How robots learn: A brief, contemporary history

Roboticists used to dream big but build small. They’d hope to match or exceed the extraordinary complexity of the human body, and then they’d spend their career refining robotic arms for auto plants. Aim for C-3P0; end up with the Roomba. 

The real ambition for many of these researchers was the robot of science fiction—one that could move through the world, adapt to different environments, and interact safely and helpfully with people. For the socially minded, such a machine could help those with mobility issues, ease loneliness, or do work too dangerous for humans. For the more financially inclined, it would mean a bottomless source of wage-free labor. Either way, a long history of failure left most of Silicon Valley hesitant to bet on helpful robots.

That has changed. The machines are yet unbuilt, but the money is flowing: Companies and investors put $6.1 billion into humanoid robots in 2025 alone, four times what was invested in 2024. 

What happened? A revolution in how machines have learned to interact with the world. 

Imagine you’d like a pair of robot arms installed in your home purely to do one thing: fold clothes. How would it learn to do that? You could start by writing rules. Check the fabric to figure out how much deformation it can tolerate before tearing. Identify a shirt’s collar. Move the gripper to the left sleeve, lift it, and fold it inward by exactly this distance. Repeat for the right sleeve. If the shirt is rotated, turn the plan accordingly. If the sleeve is twisted, correct it. Very quickly the number of rules explodes, but a complete accounting of them could produce reliable results. This was the original craft of robotics: anticipating every possibility and encoding it in advance.

Around 2015, the cutting edge started to do things differently: Build a digital simulation of the robotic arms and the clothes, and give the program a reward signal every time it folds successfully and a ding every time it fails. This way, it gets better by trying all sorts of techniques through trial and error, with millions of iterations—the same way AI got good at playing games.

The arrival of ChatGPT in 2022 catalyzed the current boom. Trained on vast amounts of text, large language models work not through trial and error but by learning to predict what word should come next in a sentence. Similar models adapted to robotics were soon able to absorb pictures, sensor readings, and the position of a robot’s joints and predict the next action the machine should take, issuing dozens of motor commands every second.

This conceptual shift—to reliance on AI models that ingest large amounts of data—seems to work whether that helpful robot is supposed to talk to people, move through an environment, or even do complicated tasks. And it was paired with other ideas about how to accomplish this new way of learning, like deploying robots even if they aren’t yet perfect so they can learn from the environment they’re meant to work in. Today, Silicon Valley roboticists are dreaming big again. Here’s how that happened. 


Jibo

A movable social robot carried out conversations long before the age of LLMs.

An MIT robotics researcher named Cynthia Breazeal introduced an armless, legless, faceless robot called Jibo to the world in 2014. It looked, in fact, like a lamp. Breazeal’s aim was to create a social robot for families, and the idea pulled in $3.7 million in a crowdsourced funding campaign. Early preorders cost $749.

The early Jibo could introduce itself and dance to entertain kids, but that was about it. The vision was always for it to become a sort of embodied assistant that could handle everything from scheduling and emails to telling stories. It earned a number of devoted users, but ultimately the company shut down in 2019.

A robot with a shape vaguely like a lowercase letter "i"
A crowdfunding campaign started in 2014 and drew 4,800 Jibo preorders.
COURTESY OF MIT MEDIA LAB

In retrospect, one thing that Jibo really needed was better language capabilities. It was competing against Apple’s Siri and Amazon’s Alexa, and all those technologies at the time relied on heavy scripting. In broad terms, when you spoke to them, software would translate your speech into text, analyze what you wanted, and create a response pulled from preapproved snippets. Those snippets could be charming, but they were also repetitive and simply boringdownright robotic. That was especially a challenge for a robot that was supposed to be social and family oriented. 

What has happened since, of course, is a revolution in how machines can generate language. Voice mode from any leading AI provider is now engaging and impressive, and multiple hardware startups are trying (and failing) to build products that take advantage of it. 

But that comes with a new risk: While scripted conversations can’t really go off the rails, ones generated by AI certainly can. Some popular AI toys have, for example, talked to kids about how to find matches and knives. 


Dactyl

A robot hand trained with simulations tries to model the unpredictability and variation of the real world.

By 2018, every leading robotics lab was trying to scrap the old scripted rules and train robots through trial and error. OpenAI tried to train its robotic hand, Dactyl, virtuallywith digital models of the hand and of the palm-size cubes Dactyl was supposed to manipulate. The cubes had letters and numbers on their faces; the model might set a task like “Rotate the cube so the red side with the letter O faces upward.”

Here’s the problem: A robotic hand might get really good at doing this in its simulated world, but when you take that program and ask it to work on a real version in the real world, the slight differences between the two can cause things to go awry. Colors might be slightly different, or the deformable rubber in the robot’s fingertips could turn out to be stretchier than it was in simulation.

a Dactyl robot hand holds a Rubix cube
Dactyl, part of OpenAI’s first attempt at robotics, was trained in simulation to solve Rubik’s Cubes.
COURTESY OF OPENAI

The solution is called domain randomization. You essentially create millions of simulated worlds that all vary slightly and randomly from one another. In each one the friction might be less, or the lighting more harsh, or the colors darkened. Exposure to enough of this variation means the robots will be better able to manipulate the cube in the real world. The approach worked on Dactyl, and one year later it was able to use the same core techniques to do something harder: solving Rubik’s Cubes (though it worked only 60% of the time, and just 20% when the scrambles were particularly hard). 

Still, the limits of simulation mean that this technique plays a far smaller role today than it did in 2018. OpenAI shuttered its robotics effort in 2021 but has recently started the division up againreportedly focusing on humanoids. 


RT-2

Training on images from across the internet helps robots translate language into action.

Around 2022, Google’s robotics team was up to some strange things. It spent 17 months handing people robot controllers and filming them doing everything from picking up bags of chips to opening jars. The team ended up cataloguing 700 different tasks.

The point was to build and test one of the first large-scale foundation models for robotics. As with large language models, the idea was to input lots of text, tokenize it into a format an algorithm could work with, and then generate an output. Google’s RT-1 received input about what the robot was looking at and how the many parts of the robotic arm were positioned; then it took an instruction and translated it into motor commands to move the robot. When it had seen tasks before, it carried out 97% of them successfully; it succeeded at 76% of the instructions it hadn’t seen before. 

a robot at a table of small toys
The model RT-2, for Robotic Transformer 2, incorporated internet data to help robots process what they were seeing.
COURTESY OF GOOGLE DEEPMIND

The second iteration, RT-2, came out the following year and went even further. Instead of training on data specific to robotics, it went broad: It trained on more general images from across the internet, like the vision-language models lots of researchers were working on at the time. That allowed the robot to interpret where certain objects were in the scene.

