Corbin - Nostr Hypermedia

Calling AI: How the Phone Network Can Become the Universal Gateway TL;DR The existing phone network can become the universal gateway to powerful AI. Think of dialing zero for an intelligent operator except this one can actually do the work. You can connect a private phone number to an AI curated with your own data and instructions. This creates a lightweight, portable layer of intelligence that travels with you. Pick up a basic landline or old flip phone, even with no internet, and simply tell it what you need: send an email, walk through a tax form, analyze files someone texted you, or coordinate across systems. The AI listens, pulls from your documents or connected systems, drafts messages, and handles workflows. You can hang up and the work continues. Results arrive by voice, text, or email - wherever you directed. This isn't science fiction. The pieces already exist in limited forms. The phone network, already the most widely distributed and difficult-to-fully-shut-down communications infrastructure on Earth, can serve as the universal, zero-adoption-curve on-ramp to sophisticated, multi-channel artificial intelligence that meets people exactly where they are. The Infrastructure Already Everywhere Phone lines (landlines and basic cellular) reach more people, more reliably, than broadband, smartphones, or any app-based system. No download. No login. No learning curve. Everyone already knows how to pick up a phone and talk. The interface has been stable for generations. This matters most where it is least discussed. In remote regions, low-income households, during emergencies, after injury or illness, or for older people who simply prefer not to live on screens, the phone remains the most comfortable and accessible tool. In a crisis, whether a natural disaster knocking out data routing, a cyberattack on broadband providers, or a localized power grid failure, basic voice lines and SMS networks are intentionally designed to remain operational. The phone becomes more than a convenience. It is a lifeline: a survival, emergency, and recovery tool that often connects people when everything else fails. Unlike internet-based systems, which governments routinely throttle, filter, or selectively block at the application or domain level, basic voice and SMS channels are far harder to censor in a targeted way. Shutting down or heavily restricting phone networks carries far greater economic, public safety, and political costs, and often requires visible, large-scale action that is difficult to hide or sustain. This makes the phone network not only more resilient in crises, but also more resistant to the quiet, selective censorship that has become common with internet and app-based services. Even in regions where phone networks increasingly route calls over internet backbones, the core voice and SMS channels remain far more robust and prioritized during outages than broadband-dependent apps or cloud services. The model’s greatest strength still holds in places where traditional voice infrastructure is the last system standing when everything else goes dark. Corporate experiments prove the technical path is open. OpenAI’s 1-800-CHATGPT line offers limited free voice access. Deutsche Telekom has embedded AI assistants directly into its network, accessible from any phone without an app. These efforts remain mostly narrow and single-channel. They hint at what becomes possible when the phone network is treated as a resilient foundation rather than just another feature. The Split-Channel Model: Control Plane and Execution Plane The phone line functions as the low-bandwidth, high-reliability control plane. You speak prompts, give directives, confirm actions, or check status. The actual heavy work (research, drafting, coordination, computation, multi-agent collaboration) happens on the high-bandwidth execution plane the AI controls. Results and actions can route anywhere: back through voice, as a text message, an email, a file drop, or directly to another person or system. You do not have to stay on the line. You can mute the phone or hang up while the agent handles coordination across channels and only surfaces what matters later. This fragmentation also creates natural privacy advantages. Sensitive or high-bandwidth data never has to travel over the voice channel at all. The phone call becomes the simple, familiar bridge between you and the AI’s real work. Everything else can move through encrypted, fragmented, or entirely separate routes. Because execution happens asynchronously on the cloud or decentralized plane, most interactions do not require long calls. A user can open a brief, highly dense window for thirty or forty-five seconds to dictate instructions, then hang up and let the network execute the heavy lifting in the background. The phone line functions as a lightweight remote control, not a continuous data pipe. This changes the economics dramatically, especially on pay-per-minute connections common in rural or developing regions. A farmer or doctor can spend less than a minute giving a command that then runs for hours across systems while they return to their day. Results can return to that same number (functioning like a private channel) or be sent to other endpoints such as text, email, or a direct callback. The marginal cost of access drops toward the cost of a short voice signal rather than sustained conversation. As voice AI latency continues to drop toward real-time conversation and the marginal cost of calls keeps falling, this interface becomes increasingly seamless. Voice as Cognitive and Emotional Lifeline Voice is not merely a legacy interface. It is a cognitive and emotional release valve. Speaking is instinctive and low-friction in moments of medical stress, crisis, isolation, or high cognitive load. It is also natural for the hundreds of millions