2026-11-12 23:00:00
AI models that learn like kids could reveal the secrets of our early years and inspire more efficient machine minds.
Babies are linguistic geniuses. Born into a noisy, chaotic world, they gradually distill incomprehensible sounds into words. Within two years, they can recognize thousands of words, follow complex sentences, and shout out their needs.
How they achieve this feat has long puzzled scientists. Babies develop language skills from an intimate world—the voices of their parents, chatter from strangers, nightly bedtime stories. A few hours of speech is enough to get them started.
This is a stark contrast to large language models (LLMs), which need to devour billions of words to mimic human conversation. Chatbots and AI agents are transforming how we interact with the digital sphere. But they guzzle water and energy, and signs that read “no data centers” now line neighborhood streets.
Building AI models that learn language the way babies do could reveal the secrets of our early years—including what happens when learning goes awry—and inspire more efficient machine minds.
An unusual study from Princeton University takes us a step closer. Using recordings from cameras and microphones from eight households, the team captured everyday sounds surrounding children from birth to roughly three years of age. These sonic diaries created a dense record of their waking lives. The researchers then built an AI learner for each child. Like a digital clone, each model experienced the child’s speech environment as it unfolded over time, rather than being trained on all the data in one fell swoop. The approach is a massive departure from the way engineers train LLMs.
“For more than a century, psychology has largely studied development from the outside,” said study author Uri Hasson in a press release. “Now we can take a complementary approach: Build a learning system, give it the child’s actual experience, and ask whether it develops the way the child does.”
Decades of research suggests babies respond to language long before they can explain what they know. A landmark study in the 1970s gave one- and four-month-old infants pacifiers that measured their sucking patterns and found they sucked faster when hearing a new speech sound, like shifting from a “b” to “p.”
At eight months, babies can detect words from a continuous stream of artificial speech, even without pauses between them. Somehow, they’ve learned to pick out statistical patterns in the sounds. And studies tracking babies’ eyes show their ability to recognize spoken words grows rapidly between 15 to 25 months, even before they can repeat them back.
More recently, scientists have eavesdropped on babies’ day-to-day lives using microphones and wearable recorders, linking emerging vocabulary to real-world audio and social interactions. Meanwhile, computational models that loosely mimic language development map some of the learning trajectory, though they’re often tailored to specific linguistic skills.
More powerful AI models—such as ones with the ability to “pay attention”—could better decode language acquisition. It’s now possible to “model learning under increasingly realistic conditions,” wrote Zara Harmon and Caroline Rowland at the Max Planck Institute for Psycholinguistics in the Netherlands, who were not involved in the study.
The new study, part of Princeton’s First 1,000 Days project, did just that.
Installing cameras and microphones in your home just after welcoming a newborn is a tough sell. The team developed an ethical framework that gave parents ongoing control over the right to withdraw or delete recordings. They ultimately recruited 15 families with diverse socioeconomic backgrounds, daily routines, and linguistic environments.
Each home had between 4 and 14 devices humming away for about 12 hours a day, collectively capturing over a million hours of audio and video. An AI pipeline tackled the data avalanche, identifying when babies appeared on camera and detecting speech.
After getting rid of noise, long stretches of silence, and overlapping conversations, the team extracted up to 6,600 hours of audio per child to train individual AI models, starting with Coral.
The AI began as a blank slate, with no access to online resources, predefined English sounds, or clues about where one word ended and another began. The relatively small model had two main components. An audio encoder broke speech into 20-millisecond chunks, while a predictive decoder learned to anticipate which sounds were likely to come next. (This is a bit like how hearing the first notes of a familiar song is enough for you to know what follows.)
Rather than training on all the recordings at once, the AI “experienced” Coral’s linguistic world day by day, in much the same way Coral did.
It didn’t work.
“We were surprised to see that, despite the immense amount of speech available from a single child, the learning results were discouraging,” said study author Hadas Raviv. “Simply exposing Coral’s model to the rich stream of daily speech was not enough.”
The fix was replay. Inspired by how the brain etches each day’s learning into long-term memory during sleep, the team gave the AI a chance to revisit snippets of recent speech at the end of each simulated day. “That was enough [to boost learning],” wrote the authors in a blog post.
With replay, the AI followed Coral’s learning trajectory. It first picked up the building blocks of speech—called phonemes—and eventually recognized recurring words. Learning started slow. But as Coral grew and spent more hours awake, both the infant and their AI twin were able to grasp new words at an accelerating pace. At around 1,000 days, their vocabulary had blossomed into thousands of words.
