04. The Natural Language Revolution, the Humanities Renaissance, and AI Literacy

Published 2026-09-14

Core Theme: The moment computers learned to understand human language instead of demanding code, the world turned 180 degrees. Reading, writing, conversation, and the humanities — long pushed into the shadow of technical disciplines — have returned as the most powerful superpower in human society.

What does it mean to instruct a computer in ordinary language?

Until very recently, making a computer perform a complex task required precise commands, specialised software syntax, or programming code. The machine set the terms of the conversation, and the human complied or was excluded.

That relationship has inverted. You now give instructions in ordinary human words — natural language — in much the same way you would speak to a colleague, an office assistant, or a senior adviser whose help you need.

Figure 4.1: Prompting in Generative AI — issuing instructions in plain natural language to instantly generate complex letters, reports, and analyses.

The act of giving those instructions is called prompting, and any question, command, context or requirement you supply to a model is a prompt. The word sounds slight, and it is worth resisting that impression. A prompt now occupies the position that source code occupied for seventy years: it is the thing that determines what the machine does. The difference is that the language it is written in is the one you already speak.

This is also where a quiet inequality appears. When the interface was code, the barrier was visible and everyone knew where they stood — you had studied programming or you had not. Now the barrier is invisible. Any officer can type a sentence, but the quality of what comes back varies enormously with the quality of the sentence, and nothing on screen tells you that your instruction was the weak part. The rest of this chapter is about what separates the officer who gets excellent results from the one who concludes the technology is overrated.

How does a machine get from your words to an answer?

Something has to happen between your sentence and the model's response. Three technical layers do the work, and they are worth knowing by name because they fail in different ways — and being able to tell which layer let you down is the difference between fixing a prompt and abandoning a tool.

Figure 4.2: The AI Cognitive Pipeline — how Natural Language Processing (NLP), Natural Language Understanding (NLU), Synthesis, and Natural Language Generation (NLG) operate step by step.

Your input

It begins with a need stated in plain language — say, "Write a magical story about a forest that glows." Eight words, no technical structure, no syntax to get wrong.

Natural language processing breaks the sentence apart

The first layer, NLP, takes the sentence apart mechanically. It separates the input into individual units and tokens — [Write], [magical], [story], [forest], [glows] — and identifies grammatical categories: which words are verbs, which adjectives, which nouns.

This layer deals in structure, not meaning. It is the equivalent of a reader recognising the words on a page before understanding the sentence, and it is the most mature and reliable of the three. When NLP fails, it is usually because of something mundane: unusual spelling, mixed scripts, a poorly scanned document.

Natural language understanding works out what you meant

The second layer is where the interesting work happens. NLU goes past word recognition to capture semantic intent — what the user actually wants, as distinct from what they literally typed. From those eight words it extracts a set of conclusions: this is a request for a story, the setting is a glowing forest, the intended effect is magical.

This is the layer that makes the technology feel like it works, because this is where the system gets it. It is also the layer most affected by how you write. A vague instruction gives NLU almost nothing to work with, and the model then fills the gap with the most statistically ordinary interpretation — which is precisely the experience people describe as "the AI gave me something generic." The output was not generic. The intent was.

Synthesis composes the answer internally

Third, the model consults the patterns in its trained memory and composes internally: the story's characters, the colours of the setting, the sequence of events. Nothing is retrieved from a library of stories. This is assembly from pattern, which is why the same prompt twice produces two different stories, both new.

Natural language generation writes it out

Finally the answer is produced as fluent prose — "Once, in a hidden glade, ancient trees with luminous leaves cast soft azure light…" — and, in a multimodal system, a matching illustration can be generated from the same understanding at the same time.

A caution belongs here, and it connects directly to Chapter 03. This last layer is the only one you ever see, and it is the layer optimised for fluency. Beautiful output is evidence that NLG worked. It is not evidence that NLU understood you correctly, and it is certainly not evidence that the content is true.

The same pipeline in a government office

The abstract example becomes concrete the moment it meets real work. When a Grama Niladhari — the government's front-line administrative officer for a village division — logs a citizen complaint into such a system, all three layers engage in sequence. NLP recognises the letters and words in the document, including handwriting. NLU understands that what it is reading is a boundary dispute, and that a report from the Land Surveyor will therefore be required. NLG drafts, within seconds, the formal investigation report to be forwarded to the Divisional Secretary.

The middle step is the one that matters. Recognising the words is transcription, and machines have done it for decades. Recognising that this complaint belongs to the class of disputes requiring a surveyor is a judgment that previously existed only in an experienced officer's head — and it is exactly the kind of judgment that used to be lost every time such an officer retired.

