Narrow AI, General AI, and What Actually Exists Today

Every AI system in production today is narrow AI, meaning it handles one specific type of task well and nothing else. General AI, a system that could match human-level reasoning across all domains, does not exist. Understanding this distinction protects you from hype and helps you focus on tools that actually work.

Narrow AI: what you can use right now

Narrow AI (also called weak AI, though it is far from weak at its designated task) is software trained to do one thing. M-Pesa's fraud detection analyzes transactions. Google Translate converts text between languages. ChatGPT generates and processes text. Your phone camera's portrait mode identifies people in a frame. Each system excels at its specific task and is completely useless at anything else.

The M-Pesa fraud detector cannot translate Swahili to English. Google Translate cannot detect fraudulent transactions. ChatGPT cannot recognize objects in photographs (unless specifically given that capability through a separate vision model). These systems have no understanding of the world. They have mathematical patterns for one job.

This is not a limitation to apologize for. Narrow AI is extraordinarily useful precisely because of its focus. A tool that detects credit card fraud with nearly all accuracy is more valuable in its context than a hypothetical tool that can do everything at a majority accuracy.

For Kenyan professionals and businesses, narrow AI is the only category that matters for practical decisions today. When you evaluate an AI tool, you are evaluating a narrow AI system. The right question is not "how intelligent is it?" but "how well does it perform the specific task I need?"

Examples of narrow AI Kenyans already use

Language models (ChatGPT, Claude, Google Gemini) are narrow AI trained on text. They generate, summarize, and analyze written content. They are impressively good at language tasks and genuinely bad at tasks requiring real-world knowledge verification or mathematical precision. Their narrow domain is language patterns.

Recommendation systems on YouTube, Netflix, and Spotify predict what you want to watch or listen to next, based on your history and the history of similar users. They are narrow AI trained on viewing/listening behavior.

Navigation and traffic prediction in Google Maps uses narrow AI to predict drive times and suggest routes. It processes GPS data from millions of phones to identify congestion patterns. Ask it to write you a poem, and you will get nothing.

Agricultural advisory tools analyze satellite imagery, weather data, and soil conditions to predict crop performance. Organizations working with Kenyan farmers have deployed these tools to advise on planting windows and input application. The AI works well for crop predictions and cannot do anything else.

Customer service chatbots on banking and telco apps classify incoming messages and generate responses. They handle common queries efficiently but escalate to humans when the conversation moves outside their trained scope. That escalation point is the boundary of their narrow capability.

General AI: what does not exist yet

General AI (also called artificial general intelligence, or AGI) would be a single system that matches human cognitive ability across every domain. It would understand context, learn new tasks without specific training, reason abstractly, and apply knowledge from one area to completely different areas.

No one has built this. Despite what headlines sometimes imply, no company has demonstrated AGI. What we have are narrow AI systems that have gotten very good at language processing, which creates the illusion of general intelligence because language touches so many parts of human life.

When ChatGPT writes a business plan, debugs code, explains biology, and tells jokes in the same conversation, it looks like a generally intelligent system. It is not. It is a narrow AI (trained on text) applying language patterns to different topics. It does not understand the business plan, the code, the biology, or the humor. It generates statistically likely text for each context.

The test for general AI is not "can it do many things?" but "can it learn a genuinely new task it was never trained for, in a domain it has never encountered, the way a human can?" No existing system passes that test.

When will AGI arrive?

Honest answer: nobody knows, and anyone who gives you a confident date is speculating.

Predictions from researchers and industry leaders range from within the next decade to never. The disagreement is not over minor details. Experts fundamentally disagree on whether current approaches (scaling up language models with more data and parameters) will eventually produce general intelligence or whether entirely new technical paradigms are needed.

For practical decision-making, the timeline for AGI does not matter. The tools available today are narrow AI. The tools available next year will be narrow AI (likely more capable narrow AI, but narrow nonetheless). Planning your business or career around the arrival of AGI is like planning your commute around teleportation. It might happen eventually, but you need a plan that works with what exists now.

Superintelligence: the even further horizon

Superintelligence refers to a hypothetical AI that surpasses human intelligence across every domain. It is further out than AGI, since we would need general AI before anything could go beyond it.

