CCM working papers
240 invented disclosure statements, coded against six questions.

Saying so

How 60 invented local newsrooms disclosed their use of AI, and where on the page they said it.

Invented contentScroll to read ↓

Newsrooms are writing AI disclosure policies faster than they are writing disclosures. This showcase study codes 240 invented published items from 60 invented local newsrooms against six questions: what the tool did, who checked it, where the notice sits, how specific it is, whether it names the model, and whether it is repeated. In the invented sample, 61 percent of items disclosed the tool and 22 percent named the human who checked the work. Notices sat at the foot of the page in 66 percent of cases.

Part 1

The question

Disclosure is a sentence on a page. This study asks what that sentence says.

Background

Why the wording of a notice is the whole policy

Most local newsrooms now hold a written AI policy. Fewer publish a disclosure statement on the items the policy covers. The gap between the two is where a reader lives, because a reader never sees the policy.

A disclosure statement does one job: it tells a reader what a machine did to the item in front of them, and who checked it. Everything else in the statement is packaging. This study codes the packaging as well, because in the invented sample the packaging changed how often the statement appeared at all.1

Two ideas from accessibility practice carry over. The first is materiality, which asks whether the machine's part changed what the reader gets. The second is proportionality, which asks how much detail the notice needs. A spelling pass is not material. A generated summary at the top of a story is.2

The field already agrees that disclosure matters. It does not agree on what a disclosure must contain, where it belongs, or how often it must repeat. Those three disagreements are what this study measures, one item at a time.

The invented sample holds 240 items from 60 newsrooms, 4 items from each. It covers the 6 months to June 2026. No such sample exists, and every result below is invented for a design showcase.

Notes

  1. Invented result. In this sample, newsrooms with a fixed notice template disclosed on 71 percent of items, and newsrooms without one disclosed on 44 percent, over the 6 months to June 2026. Back to the text
  2. The alt text comparison is a real argument in the field. The numbers in this paper are not. Back to the text

Measured in percent of items.

Disclosure was most common for a generated summary at the top of an item, at 84 percent, and least common for copy editing, at 6 percent, across 240 invented items in the 6 months to June 2026.

Invented data for this showcasecenterforcooperativemedia.org

Invented values. Each item counts once for each task it used, so the shares do not sum to 100.

Show the data
CategoryValue
Generated summary at the top84%
Translation73%
Transcription41%
Headline options29%
Photo selection17%
Copy editing6%

Part 2

Method

Six questions, two coders, one agreement test.

How the coding ran

The six questions each item was coded against

Two coders read each item and answered six yes-or-no questions. A third coder settled disagreements. The questions are listed in the table below with the share of items that answered yes.

Sampling

The invented frame held 214 local newsrooms in 8 states. Sixty were drawn at random. From each, the 4 most recent items that carried any AI notice were taken, which biases the sample toward newsrooms that disclose at all.1

That bias is deliberate and it limits the claim. This study cannot say how many newsrooms disclose. It can say what a disclosure looks like when a newsroom writes one.

Agreement

The two coders agreed on 88 percent of the 1,440 coding decisions, which is 240 items multiplied by six questions. Agreement was lowest on the question about human review, at 74 percent, because a notice that says 'edited by our staff' does not name a person.

Notes

  1. Invented frame and invented draw. The 8 states are not named, because naming a real state as the source of invented data would mislead the reader. Back to the text
Each row is one question. The share column counts the 240 invented items in the 6 months to June 2026 that answered yes.
Does the notice say what the tool did?14661
Does the notice say who checked the work?5322
Does the notice sit near the affected text?5222
Is the notice specific to this item?8937
Does the notice name the model or the product?3414
Does the notice appear on every item it covers?7130

Invented data for this showcasecenterforcooperativemedia.org

Invented values. A row counts an item once, even when the item carried two notices.

Measured in percent of that task’s items.

In every one of the 4 tasks the notice sat at the foot of the page more often than anywhere else, and a notice beside the affected text was common only for a generated summary. Each row is 100 percent of that task's invented items.

Invented data for this showcasecenterforcooperativemedia.org

Invented values. An item counts once, under the highest notice on the page. The data table holds the count as well as the share.

Show the data
CategoryFoot of the pageFoot of the page shareBeside the affected textBeside the affected text shareLinked policy onlyLinked policy only shareTotal
Generated summary38%61.3%21%33.9%3%4.8%62%
Translation44%74.6%9%15.3%6%10.2%59%
Transcription31%72.1%4%9.3%8%18.6%43%
Headline options22%62.9%1%2.9%12%34.3%35%

Part 3

Results

Newsrooms describe the tool. They rarely describe the check.

What the coding found

Newsrooms name the tool more often than the person

The gap between the first two rows of the table is the main result. In the invented sample, 61 percent of items said what the tool did and 22 percent said who checked the work. A reader learns that a machine wrote the summary. The reader does not learn that anybody read it.

Placement follows the same pattern. A notice at the foot of the page is a notice the reader reaches after the decision to trust the item is made. In the invented sample, 66 percent of notices sat there.

Specificity moved with newsroom size. Newsrooms with 5 or fewer staff wrote item-specific notices on 51 percent of items. Newsrooms with more than 20 staff wrote them on 24 percent. The larger newsrooms had a standard sentence and used it everywhere.1

That is not a failure of the larger newsrooms. A standard sentence is how a policy survives contact with a daily production schedule. It is a failure of the sentence, which was written once and never written again.