Why does the AI era put reading and writing back at the centre?

For four or five decades, our society leaned hard toward technical, white-collar careers. Computer science and engineering sat at the top of the social ladder while the humanities — literature, art history, language, philosophy — were relegated to second class. Parents steered children accordingly. Even musicians and writers came to feel they ought to learn computer code and think in the machine's language, because that was where the future clearly lay.

Now that the computer speaks fluent human language, that entire calculation reverses.

There is a well-known saying in our old society: "සකස්කඩ නොකියූ කට උගේ කට හුඹස් කට" — "a mouth that has never recited the Sakaskada is a hollow anthill." (The Sakaskada is a classical Sinhala recitation text; the proverb holds that the mouth of a person without reading, language and a literary foundation is as empty as an abandoned anthill.)

The proverb was a moral observation when it was coined. In the AI era it becomes a technical one, and the chain of dependency is worth following one link at a time:

• Only the person who has read can write.
• Only the person who can write can marshal the contents of their own mind into order.
• Only the person who can marshal their thoughts can think deeply.
• And a person who cannot read, write, or think has no question — no prompt — to ask an AI at all!

That last line is the one with teeth. The machine supplies unlimited capability and zero intent. It will not tell you what to ask, notice that your question was the wrong one, or mention that the answer you accepted was shallow. Every one of those judgments comes from the human, and every one of them is built out of reading.

This produces an outcome nobody in the 1990s would have predicted: the more capable these systems become, the more valuable a literary education becomes, because articulation is now the interface. The person who reads widely and writes precisely does not merely enjoy AI more. They extract measurably better work from it. Standing mute before a machine that could have done anything you asked is the characteristic failure of this era, and it is a failure of language, not of technology.

What are the four human skills that AI cannot supply?

In his 2005 Stanford Commencement Address, Apple's co-founder Steve Jobs offered an idea about creativity that has aged unusually well:

"Creativity is just connecting dots. When you connect the varied experiences you have gathered — the books you have read, the people you have met, the things you have lived through — you can create something the future has never seen."

The reason it has aged well is that AI has automated a great deal of production and none of the connecting. Figure 4.3 sets out the four capacities that remain entirely human.

Figure 4.3: Connecting the Dots — the four foundational human skills of the AI age: Write, Read, Listen, and Think.

Think: identifying the problem

No good prompt exists without clear, logical thinking behind it, because how you think determines what you are able to command. The model will answer the question you asked with great competence, and it has no mechanism whatsoever for noticing that you asked the wrong one. Diagnosing the actual problem — as opposed to the presenting symptom — remains the most valuable and least automatable step in any professional process.

Write: expressing the idea

Writing is the act of converting a complex need in your mind into words that another party can act on: the task, the background, the constraints, set down precisely. This was always a professional skill and is now an operational one. When you cannot write down what you want, the gap does not stay hidden — it comes back as a document that is subtly not what you needed, and you will spend longer fixing it than you would have spent specifying it.

Read: auditing critically

An AI produces a document in three seconds. Only an officer reading with fine-grained attention can verify its legal accuracy and separate what is true from what is merely fluent. This is the pillar most at risk of quiet erosion, because the output arrives looking finished — formatted, confident, professionally worded — and every visual signal we have learned to trust says it needs no checking. Reading properly means reading against the grain of that impression, and it is the specific skill that determines whether AI makes an organisation faster or merely faster at being wrong.

Listen: human sensitivity

The foundation stone of public service is listening to the pain and the problems of the public. A machine cannot feel, and it is not present. It cannot notice that a citizen at a counter is frightened rather than angry, or that the request being made is not the one the citizen actually came in with.

This is worth stating plainly, because it is the most commonly underrated of the four. Listening is not a soft skill adjacent to the real work. It is the mechanism by which the real problem gets identified in the first place — and it is the input the system cannot acquire without you.

Collecting the dots before connecting them

If you are to connect the dots, you must first collect the dots.

There is only one way to do that: read widely, listen to people with empathy, and look at the world with an open mind. When you hold dots of knowledge from many fields, AI becomes the instrument with which you connect them — and create administrative solutions nobody has seen before.

The inverse is the situation to avoid. Someone with few dots and a powerful model produces exactly what the model's training data suggests is average, at high speed, in excellent prose. The tool amplifies whatever is put into it, including the absence of anything to amplify.

What does the ATM grandmother teach us about the digital divide?

One example explains why the natural language revolution matters to the public service and to ordinary citizens better than any argument about capability — and it starts with a piece of security engineering that is entirely correct on paper.

