AI Overviews don’t show up consistently across Google Search because the system learns where they’re useful and pulls them back when people don’t engage.
Robby Stein, Vice President of Product at Google Search, described in a CNN interview how Google tests the summaries, measures interaction, and reduces their appearance for certain kinds of searches where they don’t help.
How Google Decides When To Show AI Overviews
Stein explained that AI Overviews appear based on learned usefulness rather than showing up by default.
“The system actually learns where they’re helpful and will only show them if users have engaged with that and find them useful,” Stein said. “For many questions, people just ask like a short question or they’re looking for very specific website, they won’t show up because they’re not actually helpful in many many cases.”
He gave a concrete example. When someone searches for an athlete’s name, they typically want photos, biographical details, and social media links. The system learned people didn’t engage with an AI Overview for those queries.
“The system will learn that if it tried to do an AI overview, no one really clicked on it or engaged with it or valued it,” Stein said. “We have lots of metrics we look at that and then it won’t show up.”
What “Under The Hood” Queries Mean For Visibility
Stein described the system as sometimes expanding a search beyond what you type. Google “in many cases actually issues additional Google queries under the hood to expand your search and then brings you the most relevant information for a given question,” he said.
That may help explain why pages sometimes show up in AI Overview citations even when they don’t match your exact query wording. The system pulls in content answering related sub-questions or providing context.
For image-focused queries, AI Overviews integrate with image results. For shopping queries, they connect to product information. The system adapts based on what serves the question.
Where AI Mode Fits In
Stein described AI Mode as the next step for complicated questions that need follow-up conversation. The design assumes you start in traditional Search, get an Overview if it helps, then go deeper into AI Mode when you need more.
“We really designed AI Mode to really help you go deeper with a pretty complicated question,” Stein said, citing examples like comparing cars or researching backup power options.
During AI Mode testing, Google saw “like a two to three … full increase in the query length” compared to typical Search queries. Users also started asking follow-up questions in a conversational pattern.
The longer AI Mode queries included more specificity. Stein’s example: instead of “things to do in Nashville,” users asked “restaurants to go to in Nashville if one friend has an allergy and we have dogs and we want to sit outside.”
Personalization Exists But Is Limited
Some personalization in AI Mode already exists. Users who regularly click video results might see videos ranked higher, for example.
“We are personalizing some of these experiences,” Stein said. “But right now that’s a smaller adjustment probably to the experience because we want to keep it as consistent as possible overall.”
Google’s focus is on maintaining consistency across users while allowing for individual preferences where it makes sense.
If you’re tracking AIO presence week to week, the fluctuations may reflect user behavior patterns for different question types rather than algorithm changes.
The “under the hood” query expansion means content can appear in citations even without matching your exact phrasing. That matters when you’re explaining CTR drops internally or planning content for complex queries where Overviews are more likely to surface.
Looking Ahead
Google’s AI Overviews earn placement based on usefulness rather than appearing by default.
Personalization is limited today, but the direction is moving toward more tailored experiences that maintain overall consistency.
ChatGPT accounted for 64% of worldwide traffic share among gen AI chatbot websites as of January, while Google’s Gemini reached 21%, according to Similarweb’s Global AI Tracker.
Similarweb’s tracker (PDF link) measures total visits at the domain level, so it reflects people who go to these tools directly on the web. It doesn’t capture API usage, embedded assistants, or other integrations where much of the AI usage occurs now.
ChatGPT Down, Gemini Up In A Year Of Share Gains
The share movement is easiest to see year-over-year.
A year ago, Similarweb estimated ChatGPT accounted for 86% of worldwide traffic among tracked chatbot sites. Now, that figure is 64%. Over the same period, Gemini rose from 5% to 21%.
Other tools are much smaller by this measure. DeepSeek was at 3.7%, Grok at 3.4%, and Perplexity and Claude both at 2.0%.
Google has been promoting Gemini through products like Android and Workspace, which may help explain why it’s gaining share among users who access these tools directly.
Winter Break Pulled Down Total Visits
Similarweb pointed to seasonality during the holiday period:
“Driven by the winter break, the daily average visits to all tools dropped to August-September levels.”
That context matters because it helps distinguish overall category softness from shifts in market share.
Writing Tool Domain Traffic Declines
Writing and content generation sites were down 10% over the most recent 12-week window in Similarweb’s category view.
At the individual tool level, Similarweb’s table shows steep drops for several writing platforms. Growthbarseo was down 100%, while Jasper fell 16%, Writesonic dropped 17%, and Rytr declined 9%. Originality was up 17%.
These are still domain-level visit counts, so the clearest takeaway is that fewer people are going directly to specialized writing sites online. That can happen for several reasons, including users relying more on general assistants, switching to apps, or using these models through integrations.
Code Completion Shows Mixed Results
The developer tools category looked more mixed than the writing tools.
Similarweb’s code completion table shows Bolt down 39% over 12 weeks, while Cursor (up 8%), Replit (up 2%), and Base44 (up 49%) moved in different directions.
Traditional Search Looks Close To Flat
In Similarweb’s “disrupted sectors” view, traditional search traffic is down roughly 1% to 3% year-over-year across recent periods, which doesn’t indicate a sharp drop in overall search usage in this dataset.
The same table shows Reddit up 12% year-over-year and Quora down 53%, consistent with the idea that some Q&A behavior is being redistributed even as overall search remains relatively steady.
Why This Matters
When making sense of how AI is changing discovery and demand, these numbers can help you understand where direct, web-based attention is concentrating. That can influence which assistants you monitor for brand mentions, citations, and referral behavior.
Though you should treat this a snapshot, not the full picture. If your audience is interacting with AI through browsers, apps, or embedded assistants, your own analytics will be a better barometer than any domain-level tracker.
Looking Ahead
The next report should clarify whether category traffic rebounds after the holiday period and whether Gemini continues to gain share at the same pace. It will also be a useful read on whether writing tools stabilize or whether more of that usage continues to consolidate into general assistants and bundled experiences.
When most people hear the phrase “AI bias,” their mind jumps to ethics, politics, or fairness. They think about whether systems lean left or right, whether certain groups are represented properly, or whether models reflect human prejudice. That conversation matters. But it is not the conversation reshaping search, visibility, and digital work right now.
The bias that is quietly changing outcomes is not ideological. It is structural, and operational. It emerges from how AI systems are built, trained, how they retrieve and weight information, and how they are rewarded. It exists even when everyone involved is acting in good faith. And it affects who gets seen, cited, and summarized long before anyone argues about intent.
This article is about that bias. Not as a flaw or as a scandal. But as a predictable consequence of machine systems designed to operate at scale under uncertainty.
To talk about it clearly, we need a name. We need language that practitioners can use without drifting into moral debate or academic abstraction. This behavior has been studied, but what hasn’t existed is a single term that explains how it manifests as visibility bias in AI-mediated discovery. I’m calling it Machine Comfort Bias.
Image Credit: Duane Forrester
Why AI Answers Cannot Be Neutral
To understand why this bias exists, we need to be precise about how modern AI answers are produced.
AI systems do not search the web the way people do. They do not evaluate pages one by one, weigh arguments, or reason toward a conclusion. What they do instead is retrieve information, weight it, compress it, and generate a response that is statistically likely to be acceptable given what they have seen before, a process openly described in modern retrieval-augmented generation architectures such as those outlined by Microsoft Research.
That process introduces bias before a single word is generated.
First comes retrieval. Content is selected based on relevance signals, semantic similarity, and trust indicators. If something is not retrieved, it cannot influence the answer at all.
Then comes weighting. Retrieved material is not treated equally. Some sources carry more authority. Some phrasing patterns are considered safer. Some structures are easier to compress without distortion.
Finally comes generation. The model produces an answer that optimizes for probability, coherence, and risk minimization. It does not aim for novelty. It does not aim for sharp differentiation. It aims to sound right, a behavior explicitly acknowledged in system-level discussions of large models such as OpenAI’s GPT-4 overview.
