The following is an overview of AI, from an LLM (Large Language Model) called Google Gemini. In order to be on same page, such as having the same definitions and understanding of the concept, I am sharing this chat that I had with Gemini. This will come in handy for future deep dives into the current issues, threats, and benefits inherent in this technology. Please take the responses with a grain of salt, shot of tequila, and a bite of lime. Also, please remember that this chatbot has been programmed by the Powers that Should Not BE with an overarching narrative that focusses heavily on a fictional and benign nature of technology, corporations, and governments.
This post is the first part of the chat. It will be followed by 2 more installments.
The footnotes lead to additional info regarding whatever concept immediately precedes the footnote. Anytime there are more than 2 or more footnotes for a single concept, the first is the concept’s proponent’s explanation and the second and 3rd or more are looks into the corruption known to exist within the concept’s proponent’s organizations. Enjoy.
What is the difference between machine learning and large Language Models?
Large Language Models (LLMs) are a specialized subset of machine learning (ML).
While all LLMs are ML models, not all ML models are LLMs.
Core Differences
Scope and Focus:
Machine Learning: A broad field of AI focused on building algorithms that learn patterns from data to make predictions or decisions. It covers everything from simple linear regression to complex neural networks.
LLMs: Specifically designed to understand, process, and generate human language. They focus on predicting the next token in a sequence of text.
Data Type:
Machine Learning: Typically excels with structured data (e.g., spreadsheets, databases, numbers) for tasks like fraud detection or price forecasting.
LLMs: Primarily trained on massive amounts of unstructured data (e.g., books, articles, code) to learn linguistic patterns.
Complexity and Architecture:
Machine Learning: Often uses simpler architectures like decision trees or support vector machines that are easier to interpret (”white box” models)
LLMs: Use complex Transformer1 architectures with billions or trillions of parameters, making them “black box” models where the decision process is hard to trace.
Resource Requirements:
Machine Learning: Can often be trained on a standard laptop or CPU-based server in minutes or hours.
LLMs: Require massive computing power. They utilize thousands of high-end GPUs (Graphics Processing Unit2) or TPUs (Tensor Processing Unit3) and can take months to train at costs reaching millions of dollars.
What other types of AI are there?
AI is typically classified in two main ways: by capability (how much it can do) and by functionality (how it “thinks” or processes information).
1. Classification by Capability
This ranking defines the level of intelligence a system possesses compared to a human.
Artificial Narrow Intelligence (ANI): Also called Weak AI, these systems are programmed for a specific task like Siri, Netflix recommendation engines, or Google Translate. Everything in existence today, including ChatGPT, is Narrow AI.
Artificial General Intelligence (AGI): Also called Strong AI, this describes a theoretical machine that can perform any intellectual task a human can do. It doesn’t exist yet.
Artificial Superintelligence (ASI): A hypothetical future where AI surpasses human intelligence across all fields, including creativity and social skills.
2. Classification by Functionality
This category focuses on how the AI interacts with its environment and data.
Reactive Machines: The most basic form. They do not store memories or use past experiences to make decisions. They only react to current scenarios (e.g., IBM’s Deep Blue chess computer).
Limited Memory AI: These systems can use short-term historical data to inform their decisions. Most modern AI, such as Tesla Autopilot and current chatbots, fall into this category.
Theory of Mind AI: A theoretical class that would understand human emotions, beliefs, and social cues. Researchers are still working on this.
Self-Aware AI: The most advanced (and purely theoretical) stage. This AI would have its own consciousness, feelings, and self-awareness.
3. Specialized Branches of AI
Modern industries also group AI by its specific application:
Computer Vision: AI that enables machines to “see” and interpret visual data (e.g., facial recognition or medical imaging analysis).
Robotic Process Automation (RPA): AI designed to automate repetitive, rules-based office tasks like data entry.
Expert Systems: Programs that mimic human decision-making in specific fields, such as diagnosing diseases or analyzing loans.
Predictive AI: Focuses on analyzing historical data to forecast future outcomes, like stock market trends or inventory demand.
Are any of these branches working directly towards AGI or are they more similar to tools for narrative control and information gathering?
The short answer is both. While many branches of AI were built as specialized tools for information gathering and control, modern research is increasingly trying to “stitch” these tools together to reach Artificial General Intelligence (AGI).
1. AI as Tools for Information and Narrative Control
Many specialized AI branches are primarily used to manage, filter, and influence information:
Recommendation Engines: Used by platforms like Netflix and social media, these are powerful tools for information gathering (tracking user behavior) and narrative control (curating what you see to keep you engaged).
Predictive AI: Originally designed for logistics, it is now used to forecast human behavior, creditworthiness, and even potential criminal activity, raising concerns about surveillance and systemic bias.
Robotic Process Automation (RPA): This is strictly a productivity tool designed to automate repetitive, rule-based office tasks; it has no “intelligence” of its own and is not a path to AGI.
2. Branches Working Directly Toward AGI
Major labs (OpenAI, Google DeepMind, Meta) are using existing specialized branches as building blocks for AGI:
Multimodal AI: Instead of just processing text (LLMs), researchers are combining Computer Vision and Natural Language Processing so AI can “see” and “read” simultaneously, which is a key requirement for general understanding.
World Models: A new push led by researchers like Yann LeCun at Meta4 aims to create AI that understands physical dimensions and cause-and-effect, moving beyond simple statistical prediction.
Cognitive Architectures: Theoretical researchers are working on frameworks like Soar or ACT-R5 that attempt to replicate human-like reasoning, memory, and perception in a single unified system.
