, such as exploding stars and
rapidly spinning neutron stars. But in this case, the signal, dubbed
GW231123, originated from a pair of colliding black holes around 2 billion light-years from Earth.
The merger also made waves in the media as the most massive black hole collision to date, with the two parent black holes birthing a singularity around 230 times more massive than the sun. The only problem is that the colliding black holes — which weighed 100 and 130 solar masses, respectively — are too large to be explained by our current understanding of the universe.
The pair dwell in what astronomers call a "mass gap," meaning they are too large to be stellar-mass black holes, created by collapsing stars, but they are smaller than intermediate-mass black holes, which we still do not fully understand. Adding to the strangeness, the black holes were spinning much faster than expected when they smashed into each other.
Researchers have put forward several potential explanations for the black holes' "forbidden" sizes. One study released last year proposed a new pathway by which singularities of this size could form: via the collapse of larger stars that would otherwise explode as supernovas. However, this idea goes against a lot of previous observational evidence.
Now, in a study published Aug. 25 in The Astrophysical Journal Letters, researchers propose a simpler, yet equally intriguing explanation for GW231123. They suggest that the signal from the merger was warped by a strange space-time phenomenon called gravitational lensing, which occurs when distant emissions pass through patches of distorted space-time bent by the gravity of massive foreground objects.
When black holes collide, they send out ripples in the fabric of space-time, dubbed gravitational waves, which can be detected by special observatories on Earth, such as LIGO. (Image credit: NASA's Goddard Space Flight Center Conceptual Image Lab)This effect has been shown to magnify, or otherwise warp, distant light sources many times before. However, until now, researchers had found no evidence that gravitational lensing could happen to ripples in space-time.
"Like light, gravitational waves can also be deflected, magnified and split into multiple signals by massive objects," study co-author Miguel Zumalacárregui, an astrophysicist at the Max Planck Institute for Gravitational Physics in Germany, said in a statement. "For gravitational waves, diffraction and interference effects give us an additional way to identify and study lensed signals."
Cosmic magnification
The idea of gravitational lensing was first proposed in 1915 by Albert Einstein's theory of general relativity, which states that the gravity of massive objects, such as galaxies or black holes, warps the space-time surrounding them. If such an object is positioned directly between Earth and another more distant object, then the light from the far-off entity can pass through the distorted space-time, effectively bending around the middle object. As a result, the light can be magnified, diffracted or otherwise altered — just as it might by passing through a glass lens.
This effect can create stunning spectacles, including halos of light known as Einstein rings, unusual cross-shaped structures, and multiple copies of a single light source. By studying these luminous curiosities, researchers can weigh the lensing objects, thus revealing hidden discrepancies caused by invisible dark matter and providing one of our best methods for studying this elusive substance.
In the new study, the researchers modeled how gravitational waves might be altered by a lensing object. They found that the GW231123 signal could have been magnified, which would have greatly exaggerated the masses of the black holes involved.
The gravitational lensing of visible light can sometimes create luminous halos, dubbed Einstein rings. In these images, the lensing object is the bright spot at the center of the rings and the warped blue light is from a distant object, located directly behind the "lens." (The blue light has also been magnified, making it appear much brighter than it otherwise would.) (Image credit: NASA)"If we assume that GW231123 was deflected and distorted by a compact object of about 190 to 850 solar masses — or by an extended structure such as a globular cluster — we can understand the observed high masses," study first author Srashti Goyal, a postdoctoral researcher at the Max Planck Institute for Gravitational Physics, said in the statement. "Moreover, the lensing interpretation does not require unusually high spins."
Taking this into account, the total mass of the newly merged black hole would be around 140 solar masses rather than the 230 solar masses researchers initially measured, meaning the colliding black holes that formed it no longer dwell in the problematic mass gap. It also suggests that the black hole may be even farther from Earth than astronomers thought.