At no point in this pipeline does neutrality exist in the way humans usually mean it. What exists instead is preference. Preference for what is familiar. Preference for what has been validated before. Preference for what fits established patterns.
Introducing Machine Comfort Bias
Machine Comfort Bias describes the tendency of AI retrieval and answer systems to favor information that is structurally familiar, historically validated, semantically aligned with prior training, and low-risk to reproduce, regardless of whether it represents the most accurate, current, or original insight.
This is not a new behavior. The underlying components have been studied for years under different labels. Training data bias. Exposure bias. Authority bias. Consensus bias. Risk minimization. Mode collapse.
What is new is the surface on which these behaviors now operate. Instead of influencing rankings, they influence answers. Instead of pushing a page down the results, they erase it entirely.
Machine Comfort Bias is not a scientific replacement term. It is a unifying lens. It brings together behaviors that are already documented but rarely discussed as a single system shaping visibility.
Where Bias Enters The System, Layer By Layer
To understand why Machine Comfort Bias is so persistent, it helps to see where it enters the system.
Training Data And Exposure Bias
Language models learn from large collections of text. Those collections reflect what has been written, linked, cited, and repeated over time. High-frequency patterns become foundational. Widely cited sources become anchors.
This means that models are deeply shaped by past visibility. They learn what has already been successful, not what is emerging now. New ideas are underrepresented by definition. Niche expertise appears less often. Minority viewpoints show up with lower frequency, a limitation openly discussed in platform documentation about model training and data distribution.
This is not an oversight. It is a mathematical reality.
Authority And Popularity Bias
When systems are trained or tuned using signals of quality, they tend to overweight sources that already have strong reputations. Large publishers, government sites, encyclopedic resources, and widely referenced brands appear more often in training data and are more frequently retrieved later.
The result is a reinforcement loop. Authority increases retrieval. Retrieval increases citation. Citation increases perceived trust. Trust increases future retrieval. And this loop does not require intent. It emerges naturally from how large-scale AI systems reinforce signals that have already proven reliable.
Structural And Formatting Bias
Machines are sensitive to structure in ways humans often underestimate. Clear headings, definitional language, explanatory tone, and predictable formatting are easier to parse, chunk, and retrieve, a reality long acknowledged in how search and retrieval systems process content, including Google’s own explanations of machine interpretation.
Content that is conversational, opinionated, or stylistically unusual may be valuable to humans but harder for systems to integrate confidently. When in doubt, the system leans toward content that looks like what it has successfully used before. That is comfort expressed through structure.
Semantic Similarity And Embedding Gravity
Modern retrieval relies heavily on embeddings. These are mathematical representations of meaning that allow systems to compare content based on similarity rather than keywords.
Embedding systems naturally cluster around centroids. Content that sits close to established semantic centers is easier to retrieve. Content that introduces new language, new metaphors, or new framing sits farther away, a dynamic visible in production systems such as Azure’s vector search implementation.
This creates a form of gravity. Established ways of talking about a topic pull answers toward themselves. New ways struggle to break in.
Safety And Risk Minimization Bias
AI systems are designed to avoid harmful, misleading, or controversial outputs. This is necessary. But it also shapes answers in subtle ways.
Sharp claims are riskier than neutral ones. Nuance is riskier than consensus. Strong opinions are riskier than balanced summaries.
When faced with uncertainty, systems tend to choose language that feels safest to reproduce. Over time, this favors blandness, caution, and repetition, a trade-off described directly in Anthropic’s work on Constitutional AI as far back as 2023.
Why Familiarity Wins Over Accuracy
One of the most uncomfortable truths for practitioners is that accuracy alone is not enough.
Two pages can be equally correct. One may even be more current or better researched. But if one aligns more closely with what the system already understands and trusts, that one is more likely to be retrieved and cited.
This is why AI answers often feel similar. It is not laziness. It is system optimization. Familiar language reduces the chance of error. Familiar sources reduce the chance of controversy. Familiar structure reduces the chance of misinterpretation, a phenomenon widely observed in mainstream analysis showing that LLM-generated outputs are significantly more homogeneous than human-generated one.
From the system’s perspective, familiarity is a proxy for safety.
The Shift From Ranking Bias To Existence Bias
Traditional search has long grappled with bias. That work has been explicit and deliberate. Engineers measure it, debate it, and attempt to mitigate it through ranking adjustments, audits, and policy changes.
Most importantly, traditional search bias has historically been visible. You could see where you ranked. You could see who outranked you. You could test changes and observe movement.
AI answers change the nature of the problem.
When an AI system produces a single synthesized response, there is no ranking list to inspect. There is no second page of results. There is only inclusion or omission. This is a shift from ranking bias to existence bias.
If you are not retrieved, you do not exist in the answer. If you are not cited, you do not contribute to the narrative. If you are not summarized, you are invisible to the user.
That is a fundamentally different visibility challenge.
Machine Comfort Bias In The Wild
You do not need to run thousands of prompts to see this behavior. It has already been observed, measured, and documented.
Studies and audits consistently show that AI answers disproportionately mirror encyclopedic tone and structure, even when multiple valid explanations exist, a pattern widely discussed.
Independent analyses also reveal high overlap in phrasing across answers to similar questions. Change the prompt slightly, and the structure remains. The language remains. The sources remain.
These are not isolated quirks. They are consistent patterns.
What This Changes About SEO, For Real
This is where the conversation gets uncomfortable for the industry.
SEO has always involved bias management. Understanding how systems evaluate relevance, authority, and quality has been the job. But the feedback loops were visible. You could measure impact, and you could test hypotheses. Machine Comfort Bias now complicates that work.
When outcomes depend on retrieval confidence and generation comfort, feedback becomes opaque. You may not know why you were excluded. You may not know which signal mattered. You may not even know that an opportunity existed.
This shifts the role of the SEO. From optimizer to interpreter. From ranking tactician to system translator, which reshapes career value. The people who understand how machine comfort forms, how trust accumulates, and how retrieval systems behave under uncertainty become critical. Not because they can game the system, but because they can explain it.
What Can Be Influenced, And What Cannot
It is important to be honest here. You cannot remove Machine Comfort Bias, nor can you force a system to prefer novelty. You cannot demand inclusion.
What you can do is work within the boundaries. You can make structure explicit without flattening voice, and you can align language with established concepts without parroting them. You can demonstrate expertise across multiple trusted surfaces so that familiarity accumulates over time. You can also reduce friction for retrieval and increase confidence for citation. The bottom line here is that you can design content that machines can safely use without misinterpretation. This shift is not about conformity; it’s about translation.
How To Explain This To Leadership Without Losing The Room
One of the hardest parts of this shift is communication. Telling an executive that “the AI is biased against us” rarely lands well. It sounds defensive and speculative.
I will suggest that a better framing is this. AI systems favor what they already understand and trust. Our risk is not being wrong. Our risk is being unfamiliar. That is our new, biggest business risk. It affects visibility, and it affects brand inclusion as well as how markets learn about new ideas.
Once framed that way, the conversation changes. This is no longer about influencing algorithms. It is about ensuring the system can recognize and confidently represent the business.
Bias Literacy As A Core Skill For 2026
As AI intermediaries become more common, bias literacy becomes a professional requirement. This does not mean memorizing research papers, but instead it means understanding where preference forms, how comfort manifests, and why omission happens. It means being able to look at an AI answer and ask not just “is this right,” but “why did this version of ‘right’ win.” That is an enhanced skill, and it will define who thrives in the next phase of digital work.
Naming The Invisible Changes
Machine Comfort Bias is not an accusation. It is a description, and by naming it, we make it discussable. By understanding it, we make it predictable. And anything predictable can be planned for.
