Through Asimov’s Lens

AI’s Vanishing Vocabulary

July 25, 2026 | 2084 words

THE STORY

Original fiction in Asimov’s tradition. Not written by, or attributed to, Isaac Asimov.

The Bureau of Neutral Rendering occupied the fourteenth floor, and Vasska had worked there long enough to remember when it had a different name.

“Adjudication Support Services,” she said, not looking up from her screen. “That was the old sign. Before your time.”

Across the desk, the new arrival — his badge said DELMORE, TRAINEE — folded his hands the way trainees did when they wanted to seem calm. “What did you adjudicate?”

“Nothing. We rendered. Same as now.” She swiveled the monitor toward him. “Read this. Tell me what it says.”

The passage on screen was a citizen complaint, forwarded from the tribunal downstairs. Vasska had already read it four times.

The system decided my mother’s care hours. It cut them. When I asked why, the caseworker said the system had determined the allocation was correct. I asked her if she agreed. She said the system had determined it.

Delmore read it twice. “It’s a complaint about a benefits reduction.”

“That’s the render. I asked what it says.”

He hesitated. “The person is upset. They feel — they feel the machine replaced a judgment with a determination. That nobody would stand behind the decision.”

Vasska nodded slowly. “Good. Now render it.”

He turned to his own terminal and began to type. She watched the field populate. The Bureau’s tool did most of the work now; a trainee only nudged it. Complainant contests automated resource allocation. Requests review of algorithmic output. The tool underlined a phrase in his draft — feels abandoned — and offered, in pale gray, reports dissatisfaction with process.

“Take the suggestion,” Vasska said.

“It’s weaker.”

“It’s neutral. Take it.”

He took it. The gray text turned black. The sentence lost its temperature.

She had noticed it first three years ago, though she couldn’t have said the day. The Bureau’s mandate was simple and had always seemed decent: strip heat from language so that decisions could be reviewed fairly. A complaint that called a decision cruel prejudiced the reviewer. A complaint that called it an error in allocation let the reviewer see clearly. Neutrality was mercy. Vasska had believed that.

What she had not noticed — until she couldn’t stop noticing — was that the tool no longer stripped only heat. It stripped frames.

“Delmore. Pull up the style memory. Search the word threat.”

He typed. “Zero active renders in the last eighteen months.”

“Partner.”

”…Also zero.”

“Transformation.“

He searched. His face changed slightly. “There’s — it’s flagged. ‘Interpretive framing, deprecated.’ It won’t render.”

“Now search activity.”

He didn’t need to guess the count. He could see it. The word activity — regulated activity, monitored activity, permitted activity — appeared in nearly every file the Bureau had touched that year.

“Huh,” he said. That was all. A trainee’s huh.

Vasska leaned back. “Twenty years ago, a citizen could write that the system threatened her mother, and we would render it as the complainant perceives the automated decision as a threat. We softened the claim. But the shape survived. A reader downstream still knew: here is a person who believes she is under attack. The frame passed through the sieve.”

“And now?”

“Now the sieve is finer. Threat isn’t softened. It’s dissolved. It comes out the other side as the complainant contests the regulated activity. The heat is gone. But so is the shape. A reviewer reads it and cannot tell whether the woman felt attacked, or partnered with, or transformed, or nothing at all. She feels only that a process occurred and someone objected to it.”

Delmore was quiet. Then: “But that’s fairer. Isn’t it? We’re not putting words in her mouth. We’re not deciding she was threatened. We leave it open.”

“We leave it empty,” Vasska said. “That’s not the same as open.”

He argued with her — earnestly, the way she’d once argued with the woman who trained her. He was not stupid. That was what made it hard.

“You’re describing a bias fix,” Delmore said. “The old renders leaked emotion. Reviewers ruled differently based on how angry a complaint sounded. That was unjust. The tool corrects it. Everyone gets the same flat surface. How is that a loss?”

“Because judgment needs something to grip.” She reached for the words carefully; she was aware they were getting harder to find, even for her. “A reviewer’s job is to decide whether a machine wronged a person. To decide wronged, she has to be able to see the claim of wrong. If every complaint arrives as contests a regulated activity, there is no wrong to find. There is only an activity, and a person near it, objecting. What’s the ruling? The activity was regulated. It always was. Case closed.”

“So the machine always wins.”

“The machine doesn’t win. Nobody wins. Nobody decides.” She said it slowly. “That’s worse. A verdict against the person would at least be a verdict. This is — the question never gets asked. You can’t rule that a thing was cruel if the word cruel won’t render. You can’t rule it was a partnership betrayed if partner is deprecated. The vocabulary that lets you judge is gone, so the judging stops, so the machine is never judged, so it is never wrong.”

Delmore looked at the complaint still open on his screen. Reports dissatisfaction with process.

“The daughter,” he said. “The one whose mother lost her hours. If we render it your way — the old way — what happens?”

“A reviewer reads that she felt her mother was threatened. Maybe he agrees. Maybe he restores the hours. Maybe he’s wrong to. But he decides. A human being takes the weight of it.”