Across eight children, the models learned roughly 37 phonemes within 15 months and began recognizing words before age two, though at vastly different rates. To match each child’s learning trajectory, the team fine-tuned model settings, including how much speech was replayed each night.
“We’ve known for a long time that children’s developmental environments are complex,” said study author Casey Lew-Williams. “But now we can actually see this everyday complexity.”
That’s not to say the AI is a duplicate of a baby’s brain. It doesn’t process information in the parallel and energy-efficient way characteristic of neurons. And it learned solely with sound, whereas babies also draw on non-verbal cues—gazes, gestures, facial expressions, and interactions with others.
For the team, a key next step will be teasing apart which aspects of the model actually reflect infant learning, and which might be computational shortcuts that could inspire machine minds.
“The current model is only the beginning,” said Hasson. The accompanying video footage provides “an extraordinary amount of information about how children see, hear, interact, and act in the world. Our goal is to progressively give the models more of those capacities and see what emerges.”
The post This AI Learned Language From the Recordings of a Baby’s First 1,000 Days appeared first on SingularityHub.
2026-11-11 23:00:00
Some researchers say the findings show the glimmers of AI introspection—others aren’t so sure.
AI models can produce convincing descriptions of their reasoning, but there’s little way to tell whether those explanations reflect what’s actually happening inside. An emerging strand of research, however, suggests some models can detect when their internal activity has been tampered with, potentially suggesting some degree of self-reflection.
Whether AI can introspect has become a controversial question in both AI safety and cognitive science. If models could reliably describe what’s going on inside their own neural networks, it would give researchers a powerful new way to check what they’re really up to.
But there is widespread skepticism, and the idea is hard to test. Large language models are trained on vast amounts of text, much of which contains descriptions of human introspection. So AI has learned to talk convincingly about its inner life, regardless of whether those descriptions are grounded in anything.
However, a growing body of research now suggests models can tell when their neural activity has been altered by researchers. While it’s less clear whether they can detect exactly what’s changed, some researchers say the findings indicate at least a limited ability to reflect on their internal states.
“The fact that [an AI model] can notice the anomaly is insane,” Harvey Lederman, a philosopher at New York University and Anthropic, who co-authored a study investigating the phenomenon, told Quanta Magazine. “That is a mind-bending fact. It is wild, and it does seem robust to our various controls.”
Lederman’s study was published on the arXiv preprint server in March, while he was still at the University of Texas. The work is a follow-up to a 2025 report by Anthropic researcher Jack Lindsey, which pioneered the idea of injecting activity patterns associated with specific concepts into a model’s network.
In his experiments, Lindsey had found the best performing models could both detect injections and name the concept in question in about 20 percent of cases. Lederman and his colleague Kyle Mahowald, a linguist at the University of Texas, were skeptical. So they decided to test the idea on two of the largest openly available models, Alibaba’s Qwen3-235B and Meta’s Llama 3.1 405B.
They told the models that a researcher might be injecting “thoughts” into its mind and to let them know if it noticed anything. In trials where they artificially amplified patterns related to 821 concepts, Qwen flagged injections up to 53.9 percent of the time and Llama in up to 31.7 percent. In trials where they didn’t make any changes, neither model claimed a single injection.
The models sometimes named the injected concept correctly too—in up to 13.9 percent of trials for Qwen and 12.9 percent for Llama. But the pair found evidence this had little to do with the models knowing what had been injected.
For a start, wrong guesses bore no relation to the injected concept. Almost three-quarters of Qwen’s wrong guesses were simply “apple.” And when the researchers stopped amplifying a concept before the model began answering, correct guesses dropped sharply while detection rates held steady. That suggests the right answers came from a known side effect of the technique: Injecting a concept tends to make a model talk about it whatever the question.
The researchers concluded that the ability to detect tampering is distinct from whatever enables models to guess the concept. This idea was backed by a follow-up paper from Lindsey and colleagues published the same month that mapped out how anomaly detection works inside Google’s open Gemma 3 model. They suggested the mechanism emerges as raw models are trained into chatbots, and that the ability to identify concepts appears to rely on separate processes.
Not everyone is convinced the findings amount to introspection though. To probe the results, researchers at New York University revamped the experiment by simply telling a model to become obsessed with a particular concept rather than tampering with its network.
They then subjected Meta’s Llama and Google’s Gemma models to both concept injection and their new prompting trick and asked the models whether their internals had been interfered with. The models reported “injected thoughts” more often following the prompt than after a genuine injection.
When the researchers gave the models a third option to report that the prompt had been manipulated, they were unable to reliably differentiate between the two interventions. A model genuinely inspecting its own internal states should be able to spot that difference, the researchers argue.