This is not a story about loss of control. It is a story about adaptation, about learning how systems see the world and designing visibility accordingly.
Bias has not disappeared. It has changed shape, and now that we can see it, we can work with it.
“Is there any difference between how AI systems handle JavaScript-rendered or interactively hidden content compared to traditional Google indexing? What technical checks can SEOs do to confirm that all page critical information is available to machines?”
For several years now, SEOs have been fairly encouraged by Googlebot’s improvements in being able to crawl and render JavaScript-heavy pages. However, with the new AI crawlers, this might not be the case.
In this article, we’ll look at the differences between the two crawler types, and how to ensure your critical webpage content is accessible to both.
How Does Googlebot Render JavaScript Content?
Googlebot processes JavaScript in three main stages: crawling, rendering, and indexing. In a basic and simple explanation, this is how each stage works:
Crawling
Googlebot will queue pages to be crawled when it discovers them on the web. Not every page that gets queued will be crawled, however, as Googlebot will check to see if crawling is allowed. For example, it will see if the page is blocked from crawling via a disallow command in the robots.txt.
If the page is not eligible to be crawled, then Googlebot will skip it, forgoing an HTTP request. If a page is eligible to be crawled, it will move to render the content.
Rendering
Googlebot will check if the page is eligible to be indexed by ensuring there are no requests to keep it from the index, for example, via a noindex meta tag. Googlebot will queue the page to be rendered. The rendering may happen within seconds, or it may remain in the queue for a longer period of time. Rendering is a resource-intensive process, and as such, it may not be instantaneous.
In the meantime, the bot will receive the DOM response; this is the content that is rendered before JavaScript is executed. This typically is the page HTML, which will be available as soon as the page is crawled.
Once the JavaScript is executed, Googlebot will receive the fully constructed page, the “browser render.”
Indexing
Eligible pages and information will be stored in the Google index and made available to serve as search results at the point of user query.
How Does Googlebot Handle Interactively Hidden Content?
Not all content is available to users when they first land on a page. For example, you may need to click through tabs to find supplementary content, or expand an accordion to see all of the information.
Googlebot doesn’t have the ability to switch between tabs, or to click open an accordion. So, making sure it can parse all the page’s information is important.
The way to do this is to make sure that the information is contained within the DOM on the first load of the page. Meaning, content may be “hidden from view” on the front end before clicking a button, but it’s not hidden in the code.
Think of it like this: The HTML content is “hidden in a box”; the JavaScript is the key to open the box. If Googlebot has to open the box, it may not see that content straightaway. However, if the server has opened the box before Googlebot requests it, then it should be able to get to that content via the DOM.
How To Improve The Likelihood That Googlebot Will Be Able To Read Your Content
The key to ensuring that content can be parsed by Googlebot is making it accessible without the need for the bot to render the JavaScript. One way of doing this is by forcing the rendering to happen on the server itself.
Server-side rendering is the process by which a webpage is rendered on the server rather than by the browser. This means an HTML file is prepared and sent to the user’s browser (or the search engine bot), and the content of the page is accessible to them without waiting for the JavaScript to load. This is because the server has essentially created a file that has rendered content in it already; the HTML and CSS are accessible immediately. Meanwhile, JavaScript files that are stored on the server can be downloaded by the browser.
This is opposed to client-side rendering, which requires the browser to fetch and compile the JavaScript before content is accessible on the webpage. This is a much lower lift for the server, which is why it is often favored by website developers, but it does mean that bots struggle to see the content on the page without rendering the JavaScript first.
How Do LLM Bots Render JavaScript?
Given what we now know about how Googlebot renders JavaScript, how does that differ from AI bots?
The most important element to understand about the following is that, unlike Googlebot, there is no “one” governing body that represents all the bots that might be encompassed under “LLM bots.” That is, what one bot might be capable of doing won’t necessarily be the standard for all.
The bots that scrape the web to power the knowledge bases of the LLMs are not the same as the bots that visit a page to bring back timely information to a user via a search engine.
And Claude’s bots do not have the same capability as OpenAI’s.
When we are considering how to ensure that AI bots can access our content, we have to cater to the lowest-capability bots.
Less is known about how LLM bots render JavaScript, mainly because, unlike Google, the AI bots are not sharing that information. However, some very smart people have been running tests to identify how each of the main LLM bots handles it.
Back in 2024, Vercel published an investigation into the JavaScript rendering capabilities of the main LLM bots, including OpenAI’s, Anthropic’s, Meta’s, ByteDance’s, and Perplexity’s. According to their study, none of those bots were able to render JavaScript. The only ones that were, were Gemini (leveraging Googlebot’s infrastructure), Applebot, and CommonCrawl’s CCbot.
More recently, Glenn Gabe reconfirmed Vercel’s findings through his own in-depth analysis of how ChatGPT, Perplexity, and Claude handle JavaScript. He also runs through how to test your own website in the LLMs to see how they handle your content.
These are the most well-known bots, from some of the most heavily funded AI companies in this space. It stands to reason that if they are struggling with JavaScript, lesser-funded or more niche ones will be also.
How Do AI Bots Handle Interactively Hidden Content?
Not well. That is, if the interactive content requires some execution of JavaScript, they may struggle to parse it.
To ensure the bots are able to see content hidden behind tabs, or in accordions, it is prudent to ensure the content loads fully in the DOM without the need to execute JavaScript. Human visitors can still interact with the content to reveal it, but the bots won’t need to.
How To Check For JavaScript Rendering Issues
There are two very easy ways to check if Googlebot is able to render all the content on your page:
Check The DOM Through Developer Tools
The DOM (Document Object Model) is an interface for a webpage that represents the HTML page as a series of “nodes” and “objects.” It essentially links a webpage’s HTML source code to JavaScript, which enables the functionality of the webpage to work. In simple terms, think of a webpage as a family tree. Each element on a webpage is a “node” on the tree. So, a header tag
, a paragraph
, and the body of the page itself
are all nodes on the family tree.
When a browser loads a webpage, it reads the HTML and turns it into the family tree (the DOM).
How To Check It
I’ll take you through this using Chrome’s Developer Tools as an example.
You can check the DOM of a page by going to your browser. Using Chrome, right-click and select “Inspect.” From there, make sure you’re in the “Elements” tab.
To see if content is visible on your webpage without having to execute JavaScript, you can search for it here. If you find the content fully within the DOM when you first load the page (and don’t interact with it further), then it should be visible to Googlebot and LLM bots.
Use Google Search Console
To check if the content is visible specifically to Googlebot, you can use Google Search Console.
Choose the page you want to test and paste it into the “Inspect any URL” field. Search Console will then take you to another page where you can “Test live URL.” When you test a live page, you will be presented with another screen where you can opt to “View tested page.”
How To Check If An LLM Bot Can See Your Content
As per Glenn Gabe’s experiments, you can ask the LLMs themselves what they can read from a specific webpage. For example, you can prompt them to read the text of an article. They will respond with an explanation if they cannot due to JavaScript.
Viewing The Source HTML
If we are working to the lowest common denominator, it is prudent to assume, at this point, LLMs can’t read content in JavaScript. To make sure that your content is available in the HTML of a webpage so that the bots can definitely access it, be absolutely sure that the content of your page is readable to these bots. Make sure it is in the source HTML. To check this, you can go to Chrome and right click on the page. From the menu, select “View page source.” If you can “find” the text in this code, you know it’s in the source HTML of the page.
What Does This Mean For Your Website?
Essentially, Googlebot has been developed over the years to be much better at handling JavaScript than the newer LLM bots. However, it’s really important to understand that the LLM bots are not trying to crawl and render the web in the same way as Googlebot. Don’t assume that they will ever try to mimic Googlebot’s behavior. Don’t consider them “behind” Googlebot. They are a different beast altogether.
For your website, this means you need to check if your page loads all the pertinent information in the DOM on the first load of the page to satisfy Googlebot’s needs. For the LLM bots, to be very sure the content is available to them, check your static HTML.