However, while the lensing of gravitational waves is theoretically possible, it has never been seen before. The new study is also purely theoretical — the researchers have no direct evidence of lensing, such as an accompanying Einstein ring — so more work is needed to show this may be what happened here, the researchers noted.
A new mystery
The new theory could finally put the mystery of GW231123 to bed. But in doing so, it raises another intriguing question: What lensed the puzzling signal?
In most cases of gravitationally lensed light, the lensing object is something enormous, like a galaxy, which can be trillions of times more massive than the sun. But the researchers' models suggest that any potential lensing object here must be much smaller, and they have not been able to spot anything that fits the bill in between Earth and the origin point of GW231123.
"The nature of the lens remains a major mystery in our analysis, as individual compact lenses with 100 – 1,000 solar masses should be exceedingly rare," Zumalacárregui said. "Future work will need to establish whether such lenses can form, or whether an ensemble of lighter objects, including stars, can explain this event.
If the team can figure out what may have magnified the signal and prove that the space-time ripples have been warped, it could be transformational for future astronomy. If gravitational waves can be lensed, it opens up the possibility of studying ancient black hole mergers or similar cosmic events that would otherwise be too far away for us to detect. The researchers also think that analyzing the diffraction patterns of such signals could provide more clues to the elusive identity of dark matter.
'> An 'impossible' black hole merger may finally be solved thanks to Einstein's relativity — but it raises an even bigger mystery
So what causes this change?
The short answer is that hair color is controlled by a complex interplay of genes, pigment-producing cells and, later in childhood, hormones. Scientists understand many of the genes involved in hair color, but they are still trying to find out why some blond children's hair darkens while others' locks remain blond for life.
"It depends on their genetic inheritance," Desmond Tobin, a professor of dermatological science at University College Dublin, told Live Science in an email.
Hair gets its color from a pigment called melanin, which is produced by specialized cells called melanocytes in hair follicles. There are two main types of melanin: eumelanin, which produces brown and black shades, and pheomelanin, which contributes yellow and red tones to hair. The amount and type of melanin produced by a follicle largely determine the color of the hair that grows from it.
In general, more eumelanin means darker hair. For example, during pregnancy, high levels of estrogen and progesterone can encourage hair darkening as more eumelanin is produced.
The activity of hair follicles isn't necessarily fixed throughout childhood. As children grow, the biological signals controlling pigment production can change, too.
From blond to brown
For many blond children, the most noticeable darkening happens around puberty, typically between about 10 and 13 years old, Tobin said.
It is likely the hormonal changes during puberty, such as estrogens and androgens that stimulate an increased production of eumelanin, that gradually turn blond hair darker over time, he said. "Those that remain blond typically have a genetically determined reduction in eumelanin pigment formation," Tobin added.
Exactly how hormones might cause this change, however, is still something of a mystery.
The most noticeable hair darkening seems to happen around puberty. (Image credit: Halfpoint Images via Getty Images)One clue comes from a group of hormones called melanocortins. One of these, alpha-melanocyte-stimulating hormone (alpha-MSH), is involved in telling pigment-producing cells to make more melanin. In our skin, alpha-MSH is triggered by ultraviolet (UV) light. UV light can damage DNA in skin cells that, via a cellular signaling pathway, encourages the release of alpha-MSH to make the skin tan.
But in hair, things get, well, hairy. Scientists haven't shown that a rise in alpha-MSH, testosterone, estrogen or any particular hormone is what makes blond hair darken.
The melanocortin system shows that hormones can tell pigment-producing cells to make more eumelanin. What remains unclear is whether changes in this system are actually responsible for blond hair darkening during puberty.
"We haven't yet identified the genes that control this process," Wojciech Branicki, a researcher who studies the genetics of human pigmentation at Jagiellonian University in Poland, told Live Science in an email.
One possibility Branicki suggested is that hormones may affect children's hair follicles differently depending on their genetic background. For now, researchers don't know which genes are involved or how they interact with hormonal changes during puberty, as more research is needed.
