Do not fear! The government isn't going to catch you with a watermark
Someone asked me last week whether federal agencies will start filtering grant applications by detecting AI tags in the text. It is a timely question, and the answer matters because agencies ARE filtering. Just not how you may think.
On August 11, Anthropic began marking everything Claude produces. You can bet others will follow. The EU AI Act now requires it.
Two things about that watermark matter to anyone submitting a federal application. No public detector with a published algorithm exists, so an agency cannot run the check. Any metatag or mark only indicates an AI model acted on the text at some point. The standards are so loose it could be a spell check in Grammarly or Copilot.
Forget about code-based filters; the actual filtering is already in place, and it is people holding the magnifying glass, not machines.
NIH
NIH will not treat applications substantially developed by AI as original ideas of the applicant. Post-award discovery can mean referral to the Office of Research Integrity, cost disallowances, suspension, or termination.
NIH also capped each principal investigator at six applications per calendar year, after catching AI-enabled PIs filing more than forty in a single round.
National Science Foundation
NSF requires disclosure of the extent and form of AI use. Leave it out, and it is treated as misrepresentation.
DOE and DoD
Both have yet to publish standards, but both indicate disclosures will be the rule when they do.
Notice what none of that requires: programmed detection.
A watermark says a machine touched your words. An application with narratives absent human thought is just meaningless prose. Only one of those costs you the award.
Scott Wise builds federal funding applications — SBIR, STTR, and NOFOs — for small businesses with a focus on early-stage technologies. swise@wiseexecution.us · wiseexecution.us