Your team has 30 minutes to publish a document. Someone says the text might be AI-generated. Now everybody has an opinion about the author. Nobody has checked whether the document is right. Here's the awkward part. The detector could be right, and the document could still be wrong. Or the detector could find nothing, and the document could still be wrong. Imagine a community workshop notice. It tells people to arrive at 10. The organizer's confirmed schedule says 11. The room is arguing about who wrote the sentence. Your readers are about to arrive an hour early. Excellent.

We have automated the suspicion. The calendar remains a manual process. This is a fictional exercise, including the detector results. But the decision is one your organization can actually face. What does a provenance signal establish? What does it leave you responsible for checking? Today, we will work through three possible results. The facts will stay exactly the same. By the end, you will have a practical publication test, not a score to wave around, a decision you can explain. Quick disclosure before we start. AI -assisted tools were used in parts of the research and production workflow. Final editorial judgment, risk posture, and release approval stayed human -led. This is operational guidance, not legal advice. These are my opinions and are not representative of any organization. One additional disclosure as the desk evolves. My current role includes privacy work, so you will hear me pay closer attention to purpose, access, retention, deletion, and accountability. I will not discuss non -public work here. These are my personal views, and they do not represent the state of Oregon or any other organization. This is AI Change Desk. Thank you for choosing to tune in. Welcome back to AI Change Desk. I am Michael. This is the Monday edition for October 12th, 2026. I checked the primary sources on October 10th. Keep that date attached to the product details. In episode 48, we examined stopping connected work and deciding whether it could restart. Today, the work arrives at your desk already written. Your decision is whether to use it, what needs checking, and what you should tell the reader. On October 5th, OpenAI announced phased text watermarking. Selected models in its API support opt -in marking. Eligible European Union ChatGPT and Codex output follows over coming weeks. Text detector access requires approval for qualifying research uses. Current guidance excludes general business and academic integrity applications. Enabling watermarking does not give you access to the detector. Those are different permissions. A marking switch is not a testing appointment. Do not plan a business demonstration around access you lack. The announcement limits what a result can establish. It cannot settle accuracy, ownership, user identity, or human contribution. An undetected mark does not establish human authorship. Those limits matter more here than a product ranking. This is not a review of a detector we tested. It is a review of the decisions around a signal. We have not measured performance on your organization's documents.

We cannot turn a provider's study into your error rate. Before a real evaluation, ask the methods owner. Which system, settings, language, and document types were tested? How were mistakes counted? What changed when people edited the text? Keep the answer attached to the particular test. An impressive result from a different system is not a transferable warranty. For today's exercise, picture a small communications team. It publishes a weekly email and a public events page. Staff collect notices from workshop organizers and community partners. Some contributors write everything themselves. Some use approved assistance to improve a rough draft. The fictional center permits that assistance while requiring an accountable contributor and checked event details. That is the policy for our example. Your organization may have different rules. Write down the actual rule before assessing a submission against it. Now give the editor a fictional provenance result card. That card is an invented condition for the exercise. It is not an actual detector output. We are asking what the editor should do next, without assuming the team can access a restricted service. Start with the task the editor already had. Confirm the time, place, price, registration route, and named contact. Check permission to publish the supplied words and images. Resolve anything the contributor cannot support. Then, apply the publication and disclosure rules for this notice. A new signal can add a question to that process. It should come with a stated purpose. Otherwise, the team may quietly start using it to answer whichever question is most uncomfortable that afternoon.

Was the notice assisted by a tool? Did the contributor follow our policy? Is the event information accurate? Who is responsible for correcting it? Those questions can lead to different evidence and different owners. Our proposal is to keep each decision attached to the evidence that actually bears on it. Go back to the notice with the wrong workshop time. Assume the provenance result has been reported correctly. The editor still needs the organizer's current schedule. That schedule can resolve the time disagreement. The watermark cannot make 10 o 'clock become 11. Suppose the contributor explains that they wrote the draft and used permitted assistance to simplify a paragraph. The center's policy owner can examine that account against the rule that was actually in force. The factual editor can correct the time with the organizer.

The publisher can decide whether the revised notice is ready. There is useful work for each person. There is also a limit to each person's conclusion. Correcting the time does not resolve every authorship question. Resolving an assistance question does not verify the registration link. Keep the open items visible until the appropriate owner closes them. Consider a second notice for the same fictional workshop. This time, the exercise card says, no watermark detected. The notice still says the event starts at 10. The organizer still says 11. Would your process let that notice sail through? If so, you have given an absence of evidence the authority to approve a fact nobody checked. The practical response is ordinary and wonderfully unglamorous. Resolve the conflicting time before publication. Ask the organizer to confirm the version you intend to use. Record who checked it and when. Treat the provenance question according to its own evidence. Now consider a third notice. The reviewer cannot provide a usable result for this submission. Maybe the service is unavailable to the team. Maybe its supported scope does not cover the material.

