Your Grant Narrative Carries a Mark Now

There are roughly five weeks left before the September 17 deadline, and in tribal offices across the country right now, someone is writing a narrative.

NTIA opened two funding opportunities totaling $790 million on June 17. As summarized by the law firm Akin, $540 million sits in the third round of the Tribal Broadband Connectivity Program and $250 million in the Native Entities Grant Program. Applications close September 17 at 11:59 PM Eastern. NTIA expects to make awards on a rolling basis starting in Spring 2027.

Grant offices are thin, deadlines are real, and generative AI has quietly become part of how a lot of this writing gets done. That was true last month too. What changed this month is that the writing now carries a signature.

What changed on August 2

Anthropic began embedding an imperceptible watermark into text generated by Claude, and signed provenance metadata into certain files it produces. By the company's own account, the text mark is woven into the words, survives copying, and has no documented opt-out. It applies across the API, the apps, and the versions offered through the major cloud providers.

Other major model vendors are on the same path, driven by transparency requirements that took effect in the European Union on the same date. Treat this as the direction of the whole market rather than one company's policy.

Two limits on the mark are worth holding onto, because both cut against panic. Anthropic states that a detected mark "provides a signal that content was processed by Claude, but is not fully conclusive," and separately that the absence of a mark "doesn't mean the content wasn't AI-generated or processed." The signal also degrades under heavy editing and paraphrasing. And the detection tools that would let any third party read these marks are not publicly available yet.

So nobody is scanning your application this cycle. That is the accurate near-term picture, and we would rather say it than manufacture urgency.

The reason to pay attention anyway is that grant records do not disappear. A narrative submitted in September 2026 sits in a federal file, subject to audit, monitoring, and review, for years after the award. The tools to read what is embedded in it are coming. The document is not going anywhere.

The question that actually matters

It is not whether AI helped draft the narrative. It is what your tribe certified when it submitted.

Every federal application carries certifications, and a person signs them. Those certifications speak to the accuracy and completeness of what is represented, and in some programs to who prepared it. We are not going to tell you what your specific NOFO requires, because the terms differ by program and reading them is the work. Pull the certification language for both TBCP and NEGP, read it with your grants staff and your attorney, and answer one question: does anything we signed say something about how this document was produced that is not accurate?

That is a thirty-minute task. It is worth doing before September 17 rather than during a monitoring visit in 2028.

Sovereignty is the deeper issue here

At the U.S. Indigenous Data Sovereignty and Governance Summit this spring, roughly 400 people gathered at Casino Del Sol on Pascua Yaqui homelands to work through what control over data means for tribal nations, from data centers to artificial intelligence. Writing about it in Native News Online, Levi Rickert of the Prairie Band Potawatomi Nation captured the posture well: "Most A.I. models don't know what we know, yet. We are still deciding if we want them to."

Provenance marking is the same question arriving from a different direction. A tribal document now can carry metadata that the tribe did not choose, cannot see, cannot remove, and cannot read. Only the vendor can read it, at least for now. For a government whose authority rests on controlling its own records, that is not a small thing, whatever you conclude about the policy merits.

We do not think the answer is to ban these tools. Grant offices are understaffed and the deadlines are not negotiable, and a blanket prohibition mostly produces staff who use them quietly. The answer is to decide deliberately, as a government, rather than to discover the decision was made by whoever had a deadline and a browser tab open.

What we would put in place before September 17

Name what goes through AI and what does not. A short written rule. Our general view is that boilerplate, formatting, and first-pass summarizing are low-risk, while the tribe's own history, need statement, community data, and any factual assertion about conditions on your land should come from your people. That material is the part reviewers weigh, and it is the part no model knows.

Keep the human record. Interview notes, council direction, prior plans, the community meeting where the need was described. If anyone ever asks how the narrative was built, the answer should be a paper trail rather than a recollection.

Give one person the final read. Not for style. To confirm that every factual claim about your tribe is one your tribe can stand behind and document. This single step catches most of what actually goes wrong with AI-assisted writing, which is not detection. It is confident, fluent, specific claims that are not true.

Decide before you need to. Put AI use on a council or administration agenda while it is a policy question rather than an incident. Governments that write this down while the stakes are low are in a much better position than those drafting a policy in response to a finding.

The application in front of you is a statement your tribe is making to the federal government about its own needs. It should sound like your tribe, because it should have come from your tribe. The mark is only awkward for documents where that is not quite true.

If you are working through how AI should be used in your tribe's operations, that is the kind of question we help governments think through.