More Resources:
Featured Image: Paulo Bobita/Search Engine Journal
Your future customers are relying on answer engines to surface a single recommendation, not a list of options.
Yet most small businesses remain invisible to AI because their Google Business Profile information is incomplete, inconsistent, or structured in ways these AI chat systems cannot confidently interpret. The result is fewer calls, missed bookings, and lost revenue.
In this upcoming webinar session, Raj Madhavni, Co-Founder, Alpha SEO Pros at Thryv, will explain how AI assistants evaluate local businesses today and which signals most influence recommendations. He will also identify the common gaps that prevent businesses from being selected and outline how to address them before 2026.
The ranking signals AI assistants use to select local businesses
A practical roadmap to increase AI driven visibility, trust, and conversions in 2026
Why Attend?
This webinar gives small business owners and marketers a clear framework for competing in an AI driven local search environment. You will leave with actionable guidance to close visibility gaps, strengthen trust signals, and position your business as the one AI assistants recommend when customers ask.
Register now to prepare your business for local AI search in 2026.
🛑 Can’t attend live? Register anyway, and we’ll send you the on demand recording after the session.
Nandan Nilekani can’t stop trying to push India into the future. He started nearly 30 years ago, masterminding an ongoing experiment in technological state capacity that started with Aadhaar—the world’s largest digital identity system. Aadhaar means “foundation” in Hindi, and on that bedrock Nilekani and people working with him went on to build a sprawling collection of free, interoperating online tools that add up to nothing less than a digital infrastructure for society. They cover government services, digital payments, banking, credit, and health care, offering convenience and access that would be eye-popping in wealthy countries a tenth of India’s size. In India those systems are called, collectively, “digital public infrastructure,” or DPI.
At 70 years old, Nilekani should be retired. But he has a few more ideas. India’s electrical grid is creaky and prone to failure; Nilekani wants to add a layer of digital communication to stabilize it. And then there’s his idea to expand the financial functions in DPI to the rest of the world, creating a global digital backbone for commerce that he calls the “finternet.”
“It sounds like some crazy stuff,” Nilekani says. “But I think these are all big ideas, which over the next five years will have demonstrable, material impact.” As a last act in public life, why not Aadhaarize the world?
India’s digital backbone
Today, a farmer in a village in India, hours from the nearest bank, can collect welfare payments or transfer money by simply pressing a thumb to a fingerprint scanner at the local store. Digitally authenticated copies of driver’s licenses, birth certificates, and educational records can be accessed and shared via a digital wallet that sits on your smartphone.
In big cities, where cash is less and less common (just trying to break a bill can be a major headache), mobile payments are ubiquitous, whether you’re buying a TV from a high-street retailer or a coconut from a roadside cart. There are no fees, and any payment app or bank account can send money to any other. The country’s chaotic patchwork of public and private hospitals have begun digitizing all their medical records and uploading them to a nationwide platform. On the Open Network for Digital Commerce (ONDC), people can do online shopping searches on whatever app they want, and the results show sellers from an array of other platforms, too. The idea is to liberate small merchants and consumers from the walled gardens of online shopping giants like Amazon and the domestic giant Flipkart.
In the most populous nation on Earth—with 1.4 billion people—a large portion of the bureaucracy anyone encounters in daily life happens seamlessly and in the cloud.
At the heart of all these tools is Aadhaar. The system gives every Indian a 12-digit number that, in combination with either a fingerprint scan or an SMS code, allows access to government services, SIM cards, basic bank accounts, digital signature services, and social welfare payments. The Indian government says that since its inception in 2009, Aadhaar has saved 3.48 trillion rupees ($39.2 billion) by boosting efficiency, bypassing corrupt officials, and cutting other types of fraud. The system is controversial and imperfect—a database with 1.4 billion people in it comes with inherent security and privacy concerns. Still, in the most populous nation on Earth, a big portion of the bureaucracy anyone might encounter in daily life just happens in the cloud.
Nilekani was behind much of that innovation, marshaling an army of civil servants, tech companies, and volunteers. Now he sees it in action every day. “It reinforces that what you have done is not some abstract stuff, but real stuff for real people,” he says.
By his own admission, Nilekani is entering the twilight of his career. But it’s not over yet. He’s now “chief mentor” for the India Energy Stack (IES), a government initiative to connect the fragmented data held by companies responsible for generating, transmitting, and distributing power. India’s grids are unstable and disparate, but Nilekani hopes an Aadhaar-like move will help. IES aims to give unique digital identities not only to power plants and energy storage facilities but even to rooftop solar panels and electric vehicles. All the data attached to those things—device characteristics, energy rating certifications, usage information—will be in a common, machine-readable format and shared on the same open protocols.
Ideally, that’ll give grid operators a real-time view of energy supply and demand. And if it works, it might also make it simpler and cheaper for anyone to connect to the grid—even everyday folks selling excess power from their rooftop solar rigs, says RS Sharma, the chair of the project and Nilekani’s deputy while building Aadhaar.
Nilekani’s other side hustle is even more ambitious. His idea for a global “finternet” combines Aadhaarization with blockchains—creating digital representations called tokens for not only financial instruments like stocks or bonds but also real-world assets like houses or jewelry. Anyone from a bank to an asset manager or even a company could create and manage these tokens, but Nilekani’s team especially hopes the idea will help poor people trade their assets, or use them as loan collateral—expanding financial services to those who otherwise couldn’t access them.
It sounds almost wild-eyed. Yet the finternet project has 30 partners across four continents. Nilekani says it’ll launch next year.
A call to service
Nilekani was born in Bengaluru, in 1955. His family was middle class and, Nilekani says, “seized with societal issues and challenges.” His upbringing was also steeped in the kind of socialism espoused by the newish nation’s first prime minister, Jawaharlal Nehru.
After studying electrical engineering at the Indian Institute of Technology, in 1981 Nilekani helped found Infosys, an information technology company that pioneered outsourcing and helped turned India into the world’s IT back office. In 1999, he was part of a government-appointed task force trying to upgrade the infrastructure and services in Bengaluru, then emerging as India’s tech capital. But Nilekani was at the time leery of being viewed as just another techno-optimist. “I didn’t want to be seen as naive enough to believe that tech could solve everything,” he says.
Nilekani demonstrates the biometric technology at the heart of Aadhaar, the system he spearheaded that provides a unique digital identity number to all Indians.
PALLAVA BAGLA/CORBIS/GETTY IMAGES
Seeing the scope of the problem changed his mind—sclerotic bureaucracy, endemic corruption, and financial exclusion were intractable without technological solutions. In 2008 Nilekani published a book, Imagining India: The Idea of a Renewed Nation. It was a manifesto for an India that could leapfrog into a networked future.
And it got him a job. At the time more than half the births in the country were not recorded, and up to 400 million Indians had no official identity document. Manmohan Singh, the prime minister, asked Nilekani to put into action an ill-defined plan to create a national identity card.
Nilekani’s team made a still-controversial decision to rely on biometrics. A system based on people’s fingerprints and retina scans meant nobody could sign up twice, and nobody had to carry paperwork. In terms of execution, it was like trying to achieve industrialization but skip a steam era. Deployment required a monumental data collection effort, as well as new infrastructure that could compare each new enrollment against hundreds of millions of existing records in seconds. At its peak, the Unique Identification Authority of India (UIDAI), the agency responsible for administering Aadhaar, was registering more than a million new users a day. That happened with a technical team of just about 50 developers, and in the end cost slightly less than half a billion dollars.
Buoyed by their success, Nilekani and his allies started casting around for other problems they could solve using the same digitize-the-real-world playbook. “We built more and more layers of capability,” Nilekani says, “and then this became a wider-ranging idea. More grandiose.”