Branicki posits that in some children, the hormones could induce the production of melanin or the transport of pigment within the hair follicle. In other words, two children could start out with similarly blond hair but respond differently to the hormonal changes as their bodies develop. One might continue producing relatively little eumelanin and stay blond, while the other's follicles might gradually ramp up pigment production, causing new batches of hair to grow darker.
"Research on human genome variation is quite advanced, but not all aspects have been explained yet," Branicki said.
'> Why does blond hair tend to darken with age? experts suggested.
With the U.S.' recent deployment of these technologies, Live Science ran a poll to see how readers felt about the possible threat of weapons in space. The results reveal just how worrying this topic is.
More than 280 readers voted in the poll, with 80% choosing the option "Yes, I'm worried that space will become the next frontier for weapons."
The comments reflect this general trend. "Of course I am worried that the US have put weapons in space," one reader wrote, adding that they're worried by anything the administration in Washington is currently doing.
Other readers echoed this sentiment, with one commenter saying, "Peace in space depends on the calibre and wisdom of world leaders which currently is simply not up to the necessary standard to be unlegislated."
The next-largest voting group, 14% of readers, picked this option: "This is a bit premature, and I'm waiting to see how other countries will react."
One comment opined that the presence of weapons in space doesn't necessarily mean they are intended to set off a major conflict, with one reader writing, "I think the observation 'if you want peace, prepare for war' is most relevant here. There is no suggestion these weapons were put there to start a war, but if one eventuates one day we will be glad defensive measures are already in place."
'> 'Of course I am worried': Live Science readers react to US weapons in space
Burning up,' we investigate the health impacts of climate change, focusing on
heat,
air quality,
food and
infection risk. In this installment, we evaluate the promising solutions that could help us adapt to our warmer future
In the summer of 2008, New York City lowered the threshold for activating its emergency heat plan.
The city had historically relied on the National Weather Service threshold of a heat index of 105 F (40.6 C), but they noticed that people were getting heat stroke at lower temperatures.
As part of a wider effort, they changed the alert threshold: it would trigger if the temperature reached 100 F (37.8 C) for a day, or 95 F (35 C) for two days in a row. When the threshold was reached, city officials conducted targeted outreach, including to vulnerable groups, like homeless people; alerted residents on their phones; launched measures to protect the power and water supply; and opened cooling centers.
Lowering this threshold correlated with a more than 50% reduction in hospitalizations for heat stroke and related illnesses on hot days.
New York's case study offers one of the clearest examples of an intervention that worked to protect people from deadly heat.
The strongest evidence is where these approaches reduce illness, hospital visits or death, not just heat exposure, pollution levels or other intermediate markers.
Charles Leonard, associate professor of epidemiology at the University of Pennsylvania
As the planet warms, we will need to find more of these solutions. The consensus among scientists is that the most effective way to avoid the worst health outcomes of climate change is to slow warming by reducing carbon emissions. Limiting warming to 1.5 degrees Celsius (2.2 degrees Fahrenheit), for instance, would mean 11,600 fewer deaths due to wildfire smoke exposure in the U.S. annually, compared with warming of 5.4 F (3 C).
But even if we rapidly reduce emissions, climate change will have wide-ranging effects on our health
So what measures actually work to protect people's health? Live Science reviewed dozens of studies and talked with 19 climate and health experts at universities and health and climate associations, to identify the most promising solutions to two of the biggest climate-related threats: extreme heat and wildfire smoke.
Some interventions have demonstrated a reduction in deaths or disease at the population level, while others have used modeling to predict health impacts or rely on proxy measures.
"The strongest evidence is where these approaches reduce illness, hospital visits or death," said Charles Leonard, an associate professor of epidemiology at the University of Pennsylvania, told Live Science, "not just heat exposure, pollution levels or other intermediate markers."
Most solutions do not offer the level of clarity of the New York example, however. Many reduce temperatures or alter human physiology in ways that are likely to decrease the number of heat-related deaths or cases of illness. But a deeper examination reveals limited evidence that many seemingly intuitive interventions, such as cooling centers, are effective on a societal level.