The entire strand of research also raises thorny philosophical questions. Shauli Ravfogel, one of the authors of the New York University paper, points out that all a model’s workings run on its internal states, so almost anything it does could be dressed up as self-reflection. “If we say that this is introspection, the term ‘introspection’ becomes kind of meaningless,” he told Quanta.
Others argue the work cuts both ways. Steve Fleming, a University College London neuroscientist who studies introspection in humans, told Quanta the research suggests cognitive scientists may have oversold how clear the distinction between thinking and thinking about thinking really was.
Either way, the work is turning once purely philosophical questions about self-awareness into experimental ones. But we’re likely still a long way from settling the debate about what’s really going on inside the heads of today’s leading AI models.
The post AI Models Can Tell When Researchers Plant Concepts in Their Neural Networks. Is It Self-Reflection? appeared first on SingularityHub.
2026-11-11 00:18:21
“Self-driving” labs could free scientists to analyze results, refine hypotheses, and brainstorm ideas.
“Write ‘Intelligence is the ability to adapt to change’ into DNA.”
With that one prompt, the AI lab got to work. One agent scoured scientific literature to devise a workflow. Another translated it into code to control robotic arms. Still more agents guided robots to transfer tiny amounts of reagents and operate DNA synthesis and sequencing machines.
In around 200 hours, the unusual lab, called AutoDNA, completed the entire read-write cycle with a level of precision similar to expert scientists. Humans popped in occasionally to empty trash cans and replenish supplies. But for the most part, AI agents took the wheel.
Much of chemistry and molecular biology relies on complex but repetitive tasks. A tiny slip-up can torpedo hours, if not days or weeks, of painstaking work. The long hours, physical demands, and sheer boredom eventually takes a toll.
Offloading repetitive workflows to “self-driving” labs could free up time for scientists to focus on the bigger picture: Analyzing results, refining hypotheses, brainstorming ideas, and networking with peers.
AI systems have already been “tremendously useful in science,” wrote James Zou at Stanford University, who was not involved in the work. But most focus on computational tasks. Hands-on experiments involving liquids, chemicals, and biological materials present new challenges.
Rather than retrofitting a wet lab designed for humans, the team from Shanghai Jiao Tong University in China built AutoDNA from the ground up with AI in mind. Guided by an AI “coordinator,” robotic arms can expertly operate over 20 types of common lab equipment, autocorrecting any small hiccups along the way.
“It’s a teaser of what might be possible in the future,” outside expert Reshma Shetty, co-founder of Boston-based biotechnology company Ginkgo Bioworks, told Nature.
AutoAI is the latest attempt to automate pieces of the scientific discovery process.
Last year, Zou’s group introduced Virtual Lab, where an AI principle investigator leads a team of AI experts through research meetings to tackle a problem, with humans providing feedback. Within days, the team had used the system to design new proteins targeting the Covid-19 virus, some of which were later validated in wet-lab experiments by human scientists. Other systems aimed at speeding up research include FutureHouse’s Robin and Google’s Co-Scientist.
Teams have also pushed autonomy beyond the virtual world.
Scientists at the University of Wisconsin–Madison developed SAMPLE, a self-driving lab that uses a relatively simple AI model to suggest protein tweaks and then directs robots to make and test them. Results are fed back into the model for further improvement. Over a period of six months, the system built a more heat-tolerant enzyme, a feat that would have taken human scientists a year.
Another autonomous lab boosted enzyme activity using sophisticated protein language models to guide robotic experiments. And earlier this year, OpenAI and Ginkgo Bioworks showcased a GPT-5-powered lab that optimized protein synthesis from idea to end result, with minimal human help.
Scientists are also tackling another challenge: Getting pieces of lab equipment that speak different programming languages to “talk” to each other. Recently, Anthropic and collaborators released the Model Hardware Standard, a framework that lets AI agents control everything from microscopes to robotic arms and orchestrate experiments like a seasoned lab tech.
There’s no doubt self-driving labs are getting more sophisticated. But coordinating instrument clusters is one thing; building an entire molecular biology lab around AI agents is another.
AutoDNA relies on a team of agents powered by large language models, each handling a different task. A scientist first describes a task in plain language. The machine team then works out the best procedure and generates code to operate equipment (up to over 20 different types).
Guided by agents, a liquid handling robot mixes samples with reagents, shuttles them between instruments, and operates lab machines to synthesize, amplify, and sequence DNA. These seemingly routine tasks, performed hundreds of times or more by a scientist, take tons of practice to master without introducing errors or contamination along the way.