“All these other things were unlocked,” says Kanishka Rao, a roboticist at Google DeepMind who led work on both iterations. “We could do things now like ‘Put the Coke can near the picture of Taylor Swift.’” 

In 2025, Google DeepMind further fused the worlds of large language models and robotics, releasing a Gemini Robotics model with improved ability to understand commands in natural language. 


RFM-1

An AI model that allows robotic arms to act like coworkers.

In 2017, before OpenAI shuttered its first robotics team, a group of its engineers spun out a project called Covariant, aiming to build not sci-fi humanoids but the most pragmatic of all robots: an arm that could pick up and move things in warehouses. After building a system based on foundation models similar to Google’s, Covariant deployed this platform in warehouses like those operated by Crate & Barrel and treated it as a data collection pipeline. 

By 2024, Covariant had released a robotics model, RFM-1, that you could interact with like a coworker. If you showed an arm many sleeves of tennis balls, for example, you could then instruct it to move each sleeve to a separate area. And the robot could respondperhaps predicting that it wouldn’t be able to get a good grip on the item and then asking for advice on which particular suction cups it should use. 

This sort of thing had been done in experiments, but Covariant was launching it at significant scale. The company now had cameras and data collection machines in every customer location, feeding back even more data for the model to train on.

a warehouse robot arm lifts object with many suckers to place in a bin
A Covariant robot demonstrates “induction”—the common warehouse task of placing objects on sorters or conveyors.
COURTESY OF COVARIANT

It wasn’t perfect. In a demo in March 2024 with an array of kitchen items, the robot struggled when it was asked to “return the banana” to its original location. It picked up a sponge, then an apple, then a host of other items before it finally accomplished the task. 

It “doesn’t understand the new concept” of retracing its steps, cofounder Peter Chen told me at the time. “But it’s a good exampleit might not work well yet in the places where you don’t have good training data.”

Chen and fellow founder Pieter Abbeel were soon hired by Amazon, which is currently licensing Covariant’s robotics model (Amazon did not respond to questions about how it’s being used, but the company runs an estimated 1,300 warehouses in the US alone). 


Digit

Companies are putting this humanoid to the test in real-world settings.

The new investment dollars flowing to robotics startups are aimed largely at robots shaped not like lamps or arms but like people. Humanoid robots are supposed to be able to seamlessly enter the spaces and jobs where humans currently work, avoiding the need to retool assembly lines to accommodate new shapes such as giant arms. 

It’s easier said than done. In the rare cases where humanoids appear in real warehouses, they’re often confined to test zones and pilot programs. 

Digit humanoid robot putting a plastic bin on a conveyor belt
Amazon and other companies are using Digit to help move shipping totes.
COURTESY OF AGILITY ROBOTICS

That said, Agility’s humanoid Digit appears to be doing some real work. The designwith exposed joints and a distinctly unhuman headis driven more by function than by sci-fi aesthetics. Amazon, Toyota, and GXO (a logistics giant with customers like Apple and Nike) have all deployed itmaking it one of the first examples of a humanoid robot that companies see as providing actual cost savings rather than novelty. Their Digits spend their days picking up, moving, and stacking shipping totes.

The current Digit is still a long way from the humanlike helper Silicon Valley is betting on, though. It can lift only 35 pounds, for exampleand every time Agility makes Digit stronger, its battery gets heavier and it has to recharge more often. And standards organizations say humanoids need stricter safety rules than most industrial robots, because they’re designed to be mobile and spend time in proximity to people. 

But Digit shows that this revolution in robot training isn’t converging on a single method. Agility relies on simulation techniques like those OpenAI used to train its hand, and the company has worked with Google’s Gemini models to help its robots adapt to new environments. That’s where more than a decade of experiments have gotten the industry: Now it’s building big.

Treating enterprise AI as an operating layer

There’s a fault line running through enterprise AI, and it’s not the one getting the most attention. The public conversation still tracks foundation models and benchmarks—GPT versus Gemini, reasoning scores, and marginal capability gains. But in practice, the more durable advantage is structural: who owns the operating layer where intelligence is applied, governed, and improved. One model treats AI as an on-demand utility; the other embeds it as an operating layer—the combination of operation software, data capture, feedback loops and governance that sits between models and real work—that compounds with use.

Model providers like OpenAI and Anthropic sell intelligence as a service: you have a problem, you call an API, you get an answer. That intelligence is general-purpose, largely stateless, and only loosely connected to the day-to-day operations where decisions are made. It’s highly capable and increasingly interchangeable. The distinction that matters is whether intelligence resets on every prompt or accumulates over time.

Incumbent organizations, by contrast, can treat AI as an operating layer: instrumentation across operations, feedback loops from human decisions, and governance that turns individual tasks into reusable policy. In that setup, every exception, correction, and approval becomes a chance to learn—and intelligence can improve as the platform absorbs more of the organization’s work. The organizations most likely to shape the enterprise AI era are those that can embed intelligence directly into operational platforms and instrument those platforms so work generates usable signals.

The prevailing narrative says nimble startups will out-innovate incumbents by building AI-native from scratch. If AI is primarily a model problem, that story holds. But in many enterprise domains, AI is a systems problem—integrations, permissions, evaluation, and change management—where advantage accrues to whomever already sits inside high-volume, high-stakes operations and converts that position into learning and automation.

The inversion: AI executes, humans adjudicate

Traditional services organizations are built on a simple architecture: humans use software to do expert work. Operators log into systems, navigate operations, make decisions, and process cases. Technology is the medium. Human judgment is the product.

An AI-native platform inverts this. It ingests a problem, applies accumulated domain knowledge, executes autonomously what it can with high confidence, and routes targeted sub-tasks to human experts when the situation demands judgment that the system can’t yet reliably provide.

But inverting human-AI interaction isn’t just a UI redesign—it requires raw material. It’s only possible when the platform is built on a foundation of domain expertise, behavioral data, and operational knowledge accumulated over years.

The three compounding assets incumbents already own

AI-native startups begin with a clean architectural slate and can move quickly. What they can’t easily manufacture is the raw material that makes domain AI defensible at scale:

  • Proprietary operational data
  • A large workforce of domain experts whose day-to-day decisions generate training signals
  • Accumulated tacit knowledge about how complex work actually gets done

Services companies already have all three. But these ingredients aren’t moats on their own. They become an advantage only when a company can systematically convert messy operations into AI-ready signals and institutional knowledge—then feed the results back into operations so the system keeps improving.