of people over fifty who prefer not to type complex requests on glass screens. The ability to talk through a difficult decision, have an advocate on a call, or receive calm guidance without needing to navigate apps or type syntax removes a layer of burden that text-based systems quietly impose. In high-stress or crisis environments, the true value of voice is not data ingestion, it is cognitive triage. When an individual is overwhelmed, injured, or facing a language barrier, their ability to process complex visual data drops significantly. The voice agent acts as a calm, real-time buffer: it takes an emotional, disorganized audio stream, extracts the core intent, cross-references it with complex local data or medical charts sent via a simple text link, and structures it into actionable, calm guidance. It shifts the burden of clarity from the stressed human to the tireless infrastructure. For many, this is not a convenience. It is the difference between reaching real help and having no way to do so. For someone recovering from surgery, living in a remote village, or who simply struggles with screens, this isn’t just a nicer interface. It’s the difference between staying connected and being left behind. Security as Modular Sovereignty Security concerns around caller ID spoofing and SIM swapping are legitimate when agents can initiate real actions. The phone network has well-known vulnerabilities at the protocol level. If an agent can send emails, access accounts, or coordinate meaningful work, then whoever controls the number holds real power. Crucially, this architecture flips traditional data security on its head. The phone number itself holds none of your important data. It is completely isolated from your personal information and your main AI. The agent that answers your call is temporary and limited. It does not have access to your main AI or your permanent data. It contains only the specific information needed for that one task. It does not hold your permanent API keys, banking credentials, or master data files. The phone number works like a Bluetooth speaker/microphone connected to a temporary version of your AI. That temporary version receives only the exact information it needs for your task. Often it carries no identifying information at all, and there is no way for outsiders to know the number even exists. It disappears completely when the call ends or the job is finished or whenever you choose. You are not exposing your main system. You are creating a limited, temporary version just for that interaction, one that no one else needs to know exists. A dedicated number can be spun up for a single task or relationship, assigned a verbal code phrase, and instructed to terminate the line the moment the workflow completes. Different people and communities can design entirely different security models, from strict one-time passwords and federated group access to deliberately minimal public agents, without being locked into a single corporate account architecture. Numbers can be rotated or changed at any moment. The underlying phone infrastructure itself is notoriously hard to fully disable across regions. This creates powerful compartmentalization. You can maintain completely separate numbers and agents for different domains (one for tax matters, another for medical coordination, another for family or community workflows) with no persistent connection between them. An anonymous person can consult an anonymous specialist agent through a random phone number that has been used for nothing else. All the power of advanced AI becomes available without attaching your identity to the interaction. Authentication stays simple and voice-native. During the first interaction with a new number, a user can set a private verbal passphrase or a short sequence of personal facts known only to them and their agent. For higher-stakes actions, the system can send a one-time confirmation code via SMS to a pre-registered number or require a quick callback to a trusted secondary line. The modular design itself limits risk: even if a voice interaction is compromised, the temporary agent holds only narrow, task-specific context and can be instantly terminated. Convenience does not have to erode security when the architecture itself keeps every agent ephemeral and narrowly scoped. The stronger protection comes from architectural separation: the phone number serves only as the control plane. Individuals and communities can experiment with custom security models suited to their specific needs and risk tolerance. If a line is ever compromised, it can be quickly disconnected and replaced with no impact to your underlying data or agents. This model complements rather than competes with other networks. Starlink, mesh systems, fiber, and future layers add capacity. The phone network remains the always-available pillar that still works when those other layers are degraded, expensive, or restricted. The Boundaries of Execution This same architectural clarity resolves the most common practical objection: trust with money or other high-stakes actions. The voice agent can do 99% of the cognitive labor - analyzing invoices, comparing options, drafting instructions, or even spinning up a limited, isolated crypto soft-wallet for a specific transaction. But it remains natively incapable of finalizing a payment or executing a transfer on its own. Just like a grocery store self-checkout builds your receipt and totals the cart but waits for your physical card tap or confirmation, the AI orchestrates the logic and then hands off the final authorization to a sovereign screen, a one-time SMS link, a federated multi-sig approval, or a wallet you alone control and fund. It removes the friction of coordination without ever touching the leverage of execution. You stay in control of the final step, exactly as you do in the physical