Superintelligence gets attention in films, books, and think pieces. It raises genuine long-term questions about safety and control. But it has zero relevance to your work today. No product you buy, no tool you evaluate, and no career decision you make in 2026 is affected by superintelligence timelines.

We mention it only so you can recognize when a conversation has moved from practical technology into philosophical speculation. Both conversations have value, but mixing them up leads to poor decisions.

Why the distinction matters for your decisions

Understanding the types of AI protects you in several ways.

Against oversold products. When a vendor says their software uses "advanced AI," you know to ask: what specific narrow task does it perform, and how well? The framing of narrow AI prevents you from expecting magical general capabilities from a tool that classifies invoices.

Against unnecessary fear. Worries about AI replacing all jobs are rooted in a general AI scenario that does not exist. The realistic picture is that narrow AI tools change specific tasks within jobs, automating some parts and creating demand for new human skills. That is worth preparing for. A machine that replaces all human cognition is not something to plan around today.

For better tool selection. Knowing that every AI tool is narrow means you evaluate it on its specific task performance. "Does this chatbot handle our ten most common customer questions accurately?" is a better evaluation criterion than "how intelligent is this AI?"

For learning efficiently. You do not need to understand all of AI to be productive. You need to understand the narrow AI tools relevant to your work. A Kenyan marketer needs to understand language models and image generation tools. A Kenyan accountant needs to understand document processing and data extraction tools. Narrow focus matches the narrow reality of AI technology.

What narrow AI can do for Kenyan businesses this year

Rather than waiting for some future breakthrough, here is what works right now with today's narrow AI.

Language models can draft emails, proposals, and reports in seconds. They can summarize long documents, translate between languages, and answer questions about text you provide. For any Kenyan professional who writes as part of their job, this is a genuine productivity tool today.

Automation platforms with AI capabilities can sort, classify, and route information. Customer support messages can be auto-categorized. Invoice data can be extracted from PDFs. Lead inquiries can be scored by quality.

AI-powered analytics can find patterns in business data that spreadsheets miss. Sales trends, customer behavior clusters, and operational inefficiencies become visible when you apply the right narrow AI tool to your data.

Each of these is a narrow AI system doing one thing well. Combined with automation workflows and clear thinking about where AI adds value versus where simple rules suffice, they represent real, accessible capability.

Our free welcome module focuses entirely on this practical layer: what narrow AI tools exist, how to use them, and how to combine them with automation for real results.

FAQ

Is ChatGPT narrow AI or general AI?

ChatGPT is narrow AI. Its narrow domain is language: generating, processing, and analyzing text. It appears broadly capable because language touches many topics, but it is fundamentally a text prediction system. It does not reason, understand, or learn from your conversation in the way a human would. Each conversation starts from the same trained model. The impressive breadth of topics it can discuss reflects the breadth of its training data, not genuine general intelligence.

Could narrow AI become general AI through improvements?

This is one of the biggest open questions in AI research. Some researchers believe that scaling up current approaches (more data, more parameters, more computing power) will eventually produce something resembling general intelligence. Others argue that current architectures have fundamental limitations that no amount of scaling will overcome. There is no scientific consensus. What is clear is that each improvement in narrow AI produces more capable narrow tools, and those tools are worth learning regardless of whether AGI arrives.

Should I wait for better AI before learning these tools?

No. The fundamentals of how to use AI effectively (writing good prompts, evaluating AI outputs, combining AI with automation, understanding when AI is the wrong tool) stay relevant regardless of which specific model is in use. People who learned to use early language models effectively adapted to newer models faster because the core skills transferred. Learning now puts you ahead.

What about "multimodal AI" that handles text, images, and audio?

Multimodal models (like GPT-4 with vision or Google Gemini) can process multiple input types, but they are still narrow in the sense that they perform specific tasks (describing images, transcribing audio, generating text from visual inputs). They are broader narrow AI, not general AI. The practical implication is that you can now use a single tool for tasks that previously required separate tools, but the underlying technology is still pattern-matching on training data, not general understanding.