Several case studies proved AutoDNA’s prowess. In one, the system optimized a multiple-step DNA synthesis process known as polymerase chain reaction. It’s a commonly used method, but the recipe isn’t one size fits all. Scientists make subtle changes, depending on the length and letter makeup of the DNA sequence. The system tweaked reaction settings over multiple rounds, achieving a yield of nearly 98 percent, which is similar to that of human experts.
When challenged with finding mutations in a colorectal cancer sample, AutoDNA identified a candidate in KRAS, a gene that spurs cancers when mutated. It next designed and synthesized primers—short DNA strands that initiate the sequencing process—and used them to confirm the suspected mutation.
In another ambitious experiment, the team asked AutoDNA to encode and decode a message in DNA using the molecule’s four-letter alphabet. The operation required 11,382 operations across roughly eight days. Humans weren’t entirely out of the loop. Periodically, they stepped in to restock reagents and disposable labware—pipette tips or glass tubes—and empty trash.
“On detecting a shortage, the system paused the experiment and alerted the technician for replenishment,” wrote the team. In the end, all 78 synthesized DNA strands needed to encode the message were successfully synthesized and decoded, even with a few missing DNA letters.
AutoDNA showed a sort of common sense too. In one test to detect viral DNA, the planner agent instructed a robot to cap the test tubes to prevent samples from evaporating, without explicit instructions to do so.
“This is a small decision, but experimental success depends on many such details that might be obvious to an experienced researcher and are often left out of written protocols,” wrote Zou.
That’s not to say AutoDNA is ready for prime time. Manual restocking may sound trivial, but like food prep and dishwashing in cooking, it’s an integral yet time-consuming part of lab work. The system wasn’t foolproof, logging 14 hardware errors across experiments, 6 of which needed a person to fix. And it can’t yet handle more delicate experiments involving living cells or animals.
Then there’s safety. Even with knowledge about the physical world, AI agents may struggle to operate unfamiliar, expensive equipment. One solution might be to build a digital twin of the lab—a virtual sandbox that lets agents rehearse, refine their plans, and potentially catch errors before attempting experiments in the real world.
Biosecurity is another concern. Without sufficient safeguards, self-driving labs could synthesize DNA encoding potentially dangerous proteins. Building screening software directly into the system, designating access control, and keeping a ledger of who authorized what experiments could help reduce risks.
Still, “the work shows how AI systems can begin to close the gap between proposing an experiment and doing it,” wrote Zou.
The post AI Agents Ran a Biology Lab. Humans Popped in to Empty the Trash. appeared first on SingularityHub.
2026-11-07 23:00:00
The Smartest People in AI Believe This Is the World’s Most Important ChartBen Cohen | The Wall Street Journal ($)
“Anthropic refers to [recursive self-improvement] as ‘closing the loop,’ a cycle in which each improvement makes the machines better and faster at improving themselves. AI companies say they are now seeing the early signs of recursive self-improvement, which is speeding up their research—and freaking everyone out.”
An AI Agent Injected Drugs Into a Live Mouse’s BrainEmily Mullin | Wired ($)
“If you give an AI agent a robotic arm, an injector, and a live mouse, it might just perform brain surgery. Biotech founder Gabriel Ocana Santero recently did that using an agent to successfully carry out a brain injection procedure on a living mouse. Though it required some extra steps, it offers a glimpse into how AI could eventually become an active participant in laboratory research.”
Ramp Data Shows AI Spending Stalling Just as Anthropic Nears Its IPORebecca Bellan | TechCrunch
“A new report from Ramp shows AI spending flatlined over late summer even as token volumes rose, a sign that price competition and savvier customers may be capping revenue for OpenAI and Anthropic. According to Ramp’s data, token spend only grew 1.5% across August and September. To compare, in July, it grew 42% from the previous month. As Anthropic gears up for its IPO, the data suggests investors might be wrong to assume AI companies will keep growing at breakneck speed.”
The Tangled Web of AI Bets Is Reason to Fear Tech StocksJames Mackintosh | The Wall Street Journal ($)
“Circular financing would worsen any downturn. And buying shares in a company raising money in part to buy your product is as circular as it gets. The companies are essentially doubling down on AI bets, making AI stocks even more volatile because good news is better for them and bad news worse.”
I Watched a Humanoid Robot Clean My Apartment. It Felt Like a Taste of the Future.Grace Huckins | MIT Technology Review ($)
“Tau Robotics still has a long way to go: The amount of data needed to train robots to do the varied and complicated tasks of apartment cleaning is probably enormous, and the company has completed only a couple hundred cleanings so far. But it’s remarkable that there are now humanoid robots able to clean your toilets and mop your floors, even if they need quite a bit of human help to do so.”