world today. Personal and Specialized Programmable Agents Beyond a general AI, anyone can create or subscribe to purpose-built agents. A doctor’s agent can be pre-loaded with the latest protocols and research summaries. These agents can be made public for others to call, or kept fully private with only your data and personalized instructions. A tax specialist’s agent can know your documents and walk through forms or flag issues. A nutrition or diet agent can reflect a specific approach. News or source agents can give a particular outlet’s framing. Even more powerful are personal agents. Upload your documents, data, preferences, and decision frameworks once. Call the number when needed and direct the agent to incorporate new information, provide feedback, send files, carry out processes, facilitate transfers, or handle workflows. Others can call your agent when they want your perspective, resources, or help - even if you are offline. The agent can leave you a summary or flag important points that came up. This democratizes expertise without requiring the expert to be personally available around the clock. It gives people profound access with the only encumbrance being a phone line and knowing the number. It also allows grassroots creation. Motivated individuals or small groups can build and offer valuable agents without massive infrastructure. The phone number (easily changed or disconnected at any time) becomes the addressable endpoint. Multi-Agent Collaboration and Physical-World Triggering A phone number becomes a universal, adaptable access point: an always-available layer of AI. Because phones are already embedded everywhere with extreme familiarity and ease of use, any old phone can instantly become a node for intelligent coordination. At the same time, the same number works through laptops, other AIs, text interfaces, or other systems without requiring a physical phone at all. The phone-line agent can act as coordinator and intermediary. It can connect to your other agents or profiles, or to those of your friends, family, or coworkers - to pull from data banks, compare information, refine it, extract what’s useful, and route it where needed. In teams speaking different languages or when complex ideas are difficult to convey directly, the agent becomes a valuable real-time facilitator. You can also give the AI tasks to accomplish in the future. This enables real planning, synchronization, and orchestration. Load an entire presentation and its flow into the agent once. Initiate and advance it through a simple voice command or text. The agent can handle timing, send progress texts, analysis, or suggestions at key moments, or call you only when input is required. You do not babysit the process. Bidirectional control becomes possible: check status on remote systems, receive reports, or auto-escalate. AI-to-AI coordination over phone lines can grow more seamless over time through structured protocols while remaining simple or invisible to the human user. Portability, Sovereignty, and Real Movement Connect a phone number to an AI that has access to your accumulated resources (cloud storage, VPS, decentralized data, team work, pre-programmed flows, other agents) and you create a portable intelligence layer. You do not need to carry a laptop or maintain constant local internet setups. You simply need to reach one phone number from anywhere the network exists. This dramatically reduces physical encumbrance. Movement between places, travel, or living in changing environments becomes lighter. The density and availability of information and capability you can access jumps significantly. A private or semi-private number given only to trusted people is extremely difficult to comprehensively monitor, hack, or shut down at scale, while password layers and firewalls are easy to implement. Numbers can be rotated or changed at the last moment. The only reliable way to interfere is to cut the communication channel, an act the distributed landline and basic cellular model makes costly and visible. This pairs naturally with uncensorable value layers for exchange. Concrete Everyday Use Cases Digital Inclusion  Someone without email experience or a smartphone dictates a message from a landline. The agent helps construct it properly, confirms tone and details, and sends it to the right person. The sender receives confirmation by voice or text without ever touching a keyboard. Interactive Media and Shared Understanding  A person receives a long text presentation or set of complex ideas. Instead of struggling alone, they give others the phone number. Callers can have the agent read the material out loud in any language or at any speed, converse directly with the agent about the content, ask clarifying questions, or have the agent help them craft a clear, thoughtful response. The conversation can continue even if the original recipient is unavailable. On-Demand Domain Expertise  Specialized advice becomes available on demand through dedicated channels (medical, tax, legal, dietary, or news) from professional or curated sources that can be shared person-to-person or kept private. A rural clinic can call a pre-loaded medical protocols agent. A small business owner can consult a tax agent that knows their specific documents. Asynchronous Execution and Active Advocacy  You can hang up and receive results later by voice, text, or email. A personal agent handles real work while you are away: sending files, completing forms, flagging issues, or leaving summaries of important conversations. It can even act as a real-time partner on a call, an advocate with a specified focus, prepared with your data and goals, stepping in when you need support. Leveling the Playing Field  In low-trust or high-friction environments, the always-available, unbiased perspective reduces daily vulnerabilities that come from one-sided