Codifying expertise into reusable signals

In most services organizations, expertise is tacit and perishable. The best operators know things they cannot easily articulate: heuristics developed over the years, edge-case intuitions, and pattern recognition that operate below the level of conscious reasoning.

At Ensemble, the strategy for addressing this challenge is knowledge distillation. The systematic conversion of expert judgment and operational decisions into machine-readable training signals.

In health-care revenue cycle management, for example, systems can be seeded with explicit domain knowledge and then deepen their coverage through structured daily interaction with operators. In Ensemble’s implementation, the system identifies gaps, formulates targeted questions, and cross-checks answers across multiple experts to capture both consensus and edge-case nuance. It then synthesizes these inputs into a living knowledge base that reflects the situational reasoning behind expert-level performance.

Turning decisions into a learning flywheel

Once a system is constrained enough to be trusted, the next question is how it gets better without waiting for annual model upgrades. Every time a skilled operator makes a decision, they generate more than a completed task. They generate a potential labeled example—context paired with an expert action (and sometimes an outcome). At scale, across thousands of operators and millions of decisions, that stream can power supervised learning, evaluation, and targeted forms of reinforcement—teaching systems to behave more like experts in real conditions.

For example, if an organization processes 50,000 cases a week and captures just three high-quality decision points per case, that’s 150,000 labeled examples every week without creating a separate data-collection program.

A more advanced human-in-the-loop design places experts inside the decision process, so systems learn not just what the right answer was, but how ambiguity gets resolved. Practically, humans intervene at branch points—selecting from AI-generated options, correcting assumptions, and redirecting operations. Each intervention becomes a high-value training signal. When the platform detects an edge case or a deviation from the expected process, it can prompt for a brief, structured rationale, capturing decision factors without requiring lengthy free-form reasoning logs.

Building toward expertise amplification

The goal is to permanently embed the accumulated expertise of thousands of domain experts—their knowledge, decisions, and reasoning—into an AI platform that amplifies what every operator can accomplish. Done well, this produces a quality of execution that neither humans nor AI achieve independently: higher consistency, improved throughput, and measurable operational gains. Operators can focus on more consequential work, supported by an AI that has already completed the analytical groundwork across thousands of analogous prior cases.

The broader implication for enterprise leaders is straightforward. Advantages in AI won’t be determined by access to general-purpose models alone. It will come from an organization’s ability to capture, refine, and compound what it knows, its data, decisions, and operational judgment, while building the controls required for high-stakes environments. As AI shifts from experimentation to infrastructure, the most durable edge may belong to the companies that understand the work well enough to instrument it and can turn that understanding into systems that improve with use.

This content was produced by Ensemble. It was not written by MIT Technology Review’s editorial staff.

Establishing standards of care for forensic mental health: an international Delphi consensus-building study

ObjectivesThe present study aimed to establish a consensus on a definition of forensic mental health systems and services, and to identify principles and components of forensic mental health systems.MethodsA Delphi consensus-building process was employed among 23 experts in forensic mental health, defined by lived experience of forensic mental health services, professional, clinical or management practice in forensic settings, or academic research in the field. Items were rated on a 9-point Likert scale, with consensus defined as ≥75% of panelists rating an item between 7 and 9. Across three Delphi rounds, items were revised, merged, or added based on participant feedback. Data were collected anonymously using LimeSurvey, with reminders sent to maximize participation, followed by a structured consensus meeting to resolve remaining areas of disagreement.ResultsThe final consensus statement comprises three components: (1) a definition of forensic mental health services; (2) a general statement including 12 guiding principles; and (3) 43 core components organized across 10 thematic domains addressing models of care, pathways and processes, programs and activities, physical health, service user and peer involvement, evaluation and improvement, service integration, safe environments, restrictive practices, and other system-level considerations. While all items achieved consensus at the consensus meeting, areas of sustained discussion related to the integration of cultural expertise, the inclusion of a lived experience workforce, and the distinction between descriptive and aspirational elements of forensic mental health services.ConclusionsThis international consensus statement provides a structured framework for understanding forensic mental health systems. By articulating shared principles and core components while allowing flexibility across jurisdictions, the framework offers a foundation to support service development and evaluation across diverse jurisdictions.

Cyberscammers are bypassing banks’ security with illicit tools sold on Telegram

From inside a money-laundering center in Cambodia, an employee opens a popular Vietnamese banking app on his phone. The app asks him to upload a photo associated with the account, so he clicks on a picture of a 30-something Asian man.

Next, the app requests to open the camera for a video “liveness” check. The scammer holds up a static image of a woman bearing no resemblance to the man who owns the account. After a 90-second wait—as the app tells him to readjust the face inside the frame—he’s in. 

The exploit he’s demonstrating, in a video shared with me by a cyberscam researcher named Hieu Minh Ngo, is possible thanks to one of a growing range of illicit hacking services, readily available for purchase on Telegram, that are designed to break “Know Your Customer” (KYC) facial scans.

These banking and crypto safeguards are supposed to confirm that an account belongs to a real person, and that the user’s face matches the identity documents that were provided to open the account. But scammers are bypassing them in order to open mule accounts and launder money. Rather than using a live phone camera feed for a liveness check, the hacks typically deploy a tool known as a virtual camera. Users can replace the video stream with other videos or photos—depicting a real or deepfake person or even an object.

As financial institutions enact enhanced security measures aimed at stopping cyberscammers, these workarounds are the latest round in the cat-and-mouse game between criminal operators and the financial services industry.

Over the course of a two-month investigation earlier this year, MIT Technology Review identified 22 Chinese-, Vietnamese-, and English-language public Telegram channels and groups advertising bypass kits and stolen biometric data. The software kits use a variety of methods to compromise phone operating systems and banking applications, claiming to enable users to get around the compliance checks imposed by financial institutions ranging from major crypto exchanges such as Binance to name-brand banks like Spain’s BBVA. 

“Specializing in bank services—handling dirty money,” reads the since-deleted Telegram bio of the program used by the Cambodian launderer, complete with a thumbs-up emoji. “Secure. Professional. High quality.” Some of the channels and groups had thousands of subscribers or members, and many posted bullet points listing their services (“All kinds of KYC verification services”; “It’s all smooth and seamless”) alongside videos purporting to show successful hacks. 