Silicon Valley Promises AI Will Cure Disease. Drugmaker CEOs Aren’t Buying It.Xavier Martinez | The Wall Street Journal ($)
“The promises are grandiose: Anthropic’s chief executive said artificial intelligence could cure most human diseases within a decade. …Pharma executives don’t buy it. ‘That’s complete BS,’ said Jeff Leiden, executive chairman of Vertex Pharmaceuticals, recalling a recent conversation with a tech leader who claimed AI would cure 90% of human disease in the next seven years.”
This $749 Device Turns Seawater Into Drinking WaterAdele Peters | Fast Company
“If your city’s water supply gets shut off in a disaster, extra bottled water in your cupboard won’t last long. A startup founded by former SpaceX engineers proposes a different option today: a portable desalination device that can quickly purify almost any natural source of water nearby, including seawater.”
Orbital Matter 3D-Printed the First Rigid Beam in the Vacuum of SpaceTim Fernholz | TechCrunch
“A satellite built by the German startup Orbital Matter 3D-printed a half-meter long beam out of a custom plastic material, in what executives called a first for a free-flying spacecraft in orbit, and a step toward constructing large structures in space.”
Sony Shrinks Glasses-Free 3D Tech Into 6-Inch Desktop Character DisplayMonica J. White | New Atlas
“The ELF-MN1 is all about personal content: anime characters, avatars, and digital figures that appear to be standing right in front of you, with no headset or special glasses required. It hooks up to a smartphone or PC with a single cable, and Sony is pitching it as a way to bring 3D creations into everyday life on your desk.”
Google Energy Chief: Space Data Centers Cost Up to 100 Times More Than Those on EarthAmir Efrati | The Information ($)
“Launch costs for sending computing hardware into orbit are ‘incredibly, incredibly expensive, up to 100 times more expensive than building on earth,’ Peterson Corio said. Beyond cost, operators would also need to figure out how to maintain and upgrade equipment in orbit. On Earth, data center staff constantly swap out computers and rewire systems, she said.”
New Research Tests Whether AI Models Know Their Own ThoughtsBen Brubaker | Quanta Magazine
“After over a century of studying introspection in humans and animals, researchers are beginning to explore the phenomenon in unlikely new subjects—large language models. These systems, which are based on mathematical structures called neural networks, excel at cognitive tasks that until recently seemed exclusive to humans. Can they introspect as well?”
Watch: Robots Race Up 3,100 Ancient Steps on Clear Path to Beat US TechBronwyn Thompson | New Atlas
“The route—3.9 km, or just over 2.4 miles, with an elevation of around 600 m (0.37 miles)—would take an average hiker an hour or two. …But like humans, not all machine climbers conquered. The one that did, however, was a robot from Beijing’s Lingyi Technology, a subsidiary of Hangzhou’s famous Unitree Robotics. The four-legged robot scored top marks for speed, finishing the route in just 43 minutes and 41 seconds.”
Tiny Injectable Glucose Sensor Could Make Blood Sugar Monitoring 10 Times CheaperRachit Agarwal | Digital Trends
“[The sensor] relies on a protein that latches onto sugar and a near-infrared dye attached to a sugar molecule. When your glucose climbs, glucose kicks the dye-tagged molecule loose, and its glow changes. Near-infrared light travels through skin well, so a reader outside your body could pick up the signal without any electronics inside the implant.”
This Space Company Is Betting Its Future on a Propellant Available EverywhereEric Berger | Ars Technica
“The propulsion system is basically a small rocket that uses hydrogen and oxygen as propellants. However, instead of launching with separate tanks of liquid hydrogen and liquid oxygen, Chiron starts with liquid water. In space, the vehicle uses solar power to split water into hydrogen and oxygen gases and burns them to generate thrust.”
The post This Week’s Awesome Tech Stories From Around the Web (Through November 7) appeared first on SingularityHub.
2026-11-06 23:00:00
Five scientific experts discuss whether artificial intelligence is—or might one day become—truly intelligent.
This article was originally published on Undark. Read the original article.
Is it possible that AI is—or might one day become—truly intelligent? We put this question to five experts working in a variety of fields, including linguistics, computer science, cognitive science, and psychology. Not surprisingly, they offered very different answers, though if there was one point of agreement, it was that the term “intelligence” is vague and poorly suited for the conversation.
“Intelligence is a heavily contested concept,” Benjamin K. Bergen, a professor of cognitive science at the University of California, San Diego, wrote in an email to Undark. “Psychologists disagree on what it is.” Nevertheless, these five experts helped to parse just what our machines are doing—and what, if anything, separates AI from humans and from the natural world.