In cybersecurity — see the NIST Digital Identity Guidelines (NIST SP 800-63B) — identity is verified using three factors. Something you know: a password or PIN. Something you have: a bank card or national identity card. Something you are: a fingerprint, a face, a retina — the class of factor on which Sri Lanka's national digital identity, SLUDI, is being built.

Banks secure ATM withdrawals with two of the three: the card you have, plus the number you know. This is sound design. Neither factor alone is sufficient, the factors are independent, and an attacker needs both.

Now go and stand near an ATM in a rural Sri Lankan town, and watch what actually happens.

An elderly woman or man arrives to withdraw money. They do not have the digital literacy to decode the English words on the screen, the unlabelled buttons, the nested menus that assume a mental model of how such machines work.

So what does that elder do?

They hand over their ATM card and their PIN — both factors, to a stranger — usually a young man standing nearby or the security guard, with a request: "Son, please take out four thousand rupees for me."

Sit with that for a moment, because the lesson generalises far beyond banking. The cryptography was not broken. No attacker defeated the two-factor design. The system was defeated by its interface, and the person it failed was precisely the person with least capacity to absorb the loss. A security control that a citizen cannot operate is not a security control. It is a control that has been outsourced, by the citizen, to whoever happens to be standing nearby.

This is the historic failure that natural language AI actually solves.

At the ATM of the near future — or at the information counter of a Divisional Secretariat — the citizen presses nothing and fills in nothing. She says, in her own voice and her own language:

"Son, check whether my pension has come in, and give me four thousand rupees."

The machine recognises her voice, her language, and her fingerprint, and delivers the service immediately, with no intermediary. Note what has happened to the security model: the biometric factor she cannot lose or give away has replaced the PIN she was handing to strangers. The system became more secure by becoming easier, which is the opposite of the usual trade-off and the reason this example is worth remembering.

This is the concrete meaning of closing the digital divide. Not cheaper devices, not more training courses — though both help — but an interface that does not require the citizen to become someone else before it will serve them.

🎬 Watch the animated story (2 min): The ATM That Learned Our Language — Banking in the Gen AI Era

What AI literacy does the public officer actually need?

The real AI literacy a public officer must possess is not mere typing skill, and it is not knowing which buttons to press in which application. That is the smallest part of it, and the part most likely to be obsolete within a year. Figure 4.4 sets out five competencies that hold regardless of which model is currently in fashion.

Figure 4.4: The Essential Skills Model of the Generative AI Era (Gen AI Essential Skills) — AI literacy, prompt strategy, domain expertise, ethical accountability, and the human-AI partnership.

AI literacy and data fluency

Knowing what these models can and cannot do — their genuine capabilities and their technical limits — plus the data practices underneath: governance, ethical sourcing, and the ability to detect the biases embedded in training data.

The practical value of this competency is knowing when not to use the tool. Someone who understands that a language model is a plausibility engine does not ask it for a circular number, and that single piece of restraint prevents a large share of the errors that reach real documents.

Prompt strategy and analysis

Designing prompts that are complex, staged and iterative rather than single-line questions, then verifying what comes back through fact-checking and deliberate quality control.

Note that verification is listed as part of prompting rather than as a separate activity. That placement is the point: a prompt is not finished when the output appears. Chapter 06 covers the design methods in full.

Domain and strategic insight

Deep administrative and legal context of the public service, used to identify which applications would bring genuine value to the institution, and to plan adoption across an organisation rather than in one enthusiast's corner of it.

This is the competency that determines whether AI produces a measurable result or a collection of impressive demonstrations. The scarce input is not knowledge of AI. It is knowledge of the work, and of which part of it is actually the bottleneck.

Responsible AI and ethics

Ethical frameworks and transparency, citizen data privacy and system security, and compliance with national data-protection law and regulation.

For an officer handling citizen data this is not a concluding formality but a precondition, since a single careless upload can expose information that was never yours to disclose. Chapter 08 covers the mechanics.

Human–AI partnership

At the centre of the diagram, because it is what the other four are for: working with AI agents as a team, in continuous and iterative collaboration, with constant skill development — the continuous-learning approach Idasara Academy is built around.

The word "partnership" is doing real work here. It is not a tool you pick up and put down, and not a replacement you hand your judgment to. It is a working relationship in which the machine supplies speed and breadth, you supply intent and judgment, and neither half produces good work without the other.

In the next chapter, we take up the most important human skill for getting the most accurate result from an AI — clear articulation — together with the curiosity of a small child, and study the art of prompting in practice.