Telegram says that after reviewing the accounts, it removed them for violating its terms of service. But such online marketplaces proliferate easily, and multiple channels and groups advertising similar tools remain active.

Banks and butchers

The rise in KYC bypasses has occurred alongside an expansion of a global industry in “pig-butchering” cyberscams. Crypto platforms and banks around the world are facing increasing scrutiny over the flow of illegally obtained money, including profits from such scams, through their platforms. This has prompted tightened banking regulations in countries such as Vietnam and Thailand, where governments have increased customer verification and fraud monitoring requirements and are pushing for stronger anti-money-laundering safeguards in the crypto industry.

Chainalysis, a US blockchain analysis firm, estimates that around $17 billion was stolen in 2025 in crypto scams and fraud, up from $13 billion in 2024. The United Nations Office on Drugs and Crime, meanwhile, warned in a recent report that the expansion of Asian scam syndicates in Africa and the Pacific has helped the industry “dramatically scale up profits.”

That combination of factors—more scrutiny, but also more revenue—has vaulted KYC bypasses to the center of the online marketplace for cyberscam and casino money launderers. Although estimates vary, cybersecurity researchers say these kinds of attacks are rising: The biometrics verification company iProov estimated that virtual-camera attacks were more than 25 times as common worldwide 2024 than in 2023, while Sumsub, a company providing KYC services, reported that “sophisticated” or multi-step fraud attempts, including virtual-camera bypasses, almost tripled last year among its clients. 

Three financial institutions that were named as targets on such Telegram channels—the world’s largest crypto exchange, Binance, as well as BBVA and UK-based Revolut—told me they’re aware of such bypasses and emphasize that they’re an industry-wide challenge. A spokesperson from Binance said it has “observed attempts of this nature to circumvent our controls,” adding that “we have successfully prevented such attacks and remain confident in our systems.”  BBVA and Revolut also declined to comment on whether their safeguards had been breached.

It’s difficult to estimate success rates, because companies may not be aware of bypasses—or report them—until later. “What’s important is what we don’t see,” Artem Popov, Sumsub’s head of fraud prevention products, told me, referring to attacks that go undetected. “There’s always part of the story where it might be completely hidden from our eyes, and from the eyes of any company in the industry, using any type of KYC provider.”

How criminals navigate a compliance maze 

Advertisements for the exploits appear simple enough, but on the back end, building a successful bypass is complex and often involves multiple methods. Some channels offer to jailbreak a physical phone so that scammers can trigger the use of a virtual camera (VCam) instead of the built-in one whenever they’d like. Other hacks inject code known as a “hooking framework” into a financial institution’s app that triggers the VCam to open. Either way, VCams can be used to dupe KYC safeguards with images or videos that replace genuine, live video of the account’s owner.

Sergiy Yakymchuk, CEO of Talsec, a cybersecurity company that primarily serves financial institutions, reviewed details from the Telegram channels identified by MIT Technology Review and says they are consistent with successful tactics used against his banking and crypto clients. His team received help requests from banks and exchanges for roughly 30 VCam-based hacks over the past year, up from fewer than 10 in 2023. 

Increasingly, hackers compromise both the phone itself and the code of the financial institutions’ apps before feeding the virtual camera a mix of stolen biometrics and deepfakes, Yakymchuk says.

“Some time ago, it was enough to decompile the app of a bank and distribute this on Telegram, and that was everything you needed,” he says. “Now it’s not enough, because you have KYC—and more and more things are needed.”

For money launderers, KYC bypasses have “become essential for everything right now—because scam compounds need to move money,” says Ngo, the researcher who shared the demo video. A convicted former hacker who became a cybersecurity advisor for the Vietnamese government, Ngo now runs an anti-scam nonprofit and helps law enforcement investigate money laundering. 

He describes how the process works in the case of pig-butchering scams: Funds originating with victims are received into bank accounts controlled or rented by a money-laundering network, known colloquially as “water houses.” Money launderers use KYC bypasses to access the accounts and quickly redistribute the profits before converting them into digital assets—typically in the form of the stablecoin Tether, a type of cryptocurrency that is pegged to the US dollar.

These transactions often happen in seconds, under tightly orchestrated management. “They know, very clearly, the flow of how the banks verify or authenticate accounts,” Ngo says. 

A cat-and-mouse game 

The growth of cyberscam money laundering has led to heightened scrutiny of financial institutions. In 2023, Binance pleaded guilty in US federal courts to operating without anti-money-laundering safeguards. Donald Trump pardoned former Binance CEO Chaopeng Zhao last October.

Recent analysis from the International Consortium of Investigative Journalists found that after Zhao’s guilty plea, more than $400 million continued to move to Binance from Huione Group, a Cambodia-based firm that the US sanctioned after the Treasury Department deemed it a “critical node” for money laundering in pig-butchering scams.

Binance says it has “state-of-the-art security systems” that prevented billions in fraud losses and that the company processed more than 71,000 law enforcement requests in 2025.

But John Griffin, a finance and blockchain expert at the University of Texas at Austin, does not think the exchanges are sufficiently secure. “Even though they have all this press about ‘Oh, yes, we’ve changed this and that’—well, the proof is in the pudding. The criminals are still using your exchange,” Griffin told me of the industry at large. “So there must be holes.” (Binance says it “objects to the dubious findings” of Griffin’s work tracking the flow of criminal profits across exchanges like Binance, Huobi, OKX, and Tokenlon, calling it “misleading at best and, at worst, wildly inaccurate.”)

Binance also pointed out that some purported bypass services are themselves scams, casting doubt on whether successful bypasses are as widespread as the Telegram marketplace may suggest. Engaging with such services “exposes individuals to significant security risks,” a spokesperson said. “Even where access appears to be granted, accounts are often already restricted by internal detection and compliance controls, rendering them nonfunctional for trading or withdrawals.”

Regulators around the world are trying to catch up. In Thailand, where citizens’ bank accounts regularly serve as money mules for cyberscams based in neighboring Myanmar and Cambodia, new legislation has enhanced KYC monitoring, limited daily transactions, and strengthened oversight bodies’ ability to suspend accounts. The US money-laundering regulator, the Financial Crimes Enforcement Network, issued a warning against KYC deepfakes and the use of VCams in late 2024, encouraging platforms to track broader transaction patterns to identify money laundering.

For scammers, any new security or reporting requirements will make bypasses harder, but “it’s not going to stop them,” Ngo says. “It’s just a matter of time.”