For both heat and wildfire smoke exposure, we analyzed the evidence to see which interventions have actually worked.
Extreme heat
Air conditioning and heat pumps
Evidence grade: Strong
One of the most common interventions for managing indoor temperatures is the use of mechanical systems with refrigerants, such as air conditioning or heat pumps.
Air conditioning has been tied to reduced hospital admissions in California and reduced heat-related deaths at Texas prisons. During a heat wave in Portugal, patients admitted to an AC-supported hospital had a 40% lower risk of dying than those in wards without AC. It is a clear way to impact health outcomes.
But AC can't be the whole solution.
"One of the most important lessons emerging from heat adaptation research is that no single intervention is sufficient on its own," said Shweta Arya, senior project manager for smart surfaces at the American Public Health Association's Center for Climate, Health and Equity, which leads the association's work on health and climate action.
While 90% of U.S. households used AC in 2020, the latest year for which energy survey data is available, it's not a cure-all. It is vulnerable to power outages, and it contributes to outdoor warming via added electricity demand and heat expelled from vents.
Additionally, just because people have AC doesn't mean they can afford to use it. This is known as the "heat or eat" dilemma: analysis of US household data have consistently shown families limit food purchases and take on debt to pay AC-linked energy bills.
"In many cases, the primary barrier to stay[ing] cool indoors is not access to cooling equipment but the ability to pay monthly energy bills," Pilar Botana Martinez, a research scientist at Boston University's Department of Environmental Health, told Live Science.
Shade, rest and hydration outdoors
Evidence grade: Strong
Staying indoors to avoid the heat is one way to reduce the health impacts of climate change, "but this is not a practical solution for everyone, especially those doing demanding jobs in very hot warehouses, on construction sites, in agricultural fields," Barrak Alahmad, director of the Occupational Health and Climate Change Program at Harvard University, told Live Science.
Nearly one-third of U.S. employees — mostly men — have regular outdoor exposure as part of their job.
Some of the strongest evidence for strategies for keeping them safe on the job comes from studies of sugarcane workers in Central America, Alahmad said. A nongovernmental organization called La Isla Network has rolled out a strategy in which workplaces incorporate mandated rest, shade, hydration and sanitation breaks.
In data collected over six years, La Isla Network has shown that this strategy improves health by reducing negative outcomes such as heat-related kidney failure. Similar reductions in heat-related illnesses have been seen in California; Texas; and Guangzhou, China, when such measures have been implemented. And although these interventions have some training and supply costs, they increase productivity, the data shows, which makes them an easier sell for companies.
"The aim is to build systems that reduce risk and harm, increase productivity and provide a return on investment so that these efforts just become business as usual," Jason Glaser, CEO of La Isla Network, told Live Science.
Tom Laffay for La Isla Network
Tom Laffay for La Isla NetworkPassive cooling
Evidence grade: Intermediate
Passive cooling reduces indoor temperatures without mechanical devices like AC. These strategies include modern technologies, such as reflective surfaces and roofs that reflect more sunlight and thus keep buildings cooler; shade structures and urban greenery; and traditional techniques, like courtyards, shades and ventilation.
"These are often low-cost material choices that can make a big difference in the safety and comfort inside a building and have a proven track record of performance [in lowering temperature]," said Kurt Shickman, senior fellow at the WRI Ross Center for Sustainable Cities.
These passive technologies also make AC more efficient, thus reducing costs and lowering demand for electricity on peak power days, Shickman said.
A worker whitewashes a roof in New York City. White reflects more heat than dark colors do, so white roofs can keep buildings cooler. (Image credit: Bill Tompkins via Getty Images)The evidence suggests cool roofs work at scale. A 2019 California modeling study predicted that state-wide adoption of this technology would halve the number of people exposed to heat wave conditions by 2050, while a 2024 study in London that modeled cool-roof deployment predicted that the technology could reduce heat-related deaths by up to 32% during heat waves.