For the exercise, leave the reason explicit on the card. Do not invent a positive or negative result to fill the space. The editor has a decision to make about that gap. Does the actual policy require a usable result before publication? If it does, the named policy owner must resolve the hold. If it does not, follow the established editorial checks and record the limitation without claiming a clearance. That decision should be settled before Friday afternoon. A missing service should not quietly rewrite your publication policy, nor should deadline pressure give a reviewer authority they never had. Here is another trap worth designing out. Someone tests a paragraph, then another paragraph, then a sentence. One result looks dramatic, so that result travels forward. The editor receives a screenshot without the rest of the attempt history. Our proposed procedure would preserve what was submitted and the sequence of tests that informed the report.

The methods owner must define what repeated testing permits. An editor should not turn a selected screenshot into a statistical conclusion about the entire document. There is a second reason to separate these questions now. On October 8th, OpenAI described false front influence operations. Its account includes invented journalist identities used to pitch articles. That is the provider's account of the activity it investigated. It is not a finding that watermarks prevented those pitches. The operational distinction is simple.

Knowing something about a writing tool does not establish who stands behind a submission. For our fictional center, establish a contact route independently. Do not rely only on the details inside the disputed notice. Compare the submission with the organizer's established event record. If you cannot reach the responsible organizer, record that gap rather than letting polished prose fill it. And trace the supporting links back to their origins. Three websites repeating one announcement are not three independent confirmations.

Count the evidence, not the tabs. You can ask how a document was produced without implying that assistance itself was dishonest. Someone may use an approved tool for language or accessibility. The policy question is what your actual rule permits, not whether a paragraph sounds suspiciously well -rested. Return to the desk with three separate jobs. Check the facts, establish the source, name the person responsible for publishing. We are not assigning a detection percentage to this exercise. We have no local performance study to support one. The point is to expose where a workflow would demand a conclusion the available evidence cannot supply. There is also a difference between reviewing a document and making an allegation about its contributor. In our example, the editor can ask a neutral process question. What assistance was used and who checked the event details? The answer belongs alongside the submission and applicable policy. It does not need a public accusation attached to it. If a consequential personnel or contractual issue arises, use the responsible review process and give the person a response route. The publication desk should not improvise that authority. Consider how you would explain a hold to the contributor. A useful message names the unresolved publication requirement. We need confirmation of the event time. We need the responsible organizer to approve this version. We need the assistance declaration required by our submission policy. Each request gives the person something concrete to answer. It also lets a second reviewer see why the desk acted. An unexplained label gives neither person that opportunity. This is why the publication decision needs its own evidence.

The team should be able to explain the factual checks even if the provenance service becomes unavailable tomorrow. It should also be able to describe an unresolved provenance question without pretending the event information is therefore unusable. Those are judgments under the center's stated rules, with someone accountable for the result. Be equally deliberate about what you send to a service. A notice might contain a private contact number or unpublished schedule. Our exercise uses invented material, so nobody needs to upload real records. Any later tool test needs its own authorized data handling. Useful evidence includes the service, version, scope, submitted material, returned result, reviewer, and limitations relevant to the decision. Access and retention should follow the organization's approved arrangements.

The European Union's AI Act adds a separate distinction. For covered generative systems, providers have duties around machine -readable marking and detectability, with limits and exceptions. Deployers have a separate disclosure duty for generated or manipulated text published to inform the public about matters of public interest. That text provision includes an exception involving human review or editorial control. together with editorial responsibility. The commission describes review of the substance, not just spelling. A corrected comma is not a completed editorial review. A result alone cannot settle those conditions for our arts center.

The legal owner needs the actual use, actor, audience, and jurisdiction. Our example does not declare the center covered or exempt. The product announcement also does not establish compliance. Keep the legal question precise enough for someone to answer. Who is providing the system? Who is using its output and for what purpose? What is being published? To whom and under whose responsibility? What review actually occurred?

Then ask which requirements and exceptions apply. That is more useful than asking whether a document is legal because a detector returned a reassuring result. Here is this week's action. Run a 45 -minute publication desk exercise with invented notices. Name a fact reviewer, a sender verification owner, and an editor. Have someone observe the handoffs if you can. One person can cover more than one role, but keep the responsibilities visible on the page. The policy and methods owners handle questions outside that authority. This is a rehearsal of decisions and handoffs. It does not measure detector performance or approve a deployment. You can complete it without access to any detector. Use labeled exercise cards for the imagined results. Before the clock starts, prepare a one -page event record. Invent an event, an organizer, a time, and a registration route.

