Estimated reading time at 200 wpm: 14 minutes
Imagine it is 1989. I am sitting before an Amstrad PCW, a cream-coloured machine with a small green screen, a substantial keyboard and a three-inch floppy disk drive. When the disk is accessed, the machine whirs. When the printer begins, it clatters with mechanical determination. I do not think of it as a computer. People call it a “word processor”.
Whether or not you agree our Fat Disclaimer applies
That description referred to the machine itself: a dedicated appliance for creating, revising, storing and printing documents. It occupied space on a desk and did its one job well.
Compared with a typewriter, it felt liberating. I could move backwards through a document, correct a mistake without reaching for correction fluid, insert a sentence or delete a paragraph without retyping everything after it. These functions now seem elementary. At the time, they changed the relationship between writing and revision.
The typewriter had made each sentence a physical commitment. A mistake stayed visible unless covered or erased; a major revision often meant a new page. The word processor loosened the link between the order words were first written and the order they finally appeared. Composition became more fluid. A document could stay unsettled for longer.
The machine did not contribute to the substance of what I wrote. It did not offer a better sentence, spot a weak argument, or ask whether I had missed something. It recorded my words and let me rearrange them — by my conscious intention alone.
That distinction matters today because the machines we write through no longer confine themselves to that rudimentary roles of 40-odd years ago. What began as an instrument for recording language has become something that can take part in the work from which language emerges. The word processor has not simply gained more features. It has begun to change category.
From appliance to application
The dedicated word processor belonged to a transitional period, standing between the typewriter and the general-purpose computer.
As personal computers grew more capable, the functions of the dedicated machine migrated into software. The word processor stopped being a physical object with one purpose and became an application installed alongside others. The change looked practical rather than philosophical: instead of buying a machine for writing, you bought a computer that could run a writing program. WordStar, WordPerfect, MacWrite and Microsoft Word competed on how well they could construct a document.
The underlying relationship stayed familiar — the human produced the content, the software recorded and arranged it — but small changes were accumulating. Spell-checking arrived, then grammar-checking. The software could now make rudimentary judgements about the text in front of it: questioning punctuation, flagging repetition, suggesting a sentence had grown too long. These systems were often crude, and sometimes made a sentence worse. But they marked a real shift. The software was no longer entirely passive. It had begun to inspect the language it processed.
A thesaurus could propose alternatives. Templates could determine structure. Autocorrect could alter what had been typed, sometimes before the writer noticed. None of this amounted to authorship — the machine still depended on the human for the ideas, the argument, most of the language — but the boundary between recording and contribution had started to move. The word processor was no longer only preserving the writer’s choices. It was starting to influence them.
The document becomes a shared space
The next change came through networks and the internet. Documents moved from local disks to remote servers, and a file could now be opened by several people at once. The word processor became a collaborative environment.
Two or more writers could work on the same text. One could leave a comment while another altered the paragraph beneath it. Changes could be tracked, earlier versions recovered. The document became less like a sheet of paper and more like a shared working space — and authorship shifted with it. A document might no longer belong to a single identifiable writer: one person could set the structure, another supply the evidence, another improve the language, another decide what stayed in the final version.
The software did not provide the ideas, but it mediated the interaction through which ideas developed. It preserved competing versions, displayed disagreements, let several minds converge on one text. The document was no longer simply the output of thought; it had become a place where thought was aggregated and negotiated.
This matters because present-day AI did not arrive in an untouched environment. Long before generative AI, software had already turned writing from a solitary act on a page into a dynamic process conducted through digital tools. Generative AI entered a space already prepared for it.
Anticipation and the unfinished thought
Predictive text introduced a further stage, at first a modest one — a phone guessing the word being typed from a sequence of key presses, then systems predicting the next word or phrase from what had already been written. The word processor had always followed behind the writer, recording what was already decided. Predictive systems began to move slightly ahead, offering language before the person had entered it. Each accepted suggestion saved a small amount of effort, but it also meant part of the sentence had originated outside the writer’s own act of composition. That was easy to overlook because it became so ordinary — yet it was the first sign that the machine could participate in producing the sentence itself.