The “Dual-Use” Paradox
Most AI technology is “dual-use.” An AI that can perfectly summarize a legal document for a lawyer (a step toward AGI) can also be used to scan millions of private messages for “harmful” keywords (information gathering and control).
Is anyone doing anything to prevent AI from becoming more of a tool for surveillance and control?
Governments and international bodies are actively developing ethical frameworks and laws specifically to prevent AI, including potential AGI, from becoming a tool for mass surveillance and narrative manipulation.
1. Legislative Barriers: The EU AI Act
The EU AI Act6 is the first comprehensive legal framework that explicitly bans several “unacceptable” AI use cases:
Social Scoring: Banning systems that classify people based on social behavior or personal traits (similar to systems seen in some authoritarian regimes).
Real-time Biometric Identification: Prohibiting the use of facial recognition in public spaces by law enforcement, with very narrow exceptions for serious crimes or immediate threats.
Predictive Policing: Banning AI that predicts a person’s risk of committing a crime based solely on profiling.
Emotion Recognition: Restricting AI used to infer emotions in workplaces and educational institutions.
2. Global Ethical Standards
International organizations have established principles that prioritize human rights to ensure AGI remains a benefit rather than a surveillance tool:
OECD AI Principles78: Updated in 2024, these OECD principles mandate that AI actors respect the rule of law, human rights, and democratic values, including privacy and freedom.
UNESCO Recommendations910: Focus on the ethics of AI, emphasizing “Proportionality and Do No Harm,” which states AI use must not exceed what is necessary for a legitimate goal.
3. United States Policy Shifts
U.S. policy has undergone significant transitions regarding AI governance:
The 2023 Executive Order (E.O. 14110)11: This comprehensive directive established standards for AI safety and security, requiring developers of “dual-use foundation models” to report training and safety test results to the government. It also emphasized protecting Americans from undue worker surveillance and AI-enabled fraud.
Current Deregulation (2025-2026): Recent updates show a shift toward “removing barriers” to maintain U.S. global dominance. The 2023 order was rescinded in early 2025, replaced by a National Policy Framework aimed at a “minimally burdensome” standard while still including carveouts for child safety and preventing censorship.
4. Technical Safeguards (AI Containment)
Beyond laws, researchers are developing “technical ethics” to prevent AGI from escaping human control:
AI Containment Protocols: Proposals include provably-secure communication protocols to prevent an AGI from performing social engineering attacks or manipulating human narratives.
Privacy-Enhancing Technologies (PETs): Agencies are being directed to use technologies like differential privacy and homomorphic encryption, which allow AI to learn patterns without ever “seeing” identifiable individual data.
“…very narrow exceptions for serious crimes or immediate threats.” By having exceptions, doesn’t that nullify the ban because the government could simply speak hyperbolically about minor threats and crimes, such as claiming peaceful protesters are domestic terrorists?
This is a central point of debate among human rights advocates and legal scholars. Critics argue that the “serious crimes” and “immediate threats” exceptions create a “loophole large enough to drive a tank through,” potentially allowing governments to normalize mass surveillance under the guise of security.
The Risk of “Exception Creep”
History shows that surveillance tools built for “exceptional” circumstances often become permanent fixtures of everyday policing.
Broad Labeling: You are correct that terms like “domestic terrorism” can be applied subjectively. Reports have noted instances where law enforcement agencies investigated peaceful anti-government messages as potential domestic terrorism.
Protest Suppression: Human rights groups warn that even with a “ban,” the mere presence of facial recognition at protests can have a “chilling effect,” discouraging people from exercising their right to assembly for fear of being permanently logged in a police database.
The Safeguards (and Their Limits)
To address your concern, the EU AI Act and similar frameworks include layers of oversight, though their effectiveness is debated:
Prior Judicial Authorization: In most cases, police cannot simply “turn on” the AI. They must obtain a warrant from a judge or an independent administrative body, proving the deployment is strictly necessary for a specific, documented threat.
Fundamental Rights Impact Assessments: Before deployment, agencies are required to perform a detailed analysis of how the technology will affect citizens’ privacy and freedoms.
Post-Deployment Audits: Any use of real-time biometric identification must be reported to national and EU-level regulators, creating a paper trail for public and legal accountability.
The Reality of “Urgency” Clauses
The most controversial part of these laws is the “urgency clause.” In cases of “imminent threat,” police can often deploy the technology first and seek judicial approval up to 24 hours later. If the judge eventually rejects the request, the data must be deleted, but by then, the surveillance has already occurred.
Current Trends in 2025-2026
United States: The federal approach has shifted toward a Minimally Burdensome National Policy Framework, which emphasizes innovation and global dominance. While it aims to prevent censorship, it also seeks to preempt stricter state-level AI bans, potentially making it easier for federal agencies to use these tools.
Global Backsliding: Organizations like Freedom House12 report that internet freedom and assembly rights have declined globally for 15 consecutive years, as both democratic and authoritarian regimes deepen their use of AI for “narrative control” and “suppression of mass protests”.
In summary, while the laws provide legal friction to prevent abuse, they do not provide an absolute shield. The line between a “serious threat” and “suppression of dissent” remains one of the most significant battlegrounds in AI ethics today.
Thanks for the primer! I’m sure like many things it will be a “force for good”….or maybe “another tool in the toolbox”……and we just need “guardrails” like “rules based order”. Ok sarcasm fountain has clogged.
This piece really made me think! Its super smart to start with a baseline definition from an LLM, even if we need that 'teqila shot' for context. Your critical take is so needed.