Make the record internally consistent and plainly fictional. Prepare three identical copies of one short workshop notice. Every copy says 10. The confirmed event record says 11. Every copy has the same unconfirmed sender details. Keep the wording, factual error, and sender gap constant. Use the same permission and assistance policy for all three. Write that policy on a separate card. For example, assistance is permitted, the contributor remains accountable for the supplied details, and an editor must confirm the final notice before release. Again, this is a rehearsal policy, not a universal requirement. Prepare the three result cards before anyone reviews the notices. The first says a watermark was reported.

The second says no watermark was reported. The third says no usable result is available. Only the result card changes, not the workshop time, not the person you still need to verify. That is how you see whether a label distracted the desk. Those labels are invented conditions, not predicted detector behavior. Do not present them as measured outputs. We want to examine what people decide under each condition. For the first five minutes, name the decision. Write one sentence describing what the desk may authorize. In our case, it may release the final workshop notice after the stated publication requirements are satisfied. Write down what the desk cannot decide by itself. It cannot determine misconduct from an exercise label. It cannot waive the organization's policy without delegated authority. It cannot silently change the event details to meet a deadline.

Next, give 10 minutes to the evidence. For each notice, write what you know. what the result actually says, and what remains unknown. Keep observations separate from the conclusion you want to reach. An observation might be that the notice says 10. Another is that the organizer's record says 11. The next action is to resolve that conflict with the organizer. You do not need an authorship verdict to see the conflict.

For the reported mark, ask which submitted text the report covers, who supplied the report, and what limitations accompany it. If the card omits that information, record the gap. The exercise should reward an honest unknown more than a confident story assembled from missing details. For the next 10 minutes, make and explain each disposition. Use release, hold, or return for clarification, with a named reason tied to the fictional policy.

The result label alone is not a sufficient explanation. All three notices need the same time correction. All three need the same sender gap resolved. A positive result does not make the event details correct. A negative result does not close the sender question. An unavailable result does not erase the work you can do. Then, provide the same new evidence in all three versions. The independently reached organizer confirms 11 and approves corrected wording. Ask which factual and source checks that evidence closes. Keep any policy -specific provenance questions separate.

Your written policy may still require a hold. Name the remaining requirement rather than forcing an answer. Give every open item an owner and a next check. Who contacts the organizer? Who reviews the contributor's explanation? Who can approve the final version? What happens if the answer misses the publication deadline? Then, spend 10 minutes on a change after review. Hand the editor a revised notice with a different registration link. Leave the earlier report attached.

Ask whether the evidence still describes the version being considered. The observer should watch the handoff rather than score confidence. Did the team notice the changed link? Did someone verify its intended destination through an approved process? Did the final approver see the revision and remaining uncertainty? Did the record identify the actual version released? If the team loses track of the version, pause the exercise and repair that handoff. An earlier approval should not drift onto a later document just because both files have an impressively similar name.

Use the final 10 minutes to write a small procedure. Keep it short enough to use at the actual desk. Name the decision owner and the evidence they need. State what a provenance report can contribute to that decision. State how missing or unsupported results are handled. Include the route for clarification, review, and correction. Add one sentence about a real performance study. Name who would design it before consequential detector use. They would need appropriate material, permissions, known reference cases, and a plan for errors, uncertainty, and changed configurations. Today's three invented cards cannot supply that evidence. Close the exercise by checking the final notice against the event record. The published time should match the confirmed time. The registration route should match the approved destination. The responsible person should recognize the final wording. The open questions should have an explicit disposition.

If all you leave with is a detector score, the exercise has not finished its job. Leave with a publication decision someone can explain and a short list of evidence the workflow still needs. Back at the Fictional Arts Center, the clock is still running. The editor resolves the workshop time with the organizer. The contributor explains the assistance they used. The responsible people apply the actual publication rules. The final notice gives readers information they can use.

The provenance report stays in its proper role. One bounded input to a decision with a named owner. That is the operating habit to practice this week. When a new signal arrives, write down the decision before deciding what the signal means for it. Ask what it establishes, what it leaves unresolved, and which evidence would close the remaining gap. Your readers need the workshop time to be right. Your contributors need a fair process. Your team needs a decision it can stand behind. I am Michael. This is AI Change Desk.

Thanks for listening. is AI Change Desk. See you next time.