Some of these entries are drawn from email exchanges while others came from Zoom interviews. In some cases, they have been edited for length and clarity.
Emily M. Bender
Professor of linguistics, University of Washington
Setting aside the undefined terms “AI” (a marketing term, not a coherent set of technologies) and “intelligent” (a eugenicist way to rank people), can we ever create artificial systems that engage in cognitive behaviors or experiences similar to people? Probably not, and definitely not through language modeling (like LLMs). The only evidence we have that LLMs can “think” comes from our ability to make sense of the synthetic text they extrude.
Linguistics—the study of how languages work and how we work with language—provides two crucial insights here: The first is that languages are systems of signs, which pair form (the spelling or pronunciation of a word) with meaning (what the word means in general and what it is being used to mean). Language models manipulate form, but without reference to meaning.
The second lesson concerns how we interpret language. Far from simply unpacking meaning that an author packed into the words, we instead keep in mind everything we believe about the author’s state of mind, our common ground with them, and their beliefs about their intended audience’s state of mind. Against that background, we ask ourselves: What must the author be trying to convey by using those words and in that way?
In order to interpret text, we have to imagine a mind behind the text. We do this instinctively and reflexively—we can’t help ourselves, making rigorous evaluation of these systems difficult. As we wrote in the Stochastic Parrots paper in 2021: “Our perception of natural language text, regardless of how it was generated, is mediated by our own linguistic competence and our predisposition to interpret communicative acts as conveying coherent meaning and intent, whether or not they do.”
If we don’t account for this, we will not be able to make wise decisions about this technology.
Melanie Mitchell
Computer scientist and professor, the Santa Fe Institute
There’s this notion in AI circles that there’s this easily defined split between what people will call “cognitive intelligence” versus “physical intelligence.” When these big companies, for instance, like DeepMind or OpenAI, define what they’re calling “artificial general intelligence,” they talk only about cognitive tasks, as opposed to physical tasks. So, ChatGPT is not going to come fix your roof.
But it’s not clear to me how you can so easily separate the so-called cognitive kinds of tasks from the more physical kinds of intelligence that our brains were actually evolved to enable us to have. Or the more social kinds of intelligence that we humans rely on so much.
Undark: Can you say more about why it’s hard to draw a distinction between cognitive and physical intelligence? If I ask AI, “how do I fix my dishwasher?” it will give me instructions. Then I go and do the physical act of fixing the dishwasher. That seems like a clean distinction.
But in interpreting those instructions and applying them to your dishwasher, you’re adding quite a bit of the intelligence in that. If it says, “Press the red button on the left side of the door,” you—human—are using a lot of your intelligence to figure out what that means. Maybe there’s two red buttons, but one of them is obviously the one you’re supposed to press.
Undark: My experience with AI is that it’s quite bad at social intelligence.
One notion of social intelligence is that we can coordinate with other people to get things done: So, I don’t know how to fix my car, but I know how to find someone to fix my car. I might ask my friends, “Who do you know who knows how to fix a car?” And then also figuring out, what’s the right way to explain the problem to another person? I think of that as social intelligence.
I don’t know why AI systems are so bad at that. It may be that they don’t have the right training. It may be that they don’t have the right kind of participation in the social world.
Anima Anandkumar
Professor of computing and mathematical sciences at Caltech
Let me give my definition of intelligence: the ability to learn from data, and the ability to adapt to surroundings, or environment, based on that. You have now many AI systems that can do both.
You have AI that interacts with humans as chatbots and then adapts through that feedback—what we call reinforcement learning. These models can be made to align to certain goals. Could be doing more accurate mathematical proofs. Where it’s been most popular has been language models because we interact with those daily.
Language models mirror a lot of human intelligence, but I want to talk about nature’s intelligence. The ability to learn from microscopic scales—how atoms interact, how reactions happen, that enables us to create better drugs, better materials—to planetary scales and even cosmological scales: understanding how weather systems form. Is there a hurricane forming? How quickly will it make landfall? Can we understand the fine details of it to really get the predictions correct and do that early on? Because that has huge impact on saving human lives, creating evacuation policies early on.
We created the first AI-based weather model about five years ago that’s tens of thousands of times faster than what traditional forecasting could do, and now exceeds even the accuracy of traditional forecasts. So it’s faster, it’s more accurate, and we can create better predictions of extreme weather events and also assess risk. And now with changing climate, you could also make the systems predict longer-term climate, be able to really accurately incorporate the laws of physics, so that they can be accurate, even in newer conditions.