Generative AI extended this from a word to a paragraph, an argument, a whole account. And with that came something more consequential than scale: the user no longer had to begin with a finished sentence. A request could start as a fragment —
“I have an idea about…”
“This argument doesn’t seem right…”
“What am I overlooking?”
Traditional tools generally needed the writer to know what they wanted to say before the tool could help much. A thought processor can be useful earlier than that. A person may sense something wrong with an argument without being able to name the defect; several ideas may be present but not yet connected. Human thinking doesn’t always begin with a complete statement waiting to be transcribed — we often discover what we mean by attempting to express it. The thought processor enters at exactly that unsettled stage, and the user’s own corrections to it —
“No, that’s not what I mean.”
“You’ve assumed something I don’t accept.”
— can disclose the shape of a thought more clearly than the original prompt did. This resembles conversation, but the resemblance shouldn’t be pushed too far: a human interlocutor brings experience, intention and an independent relationship to reality that the system does not. Still, something real happens at the level of use. An internal, half-formed process becomes external, visible and revisable.
Why “thought processor”?
My first inclination was to call this new tool a cognitive processor. The term has some accuracy — generative systems classify, compare, summarise, infer patterns and produce explanations. Their outputs resemble the products of cognitive work. But ‘cognitive‘ is broad and clinical. The word risks implying that the machine possesses cognition in the human sense. I settled instead on thought processor.
The term is not a claim that an AI model thinks. It implies nothing about consciousness, belief or an inner stream of experience. It describes a function within a human activity: the system receives an externalised form of human intention — a question, a fragment, an unresolved argument — and returns linguistic material that may help the person extend, organise or reconsider what they were trying to say. A calculator processes numerical representations without understanding quantity. A word processor manipulates symbols without knowing what a document means. A thought processor operates on representations of ideas without the human experience from which those ideas arise. Its output re-enters human cognition; the person reads it, judges it, and the machine’s language becomes material for another round of thinking.
The tool becomes part of a circuit: a thought is partially expressed; the system transforms or extends it. The human encounters the result, accepts or redirects it; a further expression follows. What matters is not any single response but the iteration between person and system.
Different tools in the same machine
Not everyone uses generative AI the same way. For one person it is a more powerful search box. For another, an automated writer producing a finished product on request. For another, an editor tightening existing text, or a tutor explaining unfamiliar material. All legitimate, none exhausting the possibilities.
Used as a thought processor, the system isn’t simply asked for an answer — it’s used to examine the problem. The user might ask it to identify assumptions, construct the strongest opposing argument, or show how a conclusion changes when one premise is removed. The value may lie less in the machine’s words than in the intellectual movement those words provoke, which asks more of the human, not less. Passive use invites acceptance; active use demands direction and repeated scepticism about whether the model has actually understood the task.
Knowledge doesn’t become unnecessary here — it becomes necessary differently. Without enough grounding in the subject, a user may not be able to tell a useful synthesis from an elegant fabrication. A novice tends to request an explanation; an experienced practitioner presents a tension within a field and uses the response to test a developing position. The same model can serve as a superficial answer generator or a demanding instrument of reflection — the difference lies partly in the task, and partly in who’s directing it.
The appearance of a mind, and its limits
Language has always been one of the main ways we infer the presence of another mind — we cannot observe consciousness directly, only speech and response, from which we infer an interior life. A generative system produces language that is responsive and sometimes strikingly apt, and the natural response is to attribute mentality: we say the model “thinks”, “knows”, “understands”. That’s harmless enough loosely used, but misleading once the metaphor is mistaken for an explanation. A thought processor can generate an account of grief without grieving, discuss doubt without doubting. Its language can resemble the product of an inner life without establishing that one exists — though this doesn’t make the output meaningless. Meaning arises in the encounter between the output, its context and the person interpreting it.