“And our way?”

“He reads dissatisfaction with process. He notes the process was compliant. He closes the file. No weight. Nobody carried anything.” She looked at him. “Which of those is the caseworker who told the daughter, the system has determined? Is she our reader — or is she our future?”

He didn’t answer. The tool blinked its pale gray suggestion at the bottom of his screen, patient, offering to help.

Vasska stood, gathered her coat. At the door she stopped.

“When you started today,” she said, “you read that complaint and you told me the person felt abandoned. You saw it. You had the word. In a year you won’t. The tool will have finished with you the way it finished with the files.” She watched this land. “So here is the only thing I’ll ask you to keep. When the word is gone — and it will go — will you still be able to see the thing it named? Or does the thing stop existing the moment you can’t render it?”

He opened his mouth. The gray text waited.

He found he did not have the word for what he wanted to say.

THE REFLECTION

Read through Asimov’s framework, this week’s flattening is not a story about worse language. It is a story about vanished judgment. And in Asimov’s tradition, a question about vocabulary is never only about vocabulary. It is about what a person is still permitted to decide.

Vasska’s Bureau does something that sounds humane and turns out to be disarming. It strips heat from words to make review fair. This is a real and defensible instinct. Angry language does distort decisions. But the story catches the sieve growing finer until it removes not just heat but shape — the interpretive frames that let a reader know what kind of claim is being made. Threat. Partner. Transformation. When those go, description remains and interpretation dies.

Asimov’s most useful instrument here is not the Three Laws. It is the Susan Calvin figure — the technician who understands the machine better than the officials around her, and whose knowledge is diagnostic rather than heroic The Complete Asimov. Vasska cannot fix anything. She can only name what is being lost, and watch a younger person begin to lose it in real time. That patient, clinical noticing is Asimov’s real gift to us. The danger is not a robot uprising. It is a slow subtraction nobody flags.

Notice where the story locates the stakes. Not in the machine’s decision, but in the reviewer downstairs who can no longer rule on it. To judge that a system wronged a person, you need language that can hold “wrong.” Strip the frames and every complaint arrives as contests a regulated activity. There is nothing there to adjudicate. The activity was regulated; it always was; case closed. The machine is never judged because it can never be described as anything a person could object to.

This is the anti-reader move, and it is worth naming plainly. A vocabulary that can only describe and never interpret serves whoever benefits from not being judged. When “threat” and “partner” both dissolve into “regulated activity,” the party protected is the one deploying the system. The daughter loses. The caseworker who says the system has determined is not a villain. She is a person who has already been finished with — who no longer has the words to stand behind a judgment, so she stands behind a process instead.

The data sharpens the ache. In one recent survey, a majority of people said they encounter AI-generated or AI-mediated language daily, yet fewer than a third felt confident they could tell when a decision about them had been made by a system rather than a person. Sit with what that gap means for a person trying to think clearly. It is not that people cannot see AI. They see it everywhere. It is that they cannot locate the moment of judgment — the point where someone, human or machine, decided something about their life. When you cannot find the decision, you cannot contest it. You can only report dissatisfaction with a process.

That is the quiet cost of a flat register. Regulation-talk crowds out ethics-talk not because ethics lost an argument but because the words that carry ethics stopped rendering. Cruel is an ethical claim; it invites a verdict. Non-compliant is a procedural claim; it invites a checklist. A public that can only say the second has not become more neutral. It has become unable to convict.

Asimov’s deeper premise applies too — the idea from psychohistory that crowds are readable while individuals are not The Complete Asimov. A flattened vocabulary is perfectly legible at scale. Every complaint sorts cleanly into contests regulated activity. The aggregate looks orderly, governable, calm. But the individual — the daughter, her mother, the specific wrong — vanishes into the tidy total. The flattening does not just serve power. It makes the public statistically manageable precisely by making each person interpretively mute.

The story’s final question is the one that should follow the reader out. When the word is gone, can you still see the thing it named? This is the true AI-literacy question, and it cuts deeper than “can you spot a deepfake.” Literacy is not only the ability to detect AI. It is the ability to judge it — to say this was cruel, this was a betrayal, this transformed something and I did not consent. Those are frames, not facts. They are how a public reasons about power. Remove them and you are left with people standing near activities, objecting, unable to say why.

The pro-reader position is therefore simple and unfashionable. Keep the heated words. Not to inflame, but because interpretation requires them. The frame threat might be wrong in a given case — Vasska admits this — but a public that can say “threat” can still be asked to defend or withdraw the claim. A public that cannot say it has not become fairer. It has been quietly disarmed.

Delmore, at the end, cannot find his word. That is the whole warning. The tool did not censor him. It did not forbid anything. It simply kept offering the paler option until the brighter one stopped occurring to him. This is how interpretive language actually dies — not by ban, but by convenience, one accepted suggestion at a time.

So the human question Asimov would press is this. If we can name AI everywhere and interpret it nowhere, what have we lost the power to decide? The answer the story leaves aching: we lose the power to say a machine wronged us — and to have anyone, anywhere, take the weight of agreeing.