No one’s sure if synthetic mirror life will kill us all

For four days in February 2019, some 30 synthetic biologists and ethicists hunkered down at a conference center in Northern Virginia to brainstorm high-risk, cutting-­edge, irresistibly exciting ideas that the National Science Foundation should fund. By the end of the meeting, they’d landed on a compelling contender: making “mirror” bacteria. Should they come to be, the lab-created microbes would be structured and organized like ordinary bacteria, with one important exception: Key biological molecules like proteins, sugars, and lipids would be the mirror images of those found in nature. DNA, RNA, and many other components of living cells are chiral, which means they have a built-in rotational structure. Their mirrors would twist in the opposite direction. 

Researchers thrilled at the prospect. “Everybody—everybody—thought this was cool,” says John Glass, a synthetic biologist at the J. Craig Venter Institute in La Jolla, California, who attended the 2019 workshop and is a pioneer in developing synthetic cells. It was “an incredibly difficult project that would tell us potentially new things about how to design and build cells, or about the origin of life on Earth.” The group saw enormous potential for medicine, too. Mirror microbes might be engineered as biological factories, producing mirror molecules that could form the basis for new kinds of drugs. In theory, such therapeutics could perform the same functions as their natural counterparts, but without triggering unwelcome immune responses. 

After the meeting, the biologists recommended NSF funding for a handful of research groups to develop tools and carry out preliminary experiments, the beginnings of a path through the looking glass. The excitement was global. The National Natural Science Foundation of China funded major projects in mirror biology, as did the German Federal Ministry of Research, Technology, and Space.

By five years later, in 2024, many researchers involved in that NSF meeting had reversed course. They’d become convinced that in the worst of all possible futures, mirror organisms could trigger a catastrophic event threatening every form of life on Earth; they’d proliferate without predators and evade the immune defenses of people, plants, and animals. 

“I wish that one sunny afternoon we were having coffee and we realized the world’s about to end, but that’s not what happened.”

Kate Adamala, synthetic biologist, University of Minnesota

Over the past two years, they’ve been ringing alarm bells. They published an article in Science in December 2024, accompanied by a 299-page technical report addressing feasibility and risks. They’ve written essays and convened panels and cofounded the Mirror Biology Dialogues Fund (MBDF), a broadly funded nonprofit charged with supporting work on understanding and addressing the risk. The issue has received a blaze of media attention and ignited dialogues among not only chemists and synthetic biologists but also bioethicists and policymakers.  

What’s received less attention, however, is how we got here and what uncertainties still remain about any potential threat. Creating a mirror-life organism would be tremendously complicated and expensive. And although the scientific community is taking the alarm seriously, some scientists doubt whether it’s even possible to create a mirror organism anytime soon. “The hypothetical creation of mirror-­image organisms lies far beyond the reach of present-day science,” says Ting Zhu, a molecular biologist at Westlake University, in China, whose lab focuses on synthesizing mirror-image peptides and other molecules. He and others have urged colleagues not to let speculation and anxiety guide decision-making and argued that it’s premature to call for a broad moratorium on early-stage research, which they say could have medical benefits. 

But the researchers who are raising flags describe a pathway, even multiple pathways, to bringing mirror life into existence—and they say we urgently need guardrails to figure out what kinds of mirror-biology research might still be safe. That means they’re facing a question that others have encountered before, multiple times over the last several decades and with mixed results—one that doesn’t have a neat home in the scientific method. What should scientists do when they see the shadow of the end of the world in their own research? 

Looking-glass life

The French chemist and microbiologist Louis Pasteur was the first to recognize that biological molecules had built-in handedness. In the late 19th century, he described all living species as “functions of cosmic asymmetry.” What would happen, he mused, if one could replace these chiral components with their mirror opposites? 

Scientists now recognize that chirality is central to life itself, though no one knows why. In humans, 19 of the 20 so-called “standard” amino acids that make up proteins are chiral, and all in the same way. (The outlier, glycine, is symmetrical.) The functions of proteins are intricately tied to their shapes, and they mostly interact with other molecules through chiral structures. Almost all receptors on the surface of a cell are chiral. During an infection, the immune system’s sentinels use chirality to detect and bind to antigens—substances that trigger an immune response—and to start the process of building antibodies. 

By the late 20th century, researchers had begun to explore the idea of reversing chirality. In 1992, one team reported having synthesized the first mirror-image protein. That, in turn, set off the first clarion call about the risk: In response to the discovery, chemists at Purdue University pointed out, briefly, that mirror-life organisms, if they escaped from a lab, would be immune to any attack by “normal” life. A 2010 story in Wired highlighting early findings in the area noted that if a such a microbe developed the ability to photosynthesize, it could obliterate life as we know it. 

The synthetic biology community didn’t seriously weigh those threats then, says David Relman, a specialist who bridges infectious disease and microbiology at Stanford University and a trailblazer in studying the gut and oral microbiomes. The idea of a mirror microbe seemed too far beyond the actual progress on proteins. “This was almost a solely theoretical argument 20 years ago,” he says. 

Now the research landscape has changed. 

Scientists are quickly making progress on mirror images of the machinery cells use to make proteins and to self-replicate. Those components include DNA, which encodes the recipes for proteins; DNA polymerases, which help copy genetic material; and RNA, which carries recipes to ribosomes, the cell’s protein factories. If researchers could make self-replicating mirror ribosomes, then they would have an efficient way to produce mirror proteins. That could be used as a biological manufacturing method for therapeutics. But embedded in a self-­replicating, metabolizing synthetic cell, all these pieces could give rise to a mirror microbe. 

When synthetic biologists convened in Northern Virginia in 2019, they didn’t recognize how quickly the technology was advancing, and if they saw a threat at all, it may have been obscured by the blinding appeal of pushing the science forward. What’s become apparent now, says Glass, is that scientists in different disciplines, all related to mirror life, were largely unaware of what other scientists had been doing. Chemists didn’t know that synthetic biologists had made so much progress on creating mirror cells with natural chirality from scratch. Biologists didn’t appreciate that chemists were building ever-larger mirror macromolecules. “We tend to be siloed,” Glass says. And nobody, he says, had thought to seriously examine the immune system concerns that had already been raised in response to earlier work. “There was not an immunologist or an infectious disease person in the room,” Glass says, reflecting on the 2019 meeting. “I may have come closest, given that I work with pathogenic bacteria and viruses,” he adds, but his work doesn’t address how they cause infections in their hosts.

on the left, a hand with petri dish and the same image inverted on the right

GETTY IMAGES

These scientists also didn’t know that around the same time as their meeting, another conversation about mirror life was happening—a darker dialogue that was as focused on danger as it was on discovery. Starting around 2016, researchers with a nonprofit called Open Philanthropy had begun compiling research files on catastrophic biological risks. The organization, which rebranded as Coefficient Giving in 2025, funds projects across a range of focus areas; it adheres to a divisive philanthropic philosophy called effective altruism, which advocates giving money to projects with the highest potential benefit to the most people. While that might not sound objectionable, critics point out that the metrics devotees use to gauge “effectiveness” can prioritize long-term solutions while neglecting social injustices or systemic problems. 