People focus so much on human intelligence. I want to talk about the intelligence that’s all around us—that’s invisible to us—but really, all of life is built upon it.
Alison Gopnik
Professor of psychology at UC Berkeley, author of popular books about babies and cognitive development
This question, and the phrase “artificial general intelligence” assume that there is a single thing called “intelligence” and that creatures, artificial or natural, have more or less of it. This is a common kind of explanation in folk theories and in early scientific ones. But as science progresses, explanations become more complex and cognitive science describes many varied, complex, cognitive capacities that serve different, and often conflicting, functions—many different “intelligences.” The intelligences that allow human children to learn so much are still far beyond current models.
The remarkable success of the current large models depends on the ability to detect statistical patterns in the text and pictures that humans have already produced at scale and produce new text and pictures based on those patterns. This allows human users to take advantage of the information previous humans have produced—these models are a technology for cultural transmission. But this is fundamentally different from the capacities that allow people to produce that information in the first place.
Another kind of “intelligence” is the “exploit” ability to take actions to achieve particular goals. This kind of intelligence is captured by reinforcement learning. AI systems use reinforcement learning to master skills like Atari games or coding, and it plays a crucial role in “post-training” the large models.
Cultural transmission and reinforcement learning play a role in human intelligence. But very young children actively explore a changing world and build new models of it—“explore” capacities that are at odds with “exploit” intelligence. They also rely on a very different kind of intelligence—the intelligence of care. Children can explore because other people take on the cognitively challenging task of caring for them. These capacities go beyond current large models, although different types of artificial intelligences might begin to approach them in the future.
Benjamin K. Bergen
Professor of cognitive science at UC San Diego
For most definitions of intelligence, modern AIs already display behaviors consistent with at least some degree of intelligence. They react to their environment, learn from data, and generalize beyond their training materials. By some measures, they already appear to reason logically and display sensitivity to the context and feelings of the humans they’re interacting with.
But whether these behaviors make them “truly intelligent” … depends, in part, on what you think behavioral evidence can actually tell you. Let’s say you’re trying to decide whether a horse is truly intelligent. When you pose arithmetic problems, the horse stomps its hoof the correct number of times to indicate the answer. Is the horse intelligent? Well, if the horse is able to understand problems articulated in human language, perform mathematical operations, and respond appropriately, you might reasonably conclude that it has some degree of intelligence.
But if you learn instead that the horse is actually just taking subtle cues from its trainer, who knowingly or not makes subtle movements when the horse has reached the correct number of stomps, this horse is displaying a different kind of intelligence. It has learned to stomp in a particular kind of situation, to attend to the trainer for subtle cues, and to stop appropriately. This is an example of learning from experience, generalizing to new situations, and perhaps sensitivity to the emotions of others in a social context.
But the lesson from this case is that intelligent-seeming behavior can be produced by many different kinds of underlying mechanisms, which may be different in the kinds or degree of intelligence that they engage. Similarly, AIs can answer questions that seem to require profound reasoning but which upon further inspection are actually spelled out in their entirety in the voluminous training data the model has consumed.
The post What Does the ‘I’ in AI Really Mean? appeared first on SingularityHub.
2026-11-05 23:00:00
Rather than relying on text alone, the algorithm looks for clues in a paper’s reasoning.
Science is no longer a purely human endeavor. Machine minds are increasingly merging with human creativity to generate hypotheses, experiments, and ideas.
But with promise comes misuse. Like TikTok videos and AI-generated images splattered across the internet, AI slop has infiltrated science. Academic journals and conferences are being flooded with AI-generated content, and editors are drawing the line. As long as they disclose it, scientists can use chatbots to help write papers, as long as the underlying ideas are their own. Submitting a whole paper dreamed up by AI, with little human input, is another matter.
Publishers are turning to AI detectors to weed out slop. But detectors aren’t perfect. Most scrutinize turns of phrase, word choice, and other intricacies associated with machine-generated text. They can mislabel legitimate research because of how the paper is written, rather than on the merit or origin of its ideas. That’s especially problematic for scientists whose first language isn’t English and who rely on AI for heavy editing.
A team at the University of Maryland in College Park is now taking a different approach: Strip away the prose until only a skeleton of concepts remains. Their detector, called IdeaLens, learned to find patterns in how human and AI models organize ideas, build narratives, and construct arguments, all of which are integral to every scientific paper.
IdeaLens alone won’t stem the deluge of AI slop polluting scientific literature. But it joins a growing toolset that recognizes human contributions, even as the boundary between scientist- and AI-generated ideas grows blurry.