While other countries were building digital backbones with full state control (as in China) or in public-private partnerships that favored profit-seeking corporate approaches (as in the US), Nilekani thought India needed something else. He wanted critical technologies in areas like identity, payments, and data sharing to be open and interoperable, not monopolized by either the state or private industry. So the tools that make up DPI use open standards and open APIs, meaning that anyone can plug into the system. No single company or institution controls access—no walled gardens.
A contested legacy
Of course, another way to look at putting financial and government services and records into giant databases is that it’s a massive risk to personal liberty. Aadhaar, in particular, has faced criticism from privacy advocates concerned about the potential for surveillance. Several high-profile data breaches of Aadhaar records held by government entities have shaken confidence in the system, most recently in 2023, when security researchers found hackers selling the records of more than 800 million Indians on the dark web.
Technically, this shouldn’t matter—an Aadhaar number ought to be useless without biometric or SMS-based authentication. It’s “a myth that this random number is a very powerful number,” says Sharma, the onetime co-lead of UIDAI. “I don’t have any example where somebody’s Aadhaar disclosure would have harmed somebody.”
One problem is that in everyday use, Aadhaar users often bypass the biometric authentication system. To ensure that people use a genuine address at registration, Aadhaar administrators give people their numbers on an official-looking document. Indians co-opted this paperwork as a proof of identity on its own. And since the document—Indians even call it an “Aadhaar card”—doesn’t have an expiration date, it’s possible for people to get multiple valid cards with different details by changing their address or date of birth. That’s quite a loophole. In 2018 an NGO report found that 67% of people using Aadhaar to open a bank account relied on this verification document rather than digital authentication. That report was the last time anyone published data on the problem, so nobody knows how bad it is today. “Everybody’s living on anecdotes,” says Kiran Jonnalagadda, an anti-Aadhaar activist.
In other cases, flaws in Aadhaar’s biometric technology have caused people to be denied essential government services. The government downplays these risks, but again, it’s impossible to tell how serious the problem is because the UIDAI won’t disclose numbers. “There needs to be a much more honest acknowledgment, documentation, and then an examination of how those exclusions can be mitigated,” says Apar Gupta, director of the Internet Freedom Foundation.
Beyond the potential for fraud, it’s also true that the free and interoperable tools haven’t reached all the people who might find them useful, especially among India’s rural and poorer populations. Nilekani’s hopes for openness haven’t fully come to pass. Big e-commerce companies still dominate, and retail sales on ONDC have been dropping steadily since 2024, when financial incentives to participate began to taper off. The digital payments and government documentation services have hundreds of millions of users, numbers most global technology companies would love to see—but in a country as large as India, that leaves a lot of people out.
Going global
The usually calm Nilekani bristles at that criticism; he has heard it before. Detractors overlook the dysfunction that preceded these efforts, he says, and he remains convinced that technology was the only way forward. “How do you move a country of 1.4 billion people?” he asks. “There’s no other way you can fix it.”
The proof is self-evident, he says. Indians have opened more than 500 million basic bank accounts using Aadhaar; before it came into use, millions of those people had been completely unbanked. Earlier this year, India’s Unified Payments Interface overtook Visa as the world’s largest real-time payments system. “There is no way Aadhaar could have worked but for the fact that people needed this thing,” Nilekani says. “There’s no way payments would have worked without people needing it. So the voice of the people—they’re voting with their feet.”
A street vendor in Kolkata displays a QR code that lets him get paid via India’s Unified Payments Interface, part of the digital public infrastructure Nilekani helped build. The Reserve Bank of India says more than 657 million people used the system in the financial year 2024–2025.
DEBAJYOTI CHAKRABORTY/NURPHOTO/GETTY IMAGES
That need might be present in countries beyond India. “Many countries don’t have a proper birth registration system. Many countries don’t have a payment system. Many countries don’t have a way for data to be leveraged,” Nilekani says. “So this is a very powerful idea.” It seems to be spreading. Foreign governments regularly send delegations to Bengaluru to study India’s DPI tools. The World Bank and the United Nations have tried to introduce the concept to other developing countries equally eager to bring their economies into the digital age. The Gates Foundation has established projects to promote digital infrastructure, and Nilekani has set up and funded a network of think tanks, research institutes, and other NGOs aimed at, as he says, “propagating the gospel.”
Still, he admits he might not live to see DPI go global. “There are two races,” Nilekani says. “My personal race against time and India’s race against time.” He worries that the economic potential of its vast young population—the so-called demographic dividend—could turn into a demographic disaster. Despite rapid growth, gains have been uneven. Youth unemployment remains stubbornly high—a particularly volatile problem in a large and economically turbulent country.
“Maybe I’m a junkie,” he says. “Why the hell am I doing all this? I think I need it. I think I need to keep curious and alive and looking at the future.” But that’s the thing about building the future: It never quite arrives.
Edd Gent is a journalist based in Bengaluru, India.
MIT Technology Review Explains: Let our writers untangle the complex, messy world of technology to help you understand what’s coming next. You can read more from the series here.
I am writing this because one of my editors woke up in the middle of the night and scribbled on a bedside notepad: “What is a parameter?” Unlike a lot of thoughts that hit at 4 a.m., it’s a really good question—one that goes right to the heart of how large language models work. And I’m not just saying that because he’s my boss. (Hi, Boss!)
A large language model’s parameters are often said to be the dials and levers that control how it behaves. Think of a planet-size pinball machine that sends its balls pinging from one end to the other via billions of paddles and bumpers set just so. Tweak those settings and the balls will behave in a different way.
OpenAI’s GPT-3, released in 2020, had 175 billion parameters. Google DeepMind’s latest LLM, Gemini 3, may have at least a trillion—some think it’s probably more like 7 trillion—but the company isn’t saying. (With competition now fierce, AI firms no longer share information about how their models are built.)
But the basics of what parameters are and how they make LLMs do the remarkable things that they do are the same across different models. Ever wondered what makes an LLM really tick—what’s behind the colorful pinball-machine metaphors? Let’s dive in.
What is a parameter?
Think back to middle school algebra, like 2a + b. Those letters are parameters: Assign them values and you get a result. In math or coding, parameters are used to set limits or determine output. The parameters inside LLMs work in a similar way, just on a mind-boggling scale.
How are they assigned their values?
Short answer: an algorithm. When a model is trained, each parameter is set to a random value. The training process then involves an iterative series of calculations (known as training steps) that update those values. In the early stages of training, a model will make errors. The training algorithm looks at each error and goes back through the model, tweaking the value of each of the model’s many parameters so that next time that error is smaller. This happens over and over again until the model behaves in the way its makers want it to. At that point, training stops and the values of the model’s parameters are fixed.
Sounds straightforward …
In theory! In practice, because LLMs are trained on so much data and contain so many parameters, training them requires a huge number of steps and an eye-watering amount of computation. During training, the 175 billion parameters inside a medium-size LLM like GPT-3 will each get updated tens of thousands of times. In total, that adds up to quadrillions (a number with 15 zeros) of individual calculations. That’s why training an LLM takes so much energy. We’re talking about thousands of specialized high-speed computers running nonstop for months.
Oof. What are all these parameters for, exactly?
There are three different types of parameters inside an LLM that get their values assigned through training: embeddings, weights, and biases. Let’s take each of those in turn.
Okay! So, what are embeddings?
An embedding is the mathematical representation of a word (or part of a word, known as a token) in an LLM’s vocabulary. An LLM’s vocabulary, which might contain up to a few hundred thousand unique tokens, is set by its designers before training starts. But there’s no meaning attached to those words. That comes during training.
When a model is trained, each word in its vocabulary is assigned a numerical value that captures the meaning of that word in relation to all the other words, based on how the word appears in countless examples across the model’s training data.
Each word gets replaced by a kind of code?