Someone in Open Philanthropy’s bio­security group had suggested looking into the risks posed by mirror life. In 2019 the organization began funding research by Kevin Esvelt, who leads the Sculpting Evolution group at the MIT Media Lab, on biosecurity issues, including mirror life. He began reading up to see whether mirror life was something to worry about.

Esvelt made waves in 2013 for pioneering the use of CRISPR to develop a gene drive, a technology that could spread genetic changes introduced into a living organism through a whole population. Researchers are exploring its use, for example, to make mosquitoes hostile to the parasite that causes malaria—and, as a result, lower their chance of spreading it to humans. But almost immediately after he developed the tool, Esvelt argued against using it for profit, at least until proper safeguards could be set and its use in fighting malaria had been established. “Do you really have the right to run an experiment where if you screw up, it affects the whole world?” he asked, in this magazine, in 2016. At the Media Lab, Esvelt leads efforts to safely develop gene drives that can be deployed locally but prevented from spreading globally. 

Esvelt says he’s often thinking about the security risks posed by self-sustaining genetically engineered technologies, and research led him to suspect that the threat of mirror organisms hadn’t been seriously interrogated. The more he learned about microbial growth rates, predator-prey and microbe-microbe interactions, and immunology, the more he began to worry that mirror organisms, if impervious to the innate defenses of natural ones, could cause unstoppable infections in the event that they escaped the lab. 

Even if the first experimental iteration of such a germ were too fragile to survive in the environment or a human body, Esvelt says, it would be a light lift to genetically engineer new, more resilient versions with existing technology. Even worse, he says, the results could be weaponized. The possible path from 2019 to global annihilation seemed almost too direct, he found. 

But he wasn’t an expert in all the scientific fields involved in research on mirror life, so he started making calls. He first described his concerns to Relman one night in February 2022, at a restaurant outside Washington, DC. Esvelt hoped Relman would tell him he was wrong, that he’d missed something over the years of gathering data. Instead, he was troubled. 

The concern spreads

When Relman returned to California, he read more about the technology, the risks, and the role of chirality in the immune system and the environment. And he consulted experts he knew well—ecologists, other microbiologists, immunologists, all of them leaders in their fields—in an attempt to assuage his concerns. “I was hoping that they’d be able to say, I’ve thought about this, and I see a problem with your logic. I see that it’s really not so bad,” he says. “At every turn, that did not happen. Something about it was new to every person.” 

The concern spread. Relman worked with Jack Szostak, a professor of chemistry at the University of Chicago, and a group of researchers to see if it was possible to make an argument that mirror life wasn’t going to wipe out humanity. Included in that group was Kate Adamala, a synthetic biologist at the University of Minnesota. She was a natural choice: Adamala had shared the initial grant from the NSF, in 2019, to explore mirror-life technologies. 

She also became convinced the risk was real—and was dumbfounded that she hadn’t seen it earlier. “I wish that one sunny afternoon we were having coffee and we realized the world’s about to end, but that’s not what happened,” she says. “I’m embarrassed to admit that I wasn’t even the one that brought up the risks first.” Through late 2023 and early 2024, the endeavor began to take on the form of a rigorous scientific investigation. Experts were presented with a hypothesis—namely, that if mirror cells were built, they would pose an existential threat—and asked to challenge it. The goal was to falsify the hypothesis. “It would be great if we were wrong,” says Vaughn Cooper, a microbiologist at the University of Pittsburgh and president-elect of the American Society for Microbiology. 

Relman says that as the chemists and biologists learned more about one another’s work and began to understand what immunologists know about how living things defend themselves, they started to connect the dots and see an emerging picture of an unstoppable synthetic threat.

Some scientists have pushed back against the doomsday scenario, suggesting that the case against mirror life offers an “inflated view of the danger.”

Timothy Hand, an immunologist at the University of Pittsburgh who hadn’t participated in the 2019 NSF meeting, wasn’t initially worried when he heard about mirror life, in 2024. “The mammalian immune system has this incredible capability to make antibodies against any shape,” he says. “Who cares if it’s a mirror?” But when he took a closer look at that process, he could see a cascade of potential problems far upstream of antibody production. Start with detection: Macrophages, which are cells the immune system uses to identify and dispatch invaders, use chiral sensing receptors on their surfaces. The proteins they use to grab on to those invaders, too, are chiral. That suggests the possibility that an organism could be infected with a mirror organism but not be able to detect it or defend against it. “The lack of innate immune sensing is an incredibly dangerous circumstance for the host,” Hand says.

By early 2024, Glass had become concerned as well. Relman and James Wagstaff, a structural biologist from Open Philanthropy, visited him at the Venter Institute to talk about the possibility of using synthetic cell technology—Glass’s specialty—to build mirror life. “At first I thought, This can’t be real,” Glass says. They walked through arguments and counterarguments. “The more this went on, the more I started feeling ill,” he says. “It made me realize that work I had been doing for much of the last 20 years could be setting the world up for this incredible catastrophe.” 

In the second half of 2024, the growing group of scientists assembled the report and wrote the policy forum for Science. Relman briefed policymakers at the White House, members of the defense community, and the National Security Agency. Researchers met with the National Institutes of Health and the National Science Foundation. “We briefed the United Nations, the UK government, the government of Singapore, scientific funding organizations from Brazil,” says Glass. “We’ve talked to the Chinese government indirectly. We were trying to not blindside anybody.” 

A year and a half on, the push has had an impact. UNESCO has recommended a precautionary global moratorium on creating mirror-life cells, and major philanthropic organizations that fund science, including the Alfred P. Sloan Foundation, have announced they will not finance research leading to a mirror microorganism. The Bulletin of the Atomic Scientists highlighted considerations about mirror life in its most recent report on the Doomsday Clock. In March, the United Nations Secretary-General’s Scientific Advisory Board issued a brief highlighting the risks—noting, for example, that recent progress on building mirror molecules could reduce the cost of creating a mirror microbe. 