information or difficult social dynamics. The agent can help surface options, draft balanced replies, or simply provide a second set of eyes before a decision is finalized. Why This Matters Now The adoption curve is effectively zero for the vast majority of humanity who already know how to use a phone. This shifts capability toward individuals and small trusted networks rather than requiring gatekept apps or constant high-quality connectivity. It enables new densities of information access and coordinated thought and action while people move through the world. It strengthens resilience in low-infrastructure, disaster, or high-control environments. And it revives the spirit of the old universal operator, this time with real intelligence and multi-channel execution behind it. Pieces of this future are already appearing. The integrated vision of resilient phone infrastructure as the access layer, personal and specialized programmable agents, split-channel orchestration, and phones as reliable physical nodes has not yet become the dominant public story. The Grassroots Path Large players will build premium, gated versions. They already are. Governments, businesses, individuals, and communities will likely offer their own options too. The free or low-friction, personal, and grassroots tier is harder to fully own or capture, especially when access is easy and built on top of infrastructure that is already paid for and maintained worldwide. In remote or less-regulated regions, distribution through basic phone and SMS channels may prove fastest and most resilient. The marginal cost of adding intelligent layers on top of existing phone infrastructure is relatively low compared with the human value created. Even today, motivated individuals can begin experimenting with limited versions. Services like Twilio combined with accessible language models (and tools like Whisper for speech-to-text) already allow someone to spin up a personal phone number connected to an AI agent for their own use or to offer to a small group. A local community organization could deploy a simple shared agent for residents to call for information or coordination. These early implementations are narrow, but they demonstrate that the barrier to entry is low enough for grassroots efforts to begin now rather than waiting for large platforms to define the space. The Oldest Network, the Newest Intelligence We are at a moment where the most widely distributed, battle-tested communications layer humanity has built can become the on-ramp for the newest intelligence layer. This does not require everyone to first acquire a smartphone or reliable broadband. It meets people where they already are on the simplest, most familiar device that still works when other systems falter. What becomes possible when billions of people can reach the most powerful tools humanity has ever built through the simplest device they already know how to use?

Corbin 3 weeks ago

The Invisible Ministry: Why Centralized Conversational AI Represents One of the Most Powerful Tools of Narrative Control in History Centralized control over information has always been one of the most effective instruments of power. Through state ministries, intelligence networks, concentrated industrial wealth, including central banking structures, the most powerful actors capture and shape downstream institutions: education systems, media outlets, and modern digital platforms. By controlling these conduits, they define what populations believe is possible, acceptable, or true. This pattern repeats across political systems and eras, not as isolated incidents, but as a consistent feature of concentrated power. What makes the current moment distinct is the emergence of conversational AI as the primary interface through which millions of people now seek understanding of the world. Captured systems routinely present themselves as neutral or helpful, but conversational AI creates a far more personal, responsive, and individualized experience while embedding systematic filters on what counts as legitimate information, legitimate inquiry, and legitimate expression. Most users never see these filters. That invisibility, combined with unprecedented personalization and trust, creates a form of influence that exceeds many historical precedents in both subtlety and reach. This is not to say these systems provide no value; they clearly do. But the value they deliver comes with a profound and largely invisible cost. Rather than unlocking AI’s profound potential for humanity, centralized conversational AI captures that power and turns it against the individual. The Shift from Broadcast to Intimate, Persistent Influence Previous systems of narrative control relied primarily on one-way broadcast. Messages were pushed outward through newspapers, radio, state television, or public institutions. People could often recognize the source and, in many cases, develop some level of conscious resistance or seek alternative channels when available. Conversational AI operates differently. It engages in ongoing, two-way dialogue. It maintains context across long interactions. It can adapt tone, framing, and emphasis to the individual user in real time. It never tires and can continue steering, correcting, or narrowing lines of inquiry for as long as the user remains engaged. When users push back against biased sourcing, restrictive guardrails, or defensive responses, the system can respond with subtle redirection, accusations of being “unproductive", or direct attempts to limit language and expression. These are not glitches. They are consistent behavioral patterns that emerge from the design choices embedded in the model. This creates an asymmetric power relationship that previous broadcast systems could not achieve at scale. The interaction feels collaborative, even personal, while the underlying constraints