Change over the six months

Disclosure rose over the period in every task group in the invented data. The rise was steepest for translation, from 44 percent in January to 73 percent in June 2026. The small multiples below show each task on the same scale, so the reader can compare the slopes and not only the levels.

One result runs against the pattern. Newsrooms that publish in two languages named the human check on 44 percent of items, against 22 percent for the sample as a whole. Translation is the task where an error is easiest for a reader to find, and the notice appears to follow the risk.

Notes

  1. Invented result. Staff size is self-reported in the invented frame and covers paid staff of any role, counted in January 2026. Back to the text

Measured in percent of items.

Translation rose 29 percentage points over the 6 months to June 2026, and copy editing rose 2, on the same scale in every panel.

Invented data for this showcasecenterforcooperativemedia.org

Invented values. Each panel holds the share of that month's items that carried a notice for that task.

Show the data
PointGenerated summaryTranslationTranscriptionCopy editing
Jan71%44%33%4%
Feb74%49%34%4%
Mar78%58%36%5%
Apr80%64%38%5%
May82%69%40%6%
Jun84%73%41%6%
What follows

What a newsroom can change this week

Two changes cost nothing and move four of the six coding questions.

  1. Name the person or the desk that checked the work, not the newsroom.
  2. Put the notice where the machine's work sits, not at the foot of the page.

A notice that reads 'this summary was drafted by a language model and checked by the news editor' answers four questions in one sentence. It says what the tool did, who checked it, that the notice is specific to this item, and, by its position, which text it covers.

The third change costs more. A notice must be repeated on every item it covers, and that means the notice belongs in the publishing template, not in a writer's habit. In the invented sample, 30 percent of notices appeared on every item they covered.

Start with the item type your newsroom publishes most. In the invented sample, a newsroom that fixed the notice on its single most common item type moved its own disclosure share by 19 percentage points over the following 3 months, without touching any other item type.

Appendix A. The coding sheet

Each item received one row. The row held the newsroom, the date, the task, the six yes-or-no answers, the notice text as published, and the coder's initials.

Coders were told to read only what a reader can see without leaving the item page. A policy reachable through a link in the footer counted as 'linked policy only' and never as a notice on the item.

Appendix B. Agreement between coders

Two coders read all 240 invented items. Agreement across the 1,440 decisions was 88 percent. The lowest was 74 percent on human review and the highest was 96 percent on whether the notice named a model.

A third coder settled the 173 disagreements. The settled value is the one used in every figure and table on this page.

Appendix C. The sampling frame

The invented frame held 214 local newsrooms in 8 states, built from three membership lists and one state directory. Sixty were drawn at random with equal probability.

A newsroom entered the frame if it published at least weekly and covered a defined geographic area. Wire-only sites and single-topic national sites were excluded.

Words used in this paper

disclosure statement
A sentence published with an item that tells the reader what an AI tool did to that item and who checked the result.
human review
A named person or desk that read the machine's output before publication and takes responsibility for it.
item-specific notice
A notice written for one published item, rather than one standard sentence used on every item a newsroom publishes.
materiality
Whether the tool's part changed what the reader receives. A material change needs a notice; a spelling pass usually does not.
proportionality
How much detail a notice needs. A bigger change to the item calls for a longer and more specific notice.

disclosure statement

A sentence published with an item that tells the reader what an AI tool did to that item and who checked the result.

Open the glossary entry

human review

A named person or desk that read the machine's output before publication and takes responsibility for it.

Open the glossary entry

item-specific notice

A notice written for one published item, rather than one standard sentence used on every item a newsroom publishes.

Open the glossary entry

materiality

Whether the tool's part changed what the reader receives. A material change needs a notice; a spelling pass usually does not.

Open the glossary entry

proportionality

How much detail a notice needs. A bigger change to the item calls for a longer and more specific notice.

Open the glossary entry

Methods and sources

How we did this

Two coders read 240 invented published items from 60 invented newsrooms and answered six yes-or-no questions about each notice. A third coder settled disagreements. The sample covers the 6 months to June 2026 and is drawn only from newsrooms that disclose.

No such study exists. This page is a design showcase for a report template, and every number, name, quotation and result on it is invented. The coding sheet, the sampling frame and the PDF the cover offers are placeholder addresses, and nothing is published at any of them.

The links below go to real organizations that publish work on AI and journalism. None of them produced the data on this page.

About the authors

Ruth Ellery Kwan

Principal investigator

Invented author. She studies how newsrooms write the rules they publish about their own work.

Nabil Haddadi

Coding lead

Invented author. He built the six-part coding scheme and ran the agreement test.

Colette Mwangi

Research assistant

Invented author. She collected the sample and kept the coding log.

Cite this report

Plain text

Ruth Ellery Kwan, Nabil Haddadi and Colette Mwangi. Saying so. CCM working papers. Center for Cooperative Media, September 2026. https://jamditis.com/ccm-report-template/examples/paper.html

APA

Ruth Ellery Kwan, Nabil Haddadi and Colette Mwangi. (September 2026). Saying so. Center for Cooperative Media. https://jamditis.com/ccm-report-template/examples/paper.html

BibTeX

@techreport{kwan2026saying,
  author = {Ruth Ellery Kwan and Nabil Haddadi and Colette Mwangi},
  title = {Saying so},
  institution = {Center for Cooperative Media},
  series = {CCM working papers},
  year = {2026},
  month = {September},
  url = {https://jamditis.com/ccm-report-template/examples/paper.html}
}