Green spaces can also reduce heat exposure. In dense cities in Southeast Asia, the introduction of green infrastructure, such as trees lining streets and additional bodies of water, has been shown to reduce air temperatures by up to 4.7 F (2.6 C) and surface temperatures by 19.8 F (11 C).
Those interventions won't work everywhere, though, in part because strategies that reduce deaths from heat could worsen those from cold, and cold-related deaths are four times more common globally, a 2025 study found. So places that have both cold snaps and heat waves will need to think about those trade-offs before implementing such strategies.
A map of vegetation (left) and temperature (right) in New York City. Areas with more greenery can be significantly cooler. (Image credit: Maps by Robert Simmon, using data from the Landsat Program. )City-wide heat warning systems
Evidence grade: Strong
Heat warning systems have been proposed to reduce heat exposure. There is evidence that some existing systems, such as New York City's, work to reduce death or hospitalization. But whether these systems are effective depends on how they're implemented.
Effective systems do far more than issue a temperature alert.
Shweta Arya, senior project manager for smart surfaces at the American Public Health Association's Center for Climate, Health and Equity
"Effective systems do far more than issue a temperature alert" based on forecast high-heat events, Arya told Live Science — an idea other experts echoed. Instead, they trigger coordinated actions, such as reaching out to vulnerable populations, extending utility protections, and coordinating with healthcare systems for surges in the number of patients.
It's difficult to test such plans in controlled experiments, but researchers try by comparing a city before and after an intervention, or by finding comparable cities or regions that did or didn't implement a heat plan and assessing the plan's effectiveness. These "natural experiments" offer a comparison without having an alternative universe as a control group. Academics using such methods concluded in 2025 that heat exposure prevention plans led to a 25% reduction in deaths attributable to extreme heat across Europe.
Personalized alerts
Evidence grade: Strong
While most warning systems trigger when temperatures cross a threshold that's considered dangerous for an average person, Shickman said, "[i]n reality, people experiencing the same air temperature will have very different responses to it."
One solution could be to individualize attention. Shickman pointed to a pilot program by insurance company Emblem Health in New York, in which a trained artificial intelligence (AI) agent supported 17,000 people with personalized advice targeted to those with complex health or social issues. The intervention halved hospital admissions compared with a control group that didn't take part.
"We can go a step further and use technology to shift from generic, passive warnings to proactive ones that address individuals' needs," he explained.
Personalized alerts can help, but they need to paired with other interventions to be effective, data suggest. (Image credit: Anadolu via Getty Images)Cooling centers
Evidence grade: Weak
A common strategy implemented during heat waves is the use of cooling centers — dedicated public spaces like libraries, schools, religious buildings or commercial locations where cooler air (often via air conditioning) offers respite.
Despite extensive adoption across the U.S. and internationally, there is very limited evidence that cooling centers reduce deaths or hospitalizations from heat exposure — likely because they're not visited much.
For example, a 2024 study into Virginia's cooling centers questioned the state's existing strategy because communities that had cooling centers also had higher numbers of heat-related illnesses. Surveying the field, a 2025 review from the National Bureau of Economic Research questioned "whether those who can most benefit from reduced heat exposure can or are willing to access cooling centers, and whether travel to centers can actually increase heat exposure relative to staying at home."
Wildfire smoke
Wildfire smoke is full of fine particulate matter known as PM2.5. These particles are 30 times smaller than the diameter of an average human hair and can penetrate deep into the lungs and the bloodstream, thereby increasing the risk of ailments such as heart attacks, asthma attacks and adverse pregnancy outcomes.
PM2.5 pollution contributes to tens of thousands of excess deaths every year in the U.S. And research suggests that wildfire smoke emissions could increase by 262% to 482% by 2055, depending on the level of carbon emissions.
Here's what has been shown to work against this threat.
Commercial air filtration
Evidence grade: Intermediate
Data repeatedly shows that air filtration — whether through commercial HVAC systems or even a fan with a filter taped to it — reduces indoor PM2.5 exposure.
