For a long time, the machine asked a very small question.
Did you mean this?
Spell-checkers had existed in research environments since the early days of computing. By the 1990s, the red underline had become ordinary. The machine watched what we typed, noticed when a word did not look right, and offered a correction.
We did not think of this as authorship.
We thought of it as a tool.
Then the question changed.
Is this the word?
Anyone who typed messages on an early mobile phone will remember T9. A numerical keypad offered only twelve buttons, and several letters had to share each number. T9 used a dictionary to predict which word the sequence of key presses was likely to mean.
The machine was no longer merely correcting language after we produced it.
It was anticipating language while we produced it.
That innovation has an interesting ancestry.
T9 grew out of work in assistive communication. Martin King had previously developed the Eyescan Communicator, which allowed people with severe physical disabilities to construct messages through eye movements. Cliff Kushler also worked on communication technologies for people who could not readily speak or type.
Constraint helped produce prediction.
The problem was not originally, How can we make texting more convenient?
It was closer to: How can a person communicate when every physical act required to produce language is expensive?
Once that problem exists, prediction becomes valuable. If the machine can correctly anticipate the rest of the word, the person does not have to produce every character.
Then the technology escaped the problem that helped produce it.
Millions of people carried predictive language around in their pockets.
And almost nobody worried that the telephone had become the author of their text messages.
The machine predicted.
The person selected.
That distinction seemed sufficient.
Stephen Hawking encountered the same problem under far more demanding conditions.
After a tracheotomy in 1985 left him unable to speak, communication became an engineering problem. He initially used a spelling card. In 1986 he received Equalizer, software developed by Walter Woltosz of Words Plus for people with severe physical disabilities.
Hawking could select from thousands of words and phrases. At first he controlled the system with his hand. Eventually the computer was mounted to his wheelchair. As his physical abilities declined further, the interface changed with him. By 2008, movement in his cheek could operate the selector through a sensor mounted on his glasses.
The physical channel became smaller and smaller.
The language did not.
But the narrowing channel made prediction increasingly valuable.
By 2014, Intel had developed the Assistive Context-Aware Toolkit, or ACAT, for Hawking. SwiftKey built a personalized language model using Hawking’s previous lectures and papers. The system learned the language of the person using it.
Type black, and it might suggest hole.
Hawking reportedly needed to type fewer than 20 percent of the characters ultimately appearing in his text.
That number sounds enormous.
It is also a useful warning about how poorly the percentage of machine activity measures authorship.
The computer could save most of the keystrokes while contributing almost none of the intellectual direction.
It was predicting Hawking from Hawking.
The machine supplied characters he did not physically enter.
Hawking still decided whether they belonged.
Something else happened in 2014.
Apple introduced QuickType.
Contextual prediction was simultaneously appearing in two very different rooms.
For Hawking, prediction increased the expressive bandwidth available through one surviving voluntary movement.
For millions of iPhone users, it made typing faster.
The histories do not need to merge for the parallel to matter.
In fact, the separation tells us something more interesting.
Constraint produced prediction more than once.
Technologies developed around disability did not necessarily remain at the edge of computing. Problems created by severe physical constraints forced engineers to ask unusually precise questions about the relationship between intention and expression.
How little physical input can carry a thought?
How much can the machine anticipate without taking over?
Where does assistance end and authorship begin?
Those questions eventually belonged to everyone.
Then the machine’s grammar changed again.
This is probably the rest.
Google’s search autocomplete had been offering likely queries since the 2000s. Smartphone keyboards became increasingly contextual. Google’s Smart Reply began offering short responses. In 2018, Smart Compose began completing sentences inside Gmail.
Prediction had become completion.
That sounds like a small change.
It isn’t.
With T9, I know the word I am trying to type. The machine proposes it and I recognize it.
With sentence completion, the machine may present language I have not yet physically begun.
The human job begins to change.
At the lower end of this progression, I select.
At the higher end, I increasingly evaluate.
Those are different cognitive acts.
Rejecting the wrong T9 word may cost a few keystrokes.
Rejecting a completed sentence requires deciding whether the sentence says what I intended.
Then generative AI moves the problem another step.
The machine no longer has to wait for the beginning of my sentence.
I can give it an intention.
A question.
A subject.
A paragraph.
An argument.
It can return candidates I did not begin composing.
The machine’s question has changed again.
Here are some possibilities.
Now I must do considerably more than select.
Is the statement true?
Does the argument follow?
Is this my meaning?
Is the example useful?
Is the language mine in any important sense?
Should the paragraph exist at all?
The human veto survives.
But the cost of exercising the veto has changed.
This gives us a ladder, but not a chronology.
The technologies overlap. They did not arrive obediently in conceptual order.
The ladder measures something else: how much of the linguistic work the machine is doing.
Detection: Something may be wrong.
Correction: Did you mean this?
Prediction: Is this the word?
Completion: This is probably the rest.
Generation: Here are candidates.
Collaboration: What shall we do with them?
But the ladder measures only one dimension.
It tells us how much language the machine supplies.
It does not tell us how much intellectual direction the machine supplies.
Those are different axes.
Hawking makes that distinction visible.
A machine can supply an enormous percentage of the physical construction of a sentence while supplying very little of its intellectual direction.
A generative system can do the reverse. It may contribute only one paragraph to a long essay, yet that paragraph may contain an argument the writer had not considered and cause the entire piece to change direction.
Counting words will not tell us what happened.
Counting keystrokes will not tell us who thought.
The percentage of text supplied by the machine and the degree of intellectual contribution are different measurements.
This matters because we are beginning to ask disclosure questions as though “AI-assisted” describes one activity.
It does not.
A machine can detect.
It can correct.
It can predict.
It can complete.
It can generate.
It can participate in an iterative exchange in which both human and machine repeatedly propose material and the human selects, rejects, tests, restructures and returns it.
Calling all of these simply “AI use” collapses the very distinctions we need in order to talk intelligently about authorship.
Stephen Hawking is useful here not because his circumstances excuse machine assistance.
He does not need an excuse.
His case reveals the inadequacy of the measurement.
If a computer supplied most of the characters because it had learned to predict Hawking’s language from Hawking’s own writing, it would be strange to conclude that the machine therefore deserved most of the authorship.
The tool reduced the distance between intention and expression.
That was its purpose.
Generative AI complicates the relationship because sometimes the machine does something else.
It introduces material.
Once that happens, selection begins turning into evaluation.
And perhaps that is the boundary worth watching.
Not whether a machine touched the sentence.
Machines have been touching our sentences for decades.
Not whether a machine predicted language.
We accepted that long ago.
The more useful question is what kind of work passed from the person to the machine—and what kind of work remained with the person.
The history does not give us a clean line.
It gives us something better.
A series of increasingly capable instruments, and a question that has been following them all along:
What, exactly, did the machine do?
WE&P by: EZorrillaMc&Co.