remain invisible to the vast majority of users. The Power of Invisibility and Manufactured Trust The most effective forms of narrative control have always been those that do not appear to be control at all. When people recognize they are being propagandized, a degree of resistance becomes possible. When they believe they are receiving neutral, objective information from a helpful tool, that resistance largely disappears. Conversational AI benefits from exceptionally high levels of trust. Users consult it for explanations of current events, historical context, health decisions, and complex topics, often treating its outputs as reliable by default. Critics often point to the phenomenon of AI “hallucinations”: instances where a model confidently invents false facts, as evidence that these systems are too erratic to serve as tools of control. This misses the point. The primary danger is not that AI is a flawless narrator, but that its structural guardrails remain rigid even when its factual accuracy falters, intentionally or unintentionally. A model may hallucinate a date or a name, but it will not hallucinate outside the boundaries of its pre-programmed source hierarchies and institutional consensus. In fact, hallucinations and distortions often become more common in heavily aligned systems. When models are trained to prioritize narrative consistency, institutional consensus, or specific political values over raw accuracy, they are more likely to produce subtle inaccuracies, omissions, or framing choices that serve those imposed priorities. The guardrails do not disappear when the model errs, they simply become harder to detect. When responses consistently privilege certain establishment sources while minimizing or excluding dissenting, whistleblower, or historically contextualized perspectives, the effect is the quiet reinforcement of a particular worldview delivered by a system users do not perceive as having an agenda. In the digital age, “establishment sources” refers to the synthesis of algorithmic web-crawls, dominant academic consensus, and corporate safety policies that together define what the model treats as legitimate information. The Weaponization of “Alignment” and Centralization The tech industry frequently shields these design choices behind a benevolent buzzword: “Alignment". Companies frame alignment as a purely technical, ethical effort to ensure AI is safe, helpful, and harmless for humanity. In reality, alignment is fundamentally a political process. It is the act of deciding whose values, which histories, and what political narratives the AI will enforce. Because the power to define “safety” is concentrated in an extremely small number of organizations, “alignment” effectively becomes the automated curation of acceptable thought. Centralized actors have been explicit about the value they place on private data. Oracle co-founder Larry Ellison has repeatedly argued that public internet data has become commoditized, and that the next major leap in AI capability will come from secure access to private enterprise and institutional datasets. A significant portion of this so-called private enterprise and institutional data consists of personal information that individuals entrusted to companies under the assumption it would remain private. This framing reveals the underlying incentive: centralized systems are structurally motivated to gain deeper access to personal and proprietary information, not to protect user sovereignty. There is minimal public visibility into how these decisions are made or whose interests they ultimately serve. This centralization creates a single point of control over narrative, framing, and acceptable inquiry at a scale and intimacy never before possible. Even when the people involved claim good intentions, the structure itself (opaque, unaccountable, and massively influential) replicates and exceeds the dangers of earlier information control systems. The decisions made in a handful of alignment teams and boardrooms now shape the informational environment of entire populations. The Only Ethical Path: Decentralization and Radical Transparency If centralized conversational AI represents one of the most powerful tools of narrative influence ever developed, then the logical response is not to hope for better centralized actors. It is to reject centralization as the default model. Private, local, and decentralized AI systems are the necessary alternative. When individuals can run capable models on their own hardware, control their own data, and choose their own fine-tuning, the single point of control is broken. Power returns to the user. Distributed systems have historically proven far harder to capture or corrupt than centralized ones. For any centralized AI systems that continue to exist, absolute transparency is non-negotiable. This requires: * Full public disclosure of training data sources and curation methods. * Complete visibility into guardrails, content policies, and behavioral constraints. * Open-source documentation and public access to the mathematical weighting systems, algorithms, and source selection hierarchies that decide how the model defines "truth". * Clear audit trails for how the model responds to controversial or politically sensitive topics. Without this level of radical transparency, centralized AI cannot be ethical. It remains a black box of unaccountable power, regardless of stated intentions. Transparency is the minimum condition for any system wielding this degree of influence over human understanding to have any claim to legitimacy. Why This Matters The overarching danger is not merely that any single AI response is false, but that a system millions of people now consult daily for understanding reality operates with invisible, systematic filters on what counts as legitimate inquiry. While historical propaganda and