“Discerning the extent to which AI performs critical thinking—a task that, in our view, rightly belongs to human authors—is increasingly difficult,” wrote editors of The Journal of Clinical Psychiatry, who were not involved in the new work. But, “using AI to deliberately fabricate data or automate the interpretation of study findings crosses ethical lines and undermines scientific integrity.”
There’s no doubt that AI has already made waves in science.
Algorithms are tackling particle and space physics and deciphering biology. Drug companies are betting AI can accelerate drug discovery and fight cancer, Alzheimer’s, and other medical scourges that continue to haunt millions worldwide.
Amid controversy, OpenAI recently announced an AI-powered breakthrough in fluid dynamics—one of the longest-standing, toughest puzzles in mathematics. While welcomed by some, the advance, alongside a barrage of other results, has triggered an “existential crisis” within the field and sparked debate over who deserves credit.
Generative AI has boosted research productivity over 89 percent, according to a study in late 2025. But there are growing pains. One estimate suggests roughly 12.5 percent of biomedical articles show signs of substantial AI involvement. And even AI conferences have struggled with AI submissions: Some papers accepted to the prestigious 2025 NeurIPs conference were flagged as likely AI-generated and 100 or more hallucinated citations slipped past human reviewers.
Publishers are pushing back.
The popular preprint server arXiv has tightened submission rules for computer science reviews, requiring prior peer review to combat the flood of submissions partially fueled by AI.
Others are using detectors. AI writing has particular quirks in word choice and sentence structure. In a well-known example, models tend to overuse constructions like, “It’s not X, it’s Y.” One popular tool, Pangram, claims 99 percent detection accuracy and is now integrated into alphaXiv, an arXiv mirror site, to automatically flag potential machine-written manuscripts. Meanwhile, scientists are turning papers into AI agents, making it easier to reproduce findings and spot questionable research.
Instead of focusing on wording, IdeaLens asks if the underlying concepts are human or AI.
The tool’s creators previously found AI-written fiction differs structurally from human work. Stories generated by machines tend to be tidier, more predictable, and more on the nose about their message. Humans write sloppier but more colorful tales. Beyond surface-level word choice or tone, the team found differences embedded into the narrative itself. For example, human writers often wander into subplots, let readers infer themes, explore morally debatable choices, and leave ambiguous endings. These detours create rich, lived-in worlds that reflect the messiness of real life. AI prefers straightforward narratives and neatly tied loose ends.
These findings inspired the team to develop IdeaLens, which looks past prose and focuses on structural and conceptual patterns.
The researchers first used Pangram to label roughly a million texts, such as news articles, blog posts, and academic papers, as human- or AI-generated. A large language model then stripped away the original wording, leaving a detailed outline of the main points. Scientific papers, for instance, were distilled to background information, claims, evidence, and arguments. Another model, called a classifier, then learned to find patterns that distinguished human- from AI-generated texts in these skeletal rewrites.
Like many modern AI systems, IdeaLens is a “black box.” The team can’t fully explain how it lands on verdicts, although they did find human-derived outlines tend to contain more study-specific details, rather than generic points often used by AI that could apply to other studies.
Still, it worked.
When challenged with documents written by ChatGPT and other large language models, IdeaLens mistakenly flagged them as human work just one percent of the time. For texts lightly edited by AI, it correctly recognized human origins in about 95 percent of cases. Across evaluation datasets, it missed roughly 10 percent of AI-generated work.
That might seem like a big problem. But “we would much rather have that AI document slip through than humans wrongly being flagged as AI,” study author Rishanth Rajendhran told Nature. “Because of the reputational damage that these things carry, it is important that we are careful with what we do with the prediction.” Once flagged, human reviewers should jump in for a closer look at the questionable submission.
Like other detectors, IdeaLens has blind spots. The team didn’t individually verify the origins of every document in its training data. The tool was also never tested on genuinely hybrid articles, where humans and AI models co-develop the underlying concepts, which is an increasingly common scenario in science. As scientists turn to chatbots to refine hypotheses or spark new ones, intellectual ownership becomes even murkier.
For now, IdeaLens sees the world in black and white, assigning just two labels: human or AI. Future versions may introduce more nuance, such as assigning the estimated percentages of their contributions. You can try it here.
The team stresses the detector shouldn’t discourage scientists from embracing AI. Rather, the differences it found highlight the value of human ingenuity. Without a dose of human messiness and out-the-box thinking, AI slop risks reshaping not just scientific literature, but our understanding of the world too.
The post Science Is Littered With AI Slop. A New Detector Could Filter It Out. appeared first on SingularityHub.