The system also has no direct access to thought. It receives only what the user supplies — words, documents, instructions — while unstated knowledge and tacit understanding stay outside the exchange unless expressed. A short description of a complicated situation invites the model to fill gaps with plausible assumptions that may not be true, and the response can still sound complete. Fluency intensifies the problem: a hesitant human answer signals its own uncertainty, but a generative system can present weak reasoning in polished language, so form and substance are easily confused. Coherence is not truth, and articulation is not understanding.
There’s a further risk in original work specifically. A model trained on existing language is drawn towards recognisable patterns, and may translate an unusual, developing idea into something more familiar — removing exactly what was valuable in it. Used uncritically, the thought processor can standardise thought. Used critically, it can show the user where that standardising pressure is coming from.
Error, invention and misplaced confidence
A word processor could preserve an error its user typed. A thought processor can introduce errors of its own — a false proposition, an invented source, a persuasive explanation built on a mistaken premise. These systems are designed to produce plausible continuations, not to guarantee truth, so verification remains essential wherever facts matter.
The distinctive danger isn’t that the model can be wrong — people are wrong too — but that it can be wrong with unusual fluency, speed and breadth. It can generate more material than a person can easily check. So, apparent efficiency conceals a transfer of labour: time saved in production may need to be spent in verification instead.
Responsible use isn’t a final review conducted after the machine has finished. It starts earlier, in how the task is framed, what context is supplied, and when the system shouldn’t be relied on at all. A model doesn’t bear professional duty or moral accountability. It may describe situations or things without inhabiting them. It can contribute possibilities. It cannot take responsibility for the choice among them.
Authorship after generation
The old word processor seemed to preserve a simple relationship between author and document — the person pressed the keys, the machine recorded the result. Even then authorship was rarely that pure: writers relied on editors, dictionaries, ideas encountered elsewhere. Generative AI just makes the mixed character of authorship harder to ignore. A sentence may be proposed by the machine, altered by the user, tested in dialogue, rewritten and finally incorporated into a larger argument, with no single moment at which it became “the author’s own”.
So the question isn’t only who first produced the words. It’s who established the purpose of the work, who determined what was relevant, who rejected the unsuitable material, who checked the claims, who decided the final document expressed a position worth defending — and who accepts responsibility for it. Authorship increasingly depends on intellectual control rather than manual origination of every phrase.
None of this means disclosure stops mattering, or that a person may put their name to material they haven’t understood or reviewed. The signature should signify adoption, not mere possession: to sign a document is to say the work has passed through human judgement and that you’re prepared to answer for it. A thought processor doesn’t remove that obligation. It just makes the obligation more explicit.
The human position
The arrival of the thought processor doesn’t make the human writer obsolete. It changes what’s expected of them. Less effort may be needed to produce a first formulation. More judgement is needed to decide whether that formulation is worth keeping. The scarce capacity is no longer the production of words — it may be the ability to direct, discriminate and decide.
A skilled user doesn’t just know how to get fluent output. They recognise when the model has misunderstood, when an argument has become suspiciously smooth, when an important ambiguity has been ironed out. They also know when to stop processing. An idea can be revised indefinitely — every paragraph made more polished, every objection answered — until the document is technically competent and intellectually lifeless. Human judgement includes knowing when further improvement has become dilution, and when to preserve roughness because the roughness carries meaning.
The thought processor is therefore neither an autonomous thinker nor simply a faster word processor. It is powerful because it can externalise and return our expressed ideas. It is limited because it depends on language, context and human judgement. It becomes dangerous when fluency is mistaken for knowledge; valuable when its output is treated as material for thought, not a substitute for it.
Return to the green screen
I return in memory to the Amstrad PCW. Its green screen displayed only what I had entered. It did not interrupt, propose or challenge. It could preserve my thoughts only after I had already found the words to express them.
The thought processor belongs to another world. It can meet the user before the formulation is complete, offer a structure, expose an assumption, respond to an idea while that idea is still uncertain. None of this proves the machine thinks. The thought remains human in its purpose and its consequences; the model processes representations drawn from human language and returns further representations; the person decides whether anything meaningful has occurred.
The machine on my desk once waited faithfully for my words. Its descendant can now take part in the process by which I discover what those words need to say. That is more than word processing. It is why I call it a thought processor.