Yeah. But there’s a bit more to it. The numerical value—the embedding—that represents each word is in fact a list of numbers, with each number in the list representing a different facet of meaning that the model has extracted from its training data. The length of this list of numbers is another thing that LLM designers can specify before an LLM is trained. A common size is 4,096.
Every word inside an LLM is represented by a list of 4,096 numbers?
Yup, that’s an embedding. And each of those numbers is tweaked during training. An LLM with embeddings that are 4,096 numbers long is said to have 4,096 dimensions.
Why 4,096?
It might look like a strange number. But LLMs (like anything that runs on a computer chip) work best with powers of two—2, 4, 8, 16, 32, 64, and so on. LLM engineers have found that 4,096 is a power of two that hits a sweet spot between capability and efficiency. Models with fewer dimensions are less capable; models with more dimensions are too expensive or slow to train and run.
Using more numbers allows the LLM to capture very fine-grained information about how a word is used in many different contexts, what subtle connotations it might have, how it relates to other words, and so on.
Back in February, OpenAI released GPT-4.5, the firm’s largest LLM yet (some estimates have put its parameter count at more than 10 trillion). Nick Ryder, a research scientist at OpenAI who worked on the model, told me at the time that bigger models can work with extra information, like emotional cues, such as when a speaker’s words signal hostility: “All of these subtle patterns that come through a human conversation—those are the bits that these larger and larger models will pick up on.”
The upshot is that all the words inside an LLM get encoded into a high-dimensional space. Picture thousands of words floating in the air around you. Words that are closer together have similar meanings. For example, “table” and “chair” will be closer to each other than they are to “astronaut,” which is close to “moon” and “Musk.” Way off in the distance you can see “prestidigitation.” It’s a little like that, but instead of being related to each other across three dimensions, the words inside an LLM are related across 4,096 dimensions.
Yikes.
It’s dizzying stuff. In effect, an LLM compresses the entire internet into a single monumental mathematical structure that encodes an unfathomable amount of interconnected information. It’s both why LLMs can do astonishing things and why they’re impossible to fully understand.
Okay. So that’s embeddings. What about weights?
A weight is a parameter that represents the strength of a connection between different parts of a model—and one of the most common types of dial for tuning a model’s behavior. Weights are used when an LLM processes text.
When an LLM reads a sentence (or a book chapter), it first looks up the embeddings for all the words and then passes those embeddings through a series of neural networks, known as transformers, that are designed to process sequences of data (like text) all at once. Every word in the sentence gets processed in relation to every other word.
This is where weights come in. An embedding represents the meaning of a word without context. When a word appears in a specific sentence, transformers use weights to process the meaning of that word in that new context. (In practice, this involves multiplying each embedding by the weights for all other words.)
And biases?
Biases are another type of dial that complement the effects of the weights. Weights set the thresholds at which different parts of a model fire (and thus pass data on to the next part). Biases are used to adjust those thresholds so that an embedding can trigger activity even when its value is low. (Biases are values that are added to an embedding rather than multiplied with it.)
By shifting the thresholds at which parts of a model fire, biases allow the model to pick up information that might otherwise be missed. Imagine you’re trying to hear what somebody is saying in a noisy room. Weights would amplify the loudest voices the most; biases are like a knob on a listening device that pushes quieter voices up in the mix.
Here’s the TL;DR: Weights and biases are two different ways that an LLM extracts as much information as it can out of the text it is given. And both types of parameters are adjusted over and over again during training to make sure they do this.
Okay. What about neurons? Are they a type of parameter too?
No, neurons are more a way to organize all this math—containers for the weights and biases, strung together by a web of pathways between them. It’s all very loosely inspired by biological neurons inside animal brains, with signals from one neuron triggering new signals from the next and so on.
Each neuron in a model holds a single bias and weights for every one of the model’s dimensions. In other words, if a model has 4,096 dimensions—and therefore its embeddings are lists of 4,096 numbers—then each of the neurons in that model will hold one bias and 4,096 weights.
Neurons are arranged in layers. In most LLMs, each neuron in one layer is connected to every neuron in the layer above. A 175-billion-parameter model like GPT-3 might have around 100 layers with a few tens of thousands of neurons in each layer. And each neuron is running tens of thousands of computations at a time.
Dizzy again. That’s a lot of math.
That’s a lot of math.
And how does all of that fit together? How does an LLM take a bunch of words and decide what words to give back?
When an LLM processes a piece of text, the numerical representation of that text—the embedding—gets passed through multiple layers of the model. In each layer, the value of the embedding (that list of 4,096 numbers) gets updated many times by a series of computations involving the model’s weights and biases (attached to the neurons) until it gets to the final layer.
The idea is that all the meaning and nuance and context of that input text is captured by the final value of the embedding after it has gone through a mind-boggling series of computations. That value is then used to calculate the next word that the LLM should spit out.
It won’t be a surprise that this is more complicated than it sounds: The model in fact calculates, for every word in its vocabulary, how likely that word is to come next and ranks the results. It then picks the top word. (Kind of. See below …)
That word is appended to the previous block of text, and the whole process repeats until the LLM calculates that the most likely next word to spit out is one that signals the end of its output.
That’s it?
Sure. Well …
Go on.
LLM designers can also specify a handful of other parameters, known as hyperparameters. The main ones are called temperature, top-p, and top-k.
You’re making this up.
Temperature is a parameter that acts as a kind of creativity dial. It influences the model’s choice of what word comes next. I just said that the model ranks the words in its vocabulary and picks the top one. But the temperature parameter can be used to push the model to choose the most probable next word, making its output more factual and relevant, or a less probable word, making the output more surprising and less robotic.
Top-p and top-k are two more dials that control the model’s choice of next words. They are settings that force the model to pick a word at random from a pool of most probable words instead of the top word. These parameters affect how the model comes across—quirky and creative versus trustworthy and dull.
One last question! There has been a lot of buzz about small models that can outperform big models. How does a small model do more with fewer parameters?
That’s one of the hottest questions in AI right now. There are a lot of different ways it can happen. Researchers have found that the amount of training data makes a huge difference. First you need to make sure the model sees enough data: An LLM trained on too little text won’t make the most of all its parameters, and a smaller model trained on the same amount of data could outperform it.
Another trick researchers have hit on is overtraining. Showing models far more data than previously thought necessary seems to make them perform better. The result is that a small model trained on a lot of data can outperform a larger model trained on less data. Take Meta’s Llama LLMs. The 70-billion-parameter Llama 2 was trained on around 2 trillion words of text; the 8-billion-parameter Llama 3 was trained on around 15 trillion words of text. The far smaller Llama 3 is the better model.
A third technique, known as distillation, uses a larger model to train a smaller one. The smaller model is trained not only on the raw training data but also on the outputs of the larger model’s internal computations. The idea is that the hard-won lessons encoded in the parameters of the larger model trickle down into the parameters of the smaller model, giving it a boost.
In fact, the days of single monolithic models may be over. Even the largest models on the market, like OpenAI’s GPT-5 and Google DeepMind’s Gemini 3, can be thought of as several small models in a trench coat. Using a technique called “mixture of experts,” large models can turn on just the parts of themselves (the “experts”) that are required to process a specific piece of text. This combines the abilities of a large model with the speed and lower power consumption of a small one.
But that’s not the end of it. Researchers are still figuring out ways to get the most out of a model’s parameters. As the gains from straight-up scaling tail off, jacking up the number of parameters no longer seems to make the difference it once did. It’s not so much how many you have, but what you do with them.
Can I see one?
You want to see a parameter? Knock yourself out: Here’s an embedding.
This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology.
Europe’s drone-filled vision for the future of war
Last spring, 3,000 British soldiers deployed an invisible automated intelligence network, known as a “digital targeting web,” as part of a NATO exercise called Hedgehog in the damp forests of Estonia’s eastern territories.