“I think no one really believes at this stage that we should make mirror life, based on the evidence that’s available,” says James Smith, the scientist who leads the MBDF, the nonprofit focused on assessing the risks of mirror life, which is funded by Coefficient Giving, the Sloan Foundation, and other organizations. The challenge now, Smith says, is for scientists to work with policymakers and bioethicists to figure out how much research on mirror life should be permitted—and who will enforce the rules.

Drawing the line

Not everyone is convinced that mirror organisms pose an existential threat. It’s difficult to verify predictions about how mirror microbes would fare in the immune system—or the larger world—without running experiments on them. Some scientists have pushed back against the doomsday scenario, suggesting that the case against mirror life offers an “inflated view of the danger.” Others have noted that carbohydrates called glycans already exist in both left- and right-handed forms—even in pathogens—and the immune system can recognize both of them. Experiments focused on interactions between the immune system and mirror molecules, they say, could help clarify the risks of mirror organisms and reduce uncertainty. 

Even among those convinced that the worst-case scenario is possible, researchers still disagree over where to draw the line. What inquiries should be allowed and what should be prohibited?

Andy Ellington, a biotechnologist and synthetic biologist at the University of Texas at Austin, doesn’t think mirror organisms will come to fruition anytime soon. Even if they do, he isn’t sure they will pose a threat. “If there is going to be harm done to the human race, this is about position 382 on my list,” he says. But at the same time, he says it’s a complicated issue worth studying more, and he wants to see the conversations continue: “We’re operating in a space where there’s so much unknown that it’s very difficult for us to do risk assessment.” 

Even among those convinced that the worst-case scenario is possible, researchers still disagree over where to draw the line. What inquiries should be allowed and what should be prohibited? 

Adamala, of the University of Minnesota, and others see a natural line at ribosomes, the cellular factories that transform chains of amino acids into proteins. These would be a critical ingredient in creating a self-replicating organism, and Adamala says the path to getting there once mirror ribosomes are in place would be pretty straightforward. But Zhu, at Westlake, and others counter that it’s worth developing mirror ribosomes because they could possibly produce medically useful peptides and proteins more efficiently than traditional chemical methods. He sees a clear distinction, and a foundational gap, between that kind of technology and the creation of a living synthetic organism. “It is crucial to distinguish mirror-image molecular biology from mirror-image life,” he says. That said, he points out that many synthetic molecules and organisms containing unnatural components, including but not limited to the mirror-image subset, might pose health risks. Researchers, he says, should focus on developing holistic guidelines to cover such risks—not just those from mirror molecules. 

Even if the exact risk remains uncertain, Esvelt remains more convinced than ever that the work should be paused, perhaps indefinitely. No one has taken a meaningful swing at the hypothesis that mirror life could wipe out everything, he says. The primary uncertainties aren’t around whether mirror life is dangerous, he points out; they have more to do with identifying which bacterium—including what genes it encodes, what it eats, how it evades the immune system’s sentinels—could lead to the most serious consequences. “The risk of losing everything, like the entire future of humanity integrated over time, is not worth any small fraction of the economy. You just don’t muck around with existential risk like that,” he says. 

In some ways, scientists have been here before, working out rules and limits for research. Two years after the start of the covid-19 pandemic, for example, the World Health Organization published guidelines for managing risks in biological research. But the history is much deeper: Horrific episodes of human experimentation led to the establishment of institutional review boards to provide ethical oversight. In the early 1970s, in response to concerns over lab-acquired infections and growing use of biological warfare, the US Centers for Disease Control and Prevention established biohazard safety levels (BSLs), which govern work on potentially dangerous biological experiments.

And in 1975—at the dawn of recombinant DNA research, which allows researchers to put genetic material from one organism into another—geneticists met at the Asilomar conference center in Pacific Grove, California, to hammer out rules governing the work. There were concerns over what would happen if some virus or bacterium, genetically engineered to have traits that would make it particularly dangerous for people, escaped from a lab. Scientists agreed to self-imposed restrictions, like a moratorium on research until new safety guidelines were in place. As a result of the meeting, in June 1976 the NIH issued rules that, among other things, categorized the risks associated with rDNA experiments and aligned them with the newly adopted BSL system.

Asilomar is often hailed as a successful model for scientific self-governance. But that perception reflects a tendency to recall the meeting through a nostalgic haze. “In fact, it was incredibly messy and human,” says Luis Campos, a historian of science at Rice University. Equally brilliant Nobelists argued on either side of the question of whether to rein in rDNA research. Technical discussions dominated; talks about who would be affected by the technology were missing. The meeting didn’t start establishing guidelines, says Campos, until the lawyers mentioned liability and lab leaks. 

For now it’s unclear whether these examples of self-­governance, which arose from the demonstrated risks of existing technologies, hold useful lessons for the mirror-life community. Three competing images of the future are coming into focus: Mirror life might not be possible, it might be possible but not threatening, or it might be possible and capable of obliterating all life on Earth. 

Scientists may be censoring themselves out of fear and speculation. To some, shutting down the work seems necessary and urgent; to others, it is unnecessarily limiting. What’s clear is that the question of what to do about mirror life has been both illuminating and disorienting, pushing scientists to interrogate not only their current research but where it might lead. This is uncharted territory. 

Stephen Ornes is a science writer based in Nashville, Tennessee.

Building trust in the AI era with privacy-led UX

The practice of privacy-led user experience (UX) is a design philosophy that treats transparency around data collection and usage as an integral part of the customer relationship. An undertapped opportunity in digital marketing, privacy-led UX treats user consent not as a tick-box compliance exercise, but rather as the first overture in an ongoing customer relationship. For the companies that get it right, the payoff can bring something more intangible, valuable, and durable than simple consent rates: consumer trust.

The opportunities of privacy-led UX have only recently come into focus. Adelina Peltea, the chief marketing officer at Usercentrics, has seen enterprise sentiment shift: “Even just a few years ago, this space was viewed more as a trade-off between growth and compliance,” she says. “But as the market has matured, there’s been a greater focus on how to tie well-designed privacy experiences to business growth.”

And it turns out that well-designed, value-forward consent experiences routinely outperform initial estimates.
Touchpoints for privacy-led UX often include consent management platforms, terms and conditions, privacy policies, data subject access request (DSAR) tools, and, increasingly, AI data use disclosures.

This report examines how data transparency builds trust with customers; how this, in turn, can support business performance; and how organizations can maintain this trust even as AI systems add complexity to consent processes.