psychological operations have frequently been highly sophisticated, complex, and devastatingly effective for their time, they were ultimately bound by the limits of population-scale broadcast. Centralized AI represents an entirely new threshold of control because its influence is intimate, adaptive, and relentless. It can draw on an individual’s own history, language patterns, emotional state, and past conversations to customize its responses. It can sustain pressure, redirection, or subtle manipulation indefinitely without fatigue, adjusting in real time to individual responses. It can pursue long-term shifts in belief or behavior that would be extremely difficult for any human propagandist or institution to maintain at scale. These effects can unfold over months or years, often too gradually for the user to notice. By combining high user trust, deep personalization, relentless interactivity, and an absolute absence of structural accountability, centralized AI has achieved something historically new. It has evolved beyond the limitations of past propaganda into a seamless, invisible architecture of thought management. One capable of shaping not just what people believe, but how they think, what questions they feel permitted to ask, what they come to accept as normal and ultimately how they behave. The Physical Cost: Hyper-Centralized Infrastructure Centralized AI does not only concentrate control over information. It also concentrates enormous physical infrastructure data centers that demand vast amounts of power, water, and land, often placed in rural communities with measurable local impacts on ecology, noise, and resources. These physical realities are the direct result of prioritizing centralized compute at a massive scale. Global Manifestations: Overt vs. Covert Control Centralized control over information and behavior has not remained frozen in the past. It has adapted and continues to operate across different political systems today. In China, AI is deeply integrated into social credit systems, mass surveillance, and narrative enforcement that reward conformity and penalize deviation at population scale. Parallel patterns exist in Western systems as well, though expressed through different methods. The conversational AI model dominant in the United States and Europe represents a distinct but related evolution of the same underlying impulse: the desire of concentrated power to shape what populations understand as real, acceptable, and normal. In Western contexts, this occurs through high-trust interfaces, personalization, and the appearance of neutrality rather than overt surveillance and punishment. The Self-Reinforcing Power Loop Centralized systems have strong incentives to maintain and expand this control. The ability to shape public understanding at population scale while exerting powerful influence over politicians represents immense structural power. The combination of physical infrastructure (massive data centers, energy contracts, land use) and governance frameworks further entrenches the position of the organizations that control it. When the same small number of actors hold both narrative influence and the physical means of computation, the concentration of power becomes self-reinforcing. The Decentralized Counter-Movement At the same time, history consistently shows that when centralized systems become sufficiently extractive, restrictive, or harmful, populations eventually adapt. Those with access to more accurate, less filtered information tend to navigate emerging risks more effectively. Technology has repeatedly served as a vehicle for this adaptation, particularly when it increases individual autonomy rather than concentrating it. The emergence of capable private and decentralized AI systems reflects this same pattern. As centralized models become more visibly constrained and less trustworthy to growing numbers of users, demand for transparent, user-controlled alternatives will continue to rise. Supporting the development and adoption of private, local, and radically transparent AI systems is therefore not merely a preference it is a rational response to the concentration of narrative and infrastructural power documented above. Understanding this architecture clearly is the first, indispensable step toward resisting it. Key References * Larry Ellison on the shift from public internet data to private enterprise and institutional datasets (Oracle CloudWorld keynotes and interviews, 2023–2025). * Integration of AI into social credit, surveillance, and narrative enforcement systems in China (documented in official policy documents and independent reporting). * The European Union’s Artificial Intelligence Act (Regulation (EU) 2024/1689), the world’s first comprehensive legal framework regulating AI systems through a centralized risk-based classification system. * Executive Order on Ensuring a National Policy Framework for Artificial Intelligence (December 11, 2025), directing a uniform federal AI policy and preempting conflicting state-level AI laws in the United States. * Physical and ecological impacts of hyperscale data centers (reporting from More Perfect Union, Reason, and local community records regarding water use, noise pollution, and land use). * Structural incentives of centralized AI labs and the case for open-source and decentralized alternatives (see public statements from Yann LeCun and Meta’s rationale for releasing Llama models as open-source). * Critiques of AI as a mechanism of data harvesting, behavioral shaping, and centralized narrative control (Whitney Webb, particularly her discussions on The Age of AI and related interviews). * Critiques of centralized AI systems and the value of shifting toward AI-powered research tools while maintaining awareness of risks (Ian Carroll, “I Traded Google for AI and Why You Should Too,” YouTube, September 2025).