a prototype.
As a data scientist who studies genomics and biosecurity, I research how AI is reshaping biological research and what safeguards that demands. Current safety measures and regulations have not kept pace with these capabilities, and the gap between what AI can do in biology and what governance systems are prepared to handle is growing.
The clearest example of how researchers are using AI to automate research is AI-accelerated protein design.
Proteins are the molecular machines that carry out most functions in living cells. Designing new ones has traditionally required years of trial and error because even small changes to a protein’s sequence can alter its shape and function in unpredictable ways.
Protein language models, which are AI systems trained on millions of natural protein sequences, can quickly predict how mutations will change a protein’s behavior or design new proteins. These AI models are designing potential new drugs and speeding vaccine development.
Paired with automated labs, these models create tight loops of experimentation and revision, testing thousands of variations in days rather than the months or years a human team would need.
Faster protein engineering could mean faster responses to emerging infections and cheaper drugs.
Researchers have raised concerns that these same AI tools could be misused, a challenge known as the dual-use problem: Technologies developed for beneficial purposes can also be repurposed to cause harm.
For example, researchers have found that AI models integrated with automated labs can optimize how well a virus spreads, even without specialized training. Scientists have developed a risk-scoring tool to evaluate how AI could modify a virus’s capabilities, such as altering which species it infects or helping it evade the immune system.
Current AI models are able to walk users through the technical steps of recovering live viruses from synthetic DNA. Researchers have determined that AI could lower barriers at multiple stages in the process of developing a bioweapon, and that current oversight does not adequately address this risk.
Experienced scientists are already using AI to plan and design biological experiments. The question of whether AI can help people with limited biology training carry out dangerous lab work is the subject of active research.
Two recent studies have reached different conclusions.
A study by AI company Scale AI and biosecurity nonprofit SecureBio found that when people with limited biology experience were given access to large language models, which is the type of AI behind tools like ChatGPT, they were able to complete biosecurity-related tasks, such as troubleshooting complex virology lab protocols with four times greater accuracy. In some areas, these novices outperformed trained experts. Around 90 percent of these novices reported little difficulty getting the models to provide risky biological information, such as detailed instructions on working with dangerous pathogens, despite built-in safety filters meant to block such outputs.
In contrast, a study led by Active Site, a research nonprofit that studies the use of AI in synthetic biology, found that AI help did not lead to significant differences in the ability of novices to complete the complex workflow to produce a virus in a biosafety laboratory. However, the AI-assisted group succeeded more often on most tasks and finished some steps faster, most notably on growing cells in the lab.
Hands-on work in the lab has traditionally been a bottleneck to translating designs into results. Even a brilliant study plan still depends on skilled human hands to carry out. That may not last, as cloud laboratories and robotic automation become cheaper and more accessible, allowing researchers to send AI-generated experimental designs to remote facilities for execution.
AI systems are now able to run experiments autonomously and at scale, but existing regulations were not designed for this. Rules governing biological research do not account for AI-driven automation, and rules governing AI do not specifically address its use in biology.
In the US, the Biden administration had issued a 2023 executive order on AI security that included biosecurity provisions, but the Trump administration revoked it. Screening the synthetic DNA that commercial providers make to ensure it cannot be misused to make pathogens or toxins remains mostly voluntary. A bipartisan bill introduced in 2026 to mandate DNA screening does not yet address AI-designed sequences that evade current detection methods.
The 1975 Biological Weapons Convention, an international treaty prohibiting the production and use of bioweapons, contains no provisions for AI. The UK AI Security Institute and the US National Security Commission on Emerging Biotechnology have both called for coordinated government action.
The safety evaluations that AI labs run before releasing new models are often opaque and unsuited to capture real-world risk. Researchers have estimated that even modest improvements in an AI model’s ability to help plan pathogen-related experiments could translate to thousands of additional deaths from bioterrorism per year. Timelines for when these capabilities cross critical thresholds remain unclear.
The Nuclear Threat Initiative has proposed a managed access framework for biological AI tools, matching who can use a given tool to the risk level of the model rather than blanket restrictions. The RAND Center on AI, Security and Technology outlined a set of actions researchers could take to improve biosecurity, including improved DNA synthesis screening and model evaluations before release. Researchers have also argued that biological data itself needs governance, especially genomic data that could train models with dangerous capabilities.
Some AI companies have started voluntarily imposing their own safety measures. Anthropic activated its highest safety tier when it released its most advanced model in mid-2025. At the same moment, OpenAI updated its Preparedness Framework, revising the thresholds for how much biological risk a model can pose before additional safeguards are required. But these are voluntary, company-specific steps. Anthropic’s CEO, Dario Amodei, wrote that the pace of AI development may soon outrun any single company’s ability to assess the risk of a given model.
When used in a well-controlled setting, AI can help scientists quickly reach their research goals. What happens when the same capabilities operate outside those controls is a question that policy has not yet answered. Overreact, and talent and investment may move elsewhere while the technology continues advancing anyway. Underreact, and the risks of that technology could be exploited to cause real harm.
This article is republished from The Conversation under a Creative Commons license. Read the original article.
The post AI Can Now Design and Run Thousands of Experiments Without Human Hands. We Aren’t Ready for the Risk to Biosecurity. appeared first on SingularityHub.