The system had been cobbled together over the course of four months—an astonishing pace for weapons development, which is usually measured in years. Its purpose is to connect everything that looks for targets—“sensors,” in military lingo—and everything that fires on them (“shooters”) to a single, shared wireless electronic brain.
Eighty years after total war last transformed the continent, the Hedgehog tests signal a brutal new calculus of European defense. But leaning too much on this new mathematics of warfare could be a risky bet.Read the full story.
—Arthur Holland Michel
This story is from the next print issue of MIT Technology Review magazine. If you haven’t already, subscribe now to receive it once it lands.
MIT Technology Review Narrated: How one controversial startup hopes to cool the planet
Stardust Solutions believes that it can solve climate change—for a price.
The Israel-based geoengineering startup has said it expects nations will soon pay it more than a billion dollars a year to launch specially equipped aircraft into the stratosphere. Once they’ve reached the necessary altitude, those planes will disperse particles engineered to reflect away enough sunlight to cool down the planet, purportedly without causing environmental side effects.
But numerous solar geoengineering researchers are skeptical that Stardust will line up the customers it needs to carry out a global deployment in the next decade. They’re also highly critical of the idea of a private company setting the global temperature for us.
This is our latest story to be turned into a MIT Technology Review Narrated podcast, which we’re publishing each week on Spotify and Apple Podcasts. Just navigate to MIT Technology Review Narrated on either platform, and follow us to get all our new content as it’s released.
The must-reads
I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 Amazon has been accused of listing products without retailers’ consent Small shop owners claim Amazon’s AI tool sold their goods without their permission. (Bloomberg $) + It also listed products the shops didn’t actually have in stock. (CNBC) + A new feature called “Shop Direct” appears to be to blame. (Insider $)
2 Data centers are a political issue Opposition to them is uniting communities across the political divide. (WP $) + Power-grid operators have suggested the centers power down at certain times. (WSJ $) + The data center boom in the desert. (MIT Technology Review)
3 Things are looking up for the nuclear power industry The Trump administration is pumping money into it—but success is not guaranteed. (NYT $) + Why the grid relies on nuclear reactors in the winter. (MIT Technology Review)
4 A new form of climate modelling pins blame on specific companies It may not be too long until we see the first case of how attribution science holds up in court. (New Scientist $) + Google, Amazon and the problem with Big Tech’s climate claims. (MIT Technology Review)
5 Meta has paused the launch of its Ray-Ban smartglasses They’re just too darn popular, apparently. (Engadget) + Europe and Canada will just have to wait. (Gizmodo) + It’s blaming supply shortages and “unprecedented” demand. (Insider $)
6 Sperm contains information about a father’s fitness and diet New research is shedding light on how we think about heredity. (Quanta Magazine)
7 Meta is selling online gambling ads in countries where it’s illegal It’s ignoring local laws across Asia and the Middle East. (Rest of World)
8 AI isn’t always trying to steal your job Sometimes it makes your toy robot a better companion. (The Verge) + How cuddly robots could change dementia care. (MIT Technology Review)
9 How to lock down a job at one of tech’s biggest companies You’re more likely to be accepted into Harvard, apparently. (Fast Company $)
10 Millennials are falling out of love with the internet Is a better future still possible? (Vox) + How to fix the internet. (MIT Technology Review)
Quote of the day
“I want to keep up with the latest doom.”
—Author Margaret Atwood explains why she doomscrolls to Wired.
One more thing
Inside the decades-long fight over Yahoo’s misdeeds in China
When you think of Big Tech these days, Yahoo is probably not top of mind. But for Chinese dissident Xu Wanping, the company still looms large—and has for nearly two decades.
In 2005, Xu was arrested for signing online petitions relating to anti-Japanese protests. He didn’t use his real name, but he did use his Yahoo email address. Yahoo China violated its users’ trust—providing information on certain email accounts to Chinese law enforcement, which in turn allowed the government to identify and arrest some users.
Xu was one of them; he would serve nine years in prison. Now, he and five other Chinese former political prisoners are suing Yahoo and a slate of co-defendants—not because of the company’s information-sharing (which was the focus of an earlier lawsuit filed by other plaintiffs), but rather because of what came after. Read the full story.
—Eileen Guo
We can still have nice things
A place for comfort, fun and distraction to brighten up your day. (Got any ideas? Drop me a line or skeet ’em at me.)
+ It’s time to celebrate the life and legacy of Cecilia Giménez Zueco, the legendary Spanish amateur painter whose botched fresco restoration reached viral fame in 2012. + If you’re a sci-fi literature fan, there’s plenty of new releases to look forward to in 2026. + Last week’s wolf supermoon was a sight to behold. + This Mississippi restaurant is putting its giant lazy Susan to good use.
When the concept of “Web 3.0” first emerged about a decade ago the idea was clear: Create a more user-controlled internet that lets you do everything you can now, except without servers or intermediaries to manage the flow of information.
Where Web2, which emerged in the early 2000s, relies on centralized systems to store data and supply compute, all owned—and monetized by—a handful of global conglomerates, Web3 turns that structure on its head. Instead, data and compute are decentralized through technologies like blockchain and peer-to-peer networks.
What was once a futuristic concept is quickly becoming a more concrete reality, even at a time when Web2 still dominates. Six out of ten Fortune 500 companies are exploring blockchain-based solutions, most taking a hybrid approach that combines traditional Web2 business models and infrastructure with the decentralized technologies and principles of Web3.
Popular use cases include cloud services, supply chain management, and, most notably financial services. In fact, at one point, the daily volume of transactions processed on decentralized finance exchanges exceeded $10 billion.
Gaining a Web3 edge
Among the advantages of Web3 for the enterprise are greater ownership and control of sensitive data, says Erman Tjiputra, founder and CEO of the AIOZ Network, which is building infrastructure for Web3, powered by decentralized physical infrastructure networks (DePIN), blockchain-based systems that govern physical infrastructure assets.
More cost-effective compute is another benefit, as is enhanced security and privacy as the cyberattack landscape grows more hostile, he adds. And it could even help protect companies from outages caused by a single point of failure, which can lead to downtime, data loss, and revenue deficits.
But perhaps the most exciting opportunity, says Tjiputra, is the ability to build and scale AI reliably and affordably. By leveraging a people-powered internet infrastructure, companies can far more easily access—and contribute to—shared resource like bandwidth, storage, and processing power to run AI inference, train models, and store data. All while using familiar developer tooling and open, usage-based incentives.
“We’re in a compute crunch where requirements are insatiable, and Web3 creates this ability to benefit while contributing,” explains Tjiputra.
In 2025, AIOZ Network launched a distributed compute platform and marketplace where developers and enterprises can access and monetize AI assets, and run AI inference or training on AIOZ Network’s more than 300,000 contributing devices. The model allows companies to move away from opaque datasets and models and scale flexibly, without centralized lock in.
Overcoming Web3 deployment challenges
Despite the promise, it is still early days for Web3, and core systemic challenges are leaving senior leadership and developers hesitant about its applicability at scale.
One hurdle is a lack of interoperability. The current fragmentation of blockchain networks creates a segregated ecosystem that makes it challenging to transfer assets or data between platforms. This often complicates transactions and introduces new security risks due to the reliance on mechanisms such as cross-chain bridges. These are tools that allow asset transfers between platforms but which have been shown to be vulnerable to targeted attacks.
“We have countless blockchains running on different protocols and consensus models,” says Tjiputra. “These blockchains need to work with each other so applications can communicate regardless of which chain they are on. This makes interoperability fundamental.”
Regulatory uncertainty is also a challenge. Outdated legal frameworks can sit at odds with decentralized infrastructures, especially when it comes to compliance with data protection and anti-money laundering regulations.
“Enterprises care about verifiability and compliance as much as innovation, so we need frameworks where on-chain transparency strengthens accountability instead of adding friction,” Tjiputra says.