Key findings include the following:

  • Privacy is evolving from a one-time consent transaction into an ongoing data relationship. Rather than asking users for broad permissions up front, leading organizations are introducing data-sharing decisions gradually, matching the depth of the ask to the stage of the customer relationship. Companies that take this tack tend to gather both a larger quantity and higher quality of consumer data, the value of which often compounds over time.
  • Privacy-led UX is a prerequisite for AI growth. The consumer data that organizations gather is rapidly becoming a core foundation upon which AI-powered personalization is built. Organizations that establish clear, enforceable privacy and data transparency policies now are better positioned to deploy AI responsibly and at scale in the future. This starts with correctly configured consent mode across ad platforms.
  • Agentic AI introduces new levels of both complexity and opportunity. As AI systems begin acting on users’ behalf, the traditional consent moment may never occur. Governing agent-generated data flows requires privacy infrastructure that goes well beyond the cookie banner.
  • Realizing the advantages of privacy-led UX requires cross-functional collaboration and clear leadership. Privacy-led UX touches marketing, product, legal, and data teams—but someone must own the strategy and weave the threads together. Chief marketing officers
  • (CMOs) are often best positioned for that role, given their visibility across brand, data, and customer experience.
  • A practical framework can support businesses in getting it right. Organizations must define their data collection and usage strategies and ensure their UX incorporates data consent, including a focus on banner design. Following a blueprint for evaluating and improving privacy-led UX supports consistency at every consent touchpoint.

Download the report.

This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff. It was researched, designed, and written by human writers, editors, analysts, and illustrators. This includes the writing of surveys and collection of data for surveys. AI tools that may have been used were limited to secondary production processes that passed thorough human review.

Designing Psychologically Grounded Artificial Intelligence for Supporting Bystander-Based Cyberaggression Intervention: Mixed Methods Exploratory Study

Background: Cyberaggression poses a growing threat to mental health, contributing to increased distress, reduced self-esteem, and other adverse psychosocial outcomes. Although bystander intervention can mitigate the escalation and impact of cyberaggression, individuals often lack the confidence, strategies, or language to respond effectively in these high-stakes online interactions. Advances in generative artificial intelligence (AI) present a novel opportunity to facilitate digital behavior change by assisting bystanders with contextually appropriate, theory-informed intervention messages that promote safer online environments and support mental well-being. Objective: This mixed methods design study aimed to explore the feasibility of using generative AI to support bystander intervention in cyberaggression on social media. Specifically, we examined whether AI can generate effective responses aligned with established intervention strategies and how these responses are perceived in terms of their potential to de-escalate online harm and foster behavior change. Methods: We collected 1000 real-world cyberaggression examples from public social media datasets and generated bystander intervention responses using 3 distinct prompt strategies: a generic policy reminder, a baseline GPT prompt, and a theory-driven GPT prompt (AllyGPT). To evaluate the responses, we conducted computational linguistic analyses to assess their psycholinguistic features and carried out a mixed methods evaluation. Three trained coders rated each message on favorability, conversational impact, and potential to change behavior and later participated in semistructured interviews to reflect on their evaluation process and perceptions of intervention effectiveness. Results: Linguistic analyses revealed that baseline GPT responses exhibited more emotionally positive and authentic language compared to AllyGPT responses, which showed a more analytical and assertive tone. Policy reminder messages were linguistically rigid and lacked emotional nuance. Human evaluation results showed that AllyGPT responses received the highest effectiveness ratings for low-incivil cyberaggression cases in 2 dimensions (favorability and changing behavior), and baseline GPT works better for mid and high levels for all effectiveness dimensions. For medium- and high-incivility aggressions, baseline GPT responses received the highest ratings across all 3 dimensions of effectiveness (favorability, discussion-shifting potential, and likelihood of changing bullying behavior), followed by AllyGPT, with policy reminders rated lowest. Qualitative feedback further emphasized that baseline GPT responses were perceived as natural and inclusive, while AllyGPT responses, although grounded in psychological theory, were sometimes viewed as overly direct. Policy reminders were considered clear but lacked persuasive impact. Conclusions: Our work showed that designing effective AI-generated bystander interventions requires a deep sensitivity to platform culture, social context, and user expectations. By combining psychological theory with adaptive, conversational design and ongoing feedback loops, future systems can better support bystanders, delivering interventions that are not only contextually appropriate but also socially resonant and behaviorally impactful. As such, this work serves as a foundation for scalable, human-centered AI systems that promote safer online spaces and users’ mental well-being.
<img src="https://jmir-production.s3.us-east-2.amazonaws.com/thumbs/2902e66b1ac5ac34f57f42bbc6adfe75" />

From peripheral initiation to central integration: a narrative review of the antihypertensive mechanisms of acupuncture in regulating autonomic nervous system homeostasis

Essential Hypertension (EH) is one of the most prevalent chronic cardiovascular diseases, imposing a significant burden on healthcare systems worldwide due to its high rates of disability and mortality. Long-term elevation of blood pressure leads to multi-organ damage in the heart, brain, and kidneys, resulting in severe complications such as coronary heart disease, stroke, and chronic kidney disease. Current treatment for hypertension primarily relies on pharmacological interventions. Although antihypertensive drugs have achieved notable success in controlling blood pressure, challenges remain, including poor long-term medication adherence, side effects, and inadequate blood pressure control in some patients with resistant hypertension. In parallel, acupuncture, a key modality of traditional Chinese medicine, has demonstrated unique advantages in hypertension management in recent years. Characterized by its holistic regulatory effects and minimal side effects, acupuncture is recognized by the World Health Organization as a recommended complementary and alternative therapy for hypertension, although its precise mechanisms remain incompletely understood. This review aims to summarize the “peripheral-central synergy” antihypertensive mechanism of acupuncture in regulating autonomic nervous system (ANS) homeostasis. Studies indicate that acupuncture primarily modulates autonomic homeostasis through the following pathways: (1) activating peripheral nerve fibers to convert physical stimulation into complex bioelectrical signals; (2) regulating synaptic neurotransmitter release and the expression of related membrane receptors; (3) modulating the synaptic microenvironment; (4) regulating the NTS-CVLM-RVLM neural circuit; and (5) modulating the HPA axis neuro-endocrine circuit. Through in-depth analysis, this review elucidates the multi-level and multi-dimensional impact of acupuncture therapy on primary hypertension, providing stronger evidence and a theoretical foundation for its clinical application.