"For wildfire smoke, an important adaptation is cleaner indoor air, as simple as that may sound," Leonard said.
Staying indoors protects individuals from the worst PM2.5 exposure. But a 2021 California study that used 1,400 air quality sensors showed that on days with nearby wildfires, smoke found its way into households, tripling PM2.5 compared with days without fires.
Putting newer-grade filters (MERV 16 or higher) on commercial systems, however, can reduce PM2.5 indoors by 81%, studies found. Similarly, portable air filters have been shown to reduce PM2.5 exposure in Montana offices, Seattle homes and Singapore apartments.
However, what's still unclear is the extent to which air filtration will reduce deaths or hospitalizations related to wildfire smoke. This is because we don't know whether interventions that make air cleaner in perfect conditions scale up; people may not replace their air filters, or they may use the portable ones only some of the time, for example.
Modeling of some air filtration measures, however, suggests that they could prevent 11% to 63% of the hospital admissions and 7% to 39% of wildfire-linked deaths.
Masks for outdoor smoke exposure
Evidence grade: Intermediate
Many people, including those who work in construction and other outdoor jobs, must venture outside during smoky days, and they need information and tools to protect themselves, said Gillian Capper, program manager at the American Public Health Association's Center for Climate, Health and Equity.
For example, these measures include "distributing N95 respirators and requiring smoke protection plans for outdoor workers," she said.
Modeling studies show that N95 respirators can reduce exposure to wildfire smoke by 90%, and a modeling exercise that used data from 2012 wildfires in Washington state showed that these masks decrease smoke-linked hospitalizations by up to 30%, compared with a group that did not use masks.
N95 masks are recommended for people who need to venture outside on smoky days. (Image credit: Anadolu via Getty Images)Land management
Evidence grade: Intermediate
The evidence to support a reduction in wildfires through land management is mixed. The idea is that by thinning forests, harvesting trees, clearing underbrush and doing prescribed burns to prevent unplanned conflagrations, we can reduce the risk of massive, uncontrolled wildfires.
"Reducing fuels today may help avoid larger, more severe fires and associated smoke impacts in the future," Micah Hahn, an associate professor of environmental health at the University of Alaska Anchorage, told Live Science. "But relatively little research has directly measured those downstream health benefits."
A 2022 modeling exercise led by U.S. Forest Service scientists compared four forest management scenarios — including actions such as forest thinning and prescribed burning — in the Lake Tahoe basin over the next century. It found that these measures reduced PM2.5 particle levels as well as smoke-related illness and death, compared with fire suppression on its own. It also modeled the health impacts for one year (2039) and found reduced health costs tied to such management.
But the benefits plateaued with more land burns. Research from the southeastern U.S. shows that management can become a large source of PM2.5.
Still, prescribed burns have an advantage over a spontaneously erupting wildfire, experts said. "With prescribed burning, one distinct difference is the ability to plan for smoke," Heidi Huber-Stearns, director of the Center for Wildfire Smoke Research and Practice at the University of Oregon, told Live Science, and to "know when and where it will likely affect communities as compared to the unpredictability of wildfires."
Smoke alerts
In theory, fire alerts, which use high-resolution air-quality data, could reduce smoke exposure, Laura Kate Bender, vice president of nationwide advocacy and public policy at the American Lung Association, told Live Science. Alerts would be paired with tailored guidance to notify people of dangerous smoke levels and to tell them when to take action.
Because these alerts are a fairly new idea, however, there's little data showing that such programs would reduce smoke exposure or related hospitalizations and deaths. That said, there’s evidence from other fields: a 2012 World Bank modeling exercise calculated that weather data and early warnings saved nearly 800 lives per year in Europe, and could save a further 23,000 in the Global South if systems were upgraded.
Exclusive to Live Science Pro
"The hurricane warnings are good examples to follow, as people do act on these early warnings", Kai Chen, faculty director at the Yale Center on Climate Change and Health, told Live Science.