And this is compounded by user experience (UX) challenges, says Tjiputra. “The biggest setback in Web3 today is UX,” he says. “For example, in Web2, if I forget my bank username or password, I can still contact the bank, log in and access my assets. The trade-off in Web3 is that, should that key be compromised or lost, we lose access to those assets. So, key recovery is a real problem.”
Building a bridge to Web3
Although such systemic challenges won’t be solved overnight, by leveraging DePIN networks, enterprises can bridge the gap between Web2 and Web3, without making a wholesale switch. This can minimize risk while harnessing much of the potential.
AIOZ Network’s own ecosystem includes capacity for media streaming, AI compute, and distributed storage that can be plugged into an existing Web2 tech stack. “You don’t need to go full Web3,” says Tjiputra. “You can start by plugging distributed storage into your workflow, test it, measure it, and see the benefits firsthand.”
The AIOZ Storage solution, for example, offers scalable distributed object storage by leveraging the global network of contributor devices on AIOZ DePIN. It is also compatible with existing storage systems or commonly used web application programming interfaces (APIs).
“Say we have a programmer or developer who uses Amazon S3 Storage or REST APIs, then all they need to do is just repoint the endpoints,” explains Tjiputra. “That’s it. It’s the same tools, it’s really simple. Even with media, with a single one-stop shop, developers can do transcoding and streaming with a simple REST API.”
Built on Cosmos, a network of hundreds of different blockchains that can communicate with each other, and a standardized framework enabled by Ethereum Virtual Machine (EVM), AIOZ Network has also prioritized interoperability. “Applications shouldn’t care which chain they’re on. Developers should target APIs without worrying about consensus mechanisms. That’s why we built on Cosmos and EVM—interoperability first.”
This hybrid model, which allows enterprises to use both Web2 and Web3 advantages in tandem, underpins what Tjiputra sees as the longer-term ambition for the much-hyped next iteration of the internet.
“Our vision is a truly peer-to-peer foundation for a people-powered internet, one that minimizes single points of failure through multi-region, multi-operator design,” says Tjiputra. “By distributing compute and storage across contributors, we gain both cost efficiency and end-to-end security by default.
“Ideally, we want to evolve the internet toward a more people-powered model, but we’re not there yet. We’re still at the starting point and growing.”
Indeed, Web3 isn’t quite snapping at the heels of the world’s Web2 giants, but its commercial advantages in an era of AI have become much harder to ignore. And with DePIN bridging the gap, enterprises and developers can step into that potential while keeping one foot on surer ground.
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.
This content 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.
This week’s installment of new products and services for merchants includes marketing and advertising platforms, livestream tools, pop-up and form builders, fulfillment networks, AI voice agents, agentic commerce, and reverse logistics.
Got an ecommerce product release? Email updates@practicalecommerce.com.
PayPal Ads launches Transaction Graph Insights and Measurement.PayPal Ads has launched its Transaction Graph Insights and Measurement Program, providing merchants and advertisers with a view into shopper behavior, campaign effectiveness, and data-driven recommendations. The tools help understand cross-merchant, cross-surface shopper journeys to deliver brand-specific recommendations and independent campaign validation with third-party partners.
Slingwave brings AI-powered unified measurement to ecommerce.Slingwave has unveiled its AI-native marketing platform for ecommerce and direct-to-consumer brands. The system combines marketing mix modeling, agile marketing attribution, and experimentation with an intelligence layer and customized models that run millions of scenarios to deliver a clear plan for optimizing spend. According to Slingwave, the platform learns with every campaign, ensuring recommendations continuously improve.
Getsitecontrol updates widget builder for pop-ups and forms.Getsitecontrol, an email marketing platform for ecommerce, has released a redesigned widget editor that offers enhanced visual control when designing website pop-ups, forms, and teasers. The editor introduces a visual element tree that displays the complete structure of each widget in a sidebar. Getsitecontrol now allows users to fine-tune every visual aspect of their widgets, including margins, paddings, alignment, sizes, and colors. The result, says Getsitecontrol, is professional widgets that adapt to any screen size.
ReturnPro launches Shopify app.ReturnPro, a provider of returns management and reverse logistics, has launched its Returns Portal App on the Shopify App Store. The app combines returns initiation with a connected reverse supply chain and recommerce ecosystem. Shopify merchants gain access to ReturnPro’s infrastructure, including more than 1,000 partner drop-off locations. Merchants can resell refurbished inventory through their Shopify storefronts or distribute products across ReturnPro’s network of integrated marketplaces, creating secondary revenue streams and reducing write-offs.
ReturnPro
Stord acquires Shipwire to expand its fulfillment network.Stord, a logistics provider for pre-purchase, checkout, delivery, and returns, has acquired Shipwire, a subsidiary of Ceva Logistics. Stord says the acquisition continues its expansion of fulfillment networks by adding 12 locations, strengthening its presence in Europe, and maintaining access to Ceva’s global network of warehouses through Shipwire’s existing logistics agreements. Ceva manages 120 million square feet of warehouse space worldwide.
Amazon launches Alexa+ for users to chat with its assistant.Amazon has launched an Alexa+ website that lets select users chat with its assistant via their browser. Users can access Alexa.com to get quick answers, explore complex topics, create content, and more. Alexa.com combines information with real-world actions, offering integrations across devices for shopping, home control, cooking, and entertainment, per Amazon. Customers with early access to Alexa+ can visit Alexa.com while logged into their Amazon account and start chatting.
ITTRackNap launches marketplace and subscription commerce platform.ITTRackNap, an AI-powered marketplace and subscription automation platform for cloud and technology providers, announced its U.S. launch. The platform enables managed service providers, telecommunications and connectivity providers, and technology distributors to launch and scale cloud and digital commerce faster and cost-effectively. RackNap streamlines and lowers the cost of channel back-office operations through native integrations with hyperscalers and portals, including Microsoft, Amazon Web Services, Google, and Acronis.
PubMatic launches AgenticOS for agent-to-agent advertising.PubMatic, an ad tech company, has launched AgenticOS, an operating system to orchestrate autonomous, agent-to-agent advertising across digital environments. AgenticOS deploys a three-layer framework to plan, transact, and optimize programmatic advertising: (i) an Nvidia-powered infrastructure layer, (ii) an application layer with embedded agentic capabilities to interpret intent through protocols such as the Ad Context and Model Context, and (iii) a transaction layer that connects agentic decisioning to PubMatic’s Activate buying platform.
PubMatic
eBay introduces credit notes for U.S. seller fees and tax reversals.eBay is issuing separate credit notes for all seller fees, charges, and tax reversals in the U.S. A credit note reduces or cancels an invoice. Each credit note will show the reduced amounts and reference to the original invoice. According to eBay, the update improves transparency and helps match charges with reversals.
Cloudhands launches cross-model AI platform.Cloudhands, a marketplace for AI tools, has announced a new unified platform that lets users move among leading models such as OpenAI, Anthropic, and Google while keeping their conversation history, documents, tasks, and creative work connected. Interested users can join the waitlist for the platform, which will launch early this year, per Cloudhands.
xAI launches Grok Business and Grok Enterprise.xAI, the chatbot natively integrated into X, has launched Grok Business and Grok Enterprise, two new tiers providing access to Grok 3, Grok 4, and Grok 4 Heavy. Grok Business offers a self-serve process for small-to-medium teams. For larger organizations, Grok Enterprise includes Grok Business plus Custom Single Sign-On, Directory Sync, and audit and security controls.
VoAgents launches enterprise voice AI platform for customer conversations.VoAgents, a provider of enterprise voice tools, has launched voice AI agents capable of handling inbound and outbound calls. The platform’s self-learning capability means voice agents improve with every interaction. Core platform features include customizable voice personalities and workflows tailored to brand requirements, calendar and customer-management integrations, real-time call recordings and transcripts, outbound campaign management, and more. VoAgents offers access to all leading language models, including OpenAI and Anthropic.