What AI Writing Assistants Do to Your Text

The safest AI writing jobs are the ones where a person supplies all the material and the model only rearranges it. The riskiest are the ones where the model supplies material. When that line stays clear, AI behaves like a handy, slightly overeager copy desk. When it blurs, someone spends an evening tracing claims they never made back to sources that don't exist.

This matters well beyond newsrooms. Public comments, testimony, flyers, candidate statements, letters to a board and social posts all get run through AI tools now. The failures are the same in every one.

Where it helps, where it hurts

Outlines: a test, never a source

Given an outline and told not to add facts, a model is decent at finding holes. Here is the kind of request that keeps it in that lane, using the invented Halden County budget from lesson 3:

Here is the point sentence and outline for a three-minute public
comment at the county budget hearing.
Point: Halden County's proposed jail budget rises $655,000, more
than half of the county's entire $1.27 million increase, driven by
a new medical contract.
Outline:
1. The jail increase and its share
2. The medical contract: what it costs and why
3. Sheriff's explanation
4. Commissioner who questioned it
5. What happens next: Sept. 15 hearing

Which section doesn't serve the point? What would a listener ask
that the outline doesn't answer? Don't add facts; just ask questions.

A typical useful answer: "A listener will want to know what this means for property taxes" and "Section 4 should say how the commissioner voted, if there was a vote." Both are questions only the records or the clerk can answer. The AI found gaps. It did not fill them, which is exactly what it was asked to do. Ask it to fill them and it will, with guesses.

Tightening, and how claim creep gets in

Tighten this paragraph to under 45 words. Rules: keep every fact
and number exactly. Add no facts, context or characterization.
Don't change who said what. Return the edit, then list any word
you changed that alters meaning.

A smoother sentence can hide a claim that doesn't add up. Editing tools make claims easier to read. Only a check against the numbers makes them true.

Here is what the same paragraph looks like when the request is just "make this better":

Facing mounting pressure over rising costs, Sheriff Dana Kroll defended the jail budget Tuesday, pointing to a new inmate medical contract that has nearly doubled what the county pays for care.

"Facing mounting pressure" is invented. "Defended" frames the sheriff as under attack. "Nearly doubled" is roughly right (412 divided by 198 is 2.08, so it more than doubled), which means even the true-ish part is wrong. Three problems in one sentence, all of them confident. An advocate who used that version at the hearing would hand the sheriff an easy rebuttal.

Titles and summary lines

Asked for title options, AI produces volume. It also produces drama. Its favorite verbs are "slams," "blasts" and "sparks," and it loves a question opener ("Is your water bill about to explode?"). For a water rate hike that failed on a 2-to-3 council vote, the accurate version is dull and correct: "Water rate hike fails on 3-2 council vote." The dramatic versions ("Council kills rate hike in stunning rebuke," "Residents win war over water rates") add a conflict frame the vote doesn't support.

Three tests catch most bad titles, from AI or anyone. Is every word supported by the facts? Would the person named call it fair? Does it promise something the document delivers?

Explaining terms

AI is a patient explainer. Asked what "levy rate" or "material weakness" means, it usually gives a clear paragraph. That paragraph is good for understanding the term. It is a poor thing to quote, because a model's definition can be right in general and wrong for your state. Property tax rules, for one, differ a lot from place to place, and the model may describe another state's system with total confidence. The statute, the auditor's standards or the agency's glossary is the citable version.

The claim sweep

One more mechanical job AI does well: asked to "list every factual claim in this document as a numbered list, one claim per line, including numbers, dates, titles and attributions, without evaluating them," it produces a checklist. Each line can then be matched to the evidence table. Any claim without a row either has a source waiting to be found or shouldn't be repeated. It catches the sentence someone added at 11 p.m. from memory.