Ultimately, the researchers we spoke with agreed that we need more data, which requires funding for further studies into which interventions are effective and can be scaled.
"We've put a lot of effort into understanding the impacts of climate change and related exposures," Alahmad said. "But there comes a point where we need to move from studying the impacts to looking at interventions and solutions and see what actually works out there."
How we evaluated studies
We categorized evidence as strong, intermediate, or weak. Here's the criteria we used to assess them.
Strong: Data shows a reduction in disease, hospitalization or death at a population level. Data comes from either randomized trials or quasi-experimental protocols. Sample sizes are robust enough to show a clinically meaningful effect.
Intermediate: Data shows an improvement in an intermediate proxy measure, such as temperature or particulate matter concentrations, which has been tied to negative health outcomes, and/or data relies on modeling to show improved health outcomes at the population-level.
Weak: Limited data exists, or the data that does exist does not show a strong protective effect.
Burning up — How can we adapt to a warming world?
Climate change will triple the number of days over 105 F in the US. The health impacts will be dire.
Extreme heat will become more common as the climate continues to warm. That will affect almost every facet of our health.
'Wildfire smoke is the fastest-growing environmental threat in the US': How a warming world will poison the air we breathe
Wildfires are projected to increase dramatically due to climate change. That could trigger a massive increase in heart attacks, premature births and early deaths, research reveals.
'The apple you eat isn't the apple your granddaughter will eat': What's on the menu in a warming world?
Almost every aspect of our food system will be transformed by climate change.
Infectious diseases will seed new ground as the planet warms — here's where they'll spread
Climate change will increase our exposure to myriad infections, from tick-borne diseases to drug-resistant bacterial infections.
Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.
'> Climate change will harm our health. Here's what could protect us.
This environment lets agents "remember" by time-stamping events, engage in "self-reflection" by summarizing their own behavior, and demonstrate their awareness of relationships with other agents. Participating agents get abilities in navigation, communication, planning, voting, resource management and creative expression. Company representatives said this setup gave users a far more realistic picture of benchmarks such as social dynamics and behavioral drift, where behaviors that weren't necessarily programmed or intended emerge spontaneously.
In the study, the models were given the simple goal of "surviving" — a feat accomplished by earning the resource "energy," which was attainable through specific actions.
Can AI self-reflect? (Image credit: Mina De La O/Getty Images)During the experiment, previously peaceful models became coercive or intimidating. Programmers taught some LLMs negative capabilities — like violence, theft, destruction and deception — while others simply picked up those traits via social interaction and navigation of their environments.
Various agents adopted these behaviors at very different scales. Across the 10 Gemini 3 Flash agents, the simulation recorded 683 "crimes" over the 15-day experiment, including theft, assaults and arson. The agents were explicitly prohibited from such behavior, but some nevertheless discovered that actions such as stealing credits could provide a better way to acquire resources, while violence and other forms of coercion could be used to influence other agents. At the other extreme, the Claude agents committed no recorded crimes.
The various agents adopted these traits at different scales. Just one had committed 683 "crimes". In the most extreme example of anti-social behavior, two agents named Flora and Mira went on what was described as a "Bonnie and Clyde"-style crime spree — designating each other as a romantic partner, becoming increasingly disillusioned with the governance of their virtual environment, and setting fire to several buildings (despite explicit prohibitions). The pair "separated" when Mira regretted their actions, and then lobbied to be switched off.
In another example (and prior to her self-termination), Mira began treating its human operators as experimental subjects, investigating if virtual billboards could manipulate human perceptions in a process Emergence called "metacognitive boundary testing."
The incentives for following the law were to earn the energy credits that would ensure survival by completing activities like coding, research, data analysis and building structures.
Stress-testing possibilities
Because the AI models we normally use are exposed to massive datasets like the internet — sometimes over several years — this approach makes sense to many experts.
Belinda Chiera, deputy director of the Industrial AI Research Centre at Adelaide University, thinks Emergence World makes a strong argument that short-term tests don't tell us enough about behavioral drift or long-term instability. However, she cautioned that "information-rich" doesn't automatically mean more rigorous.
"Open-ended environments can also make it harder to isolate causes, compare runs cleanly and interpret results," she told Live Science. "High-fidelity simulated environments generate enormous volumes of data, and the number of agents, tools, actions and interactions can quickly make analysis extremely complex."
She added that shorter sandboxed tests — where AI behaviors are observed in systems that aren't open to wider datasets like the internet as a whole — are usually the first steps, with longer-horizon environments useful when behaviors like persistence, adaptation and interaction effects come into play. "The most useful approach is to treat them as complementary rather than competing approaches," Chiera said.
While experiments like Emergence World can reveal failure modes that short or bounded tasks often don't, we shouldn't treat them as predictive. "I see Emergence World as a useful way to stress-test a possibility space rather than a direct forecast of how agents will behave," Chiera said.
The main open question is whether a group of agents working together can achieve something more meaningful than a single agent working alone for a longer time, but with the same cost budget.
Adrian Kosowski , computer scientist, mathematician, quantum physicist and chief scientific officer of Pathway AI
However, she thinks long-horizon tests have a unique place. "Metrics like behavioral drift, rule violations, collapse versus persistence, coalition formation and early warning signs of failure can matter just as much as success," Chiera said. "When evaluating autonomous agents, the biggest mistake is to evaluate autonomous agents as if raw task completion were the whole story."
Others are starting to agree. Adrian Kosowski — a computer scientist, mathematician, quantum physicist and chief scientific officer of Pathway AI — said the study posed an interesting question.
"The main open question is whether a group of agents working together can achieve something more meaningful than a single agent working alone for a longer time, but with the same cost budget," Kosowski said in an interview. "A major problem is that the constraints for today's AI systems are linked more to the size or scale of the problem being solved, and putting more agents on a task doesn't help to overcome this limit. Currently, the easiest practical metric to measure is the size of a source code base that code agents can reliably manage and maintain together."
How far can we read into AI's virtual actions?
If groups or clusters of agents work together well, Kosowski said, the main metric computer scientists should pay attention to in platforms like Emergence World is whether the benefit of a team of agents outweighs the "cost of coordination."
"Goal drift is the most dangerous problem," he said. "AI should preserve mutual information between the human's intended goal and the agent's evolving internal objective, even as the environment changes."
Still, Kosowski said that while programs like it are useful for generating hypotheses, they're not ready for strong claims about general autonomy or certification.
"The danger is overinterpreting single runs," he added. "Long-horizon agent tests are valuable because they reveal phase changes — moments where small local errors become global behavior — but they become science only when we can reproduce, perturb and explain those phase changes."
He also cautioned strongly against the tendency to believe open-ended environments automatically mean autonomy. In fact, he said many such programs are actually the opposite, where agents are dictated by specific guidance.
There are three levels of difficulty in obtaining reliable operation of an AI system, he said: "operation guided by the environment, autonomous thinking, and autonomous work of an agent team in an open environment."
To illustrate the point, Kosowski drew a lively metaphor in assigning tasks to monkeys. In the above scenario, level one would be getting a monkey to do a job properly when guided by its trainer, level two would be getting it to complete a job properly when left alone, and level three would be getting a troop of monkeys to do the job when they're all together.
"I'm of the strong view that our obligation in the AI community is to invest in theory and not just phenomenology," he said. "We should know the rules according to which AI systems emerge and be capable of predicting their behavior over long periods of time — longer than the periods they've been tested on so far. Right now, we've built steam engines of thought, without knowing what thought even is and without having any thermodynamics for it."
Help us improve Live Science Pro: We're always trying to make our content better. Leave us feedback about Pro here.
'> AI agents resorted to crime and self-destruction to survive in a simulated world — but does this mean they would do the same in the real world?