Public records are trusted for a simple reason: they are supposed to preserve what an agency knew, did, received, and decided. Generative AI complicates that promise. A fabricated meeting summary, altered public notice, synthetic image, or plausible-looking “leaked” memo can enter an email inbox or records system before anyone has checked whether it is authentic.
AI detection tools such as ZeroGPT can help agencies identify material that deserves closer inspection. They cannot, by themselves, prove that a document is false, prove who created it, or establish what actually happened. The useful government question is not “Can a detector catch every AI document?” It is “Can a detector help route questionable material into a defensible verification process?”
Why a suspicious document is a records problem, not just a technology problem
A government record is more than its words. Its reliability depends on context: who created it, when it was received, which system stored it, whether it was altered, and how it relates to an official transaction.
The National Archives and Records Administration describes electronic records management through six lifecycle functions: capture, maintenance and use, disposal, transfer, metadata, and reporting. That structure matters here because a detector operates mainly at one point in the lifecycle. It may flag a file, but it does not preserve the original, manage retention, or document the decision that followed.
NARA’s 2026 guidance on AI materials makes the same practical point from another direction: prompts, outputs, audit trails, and other AI-related materials may become federal records when they are used for official business, relied on in decision-making, circulated, or integrated into agency systems. An agency should therefore decide what must be retained before an incident occurs, not after a disputed document has gone viral.
Use detection as triage, not as a verdict
A detector typically looks for statistical and linguistic patterns associated with machine-generated text. ZeroGPT’s public tool describes sentence-level highlighting, an overall AI percentage, file uploads, batch processing, and generated reports. Those features can be useful for screening a large intake queue, especially when staff need a fast way to prioritize review.
But a percentage is not a fact. It is a model output whose meaning depends on the document’s language, length, subject matter, editing history, and the detector’s performance on similar material. The AI detector free option should therefore be treated as an initial signal, not an authenticity certificate.
This distinction protects both the public and the agency. A genuine staff report written in formulaic administrative language may look machine-generated. A carefully edited AI draft may evade detection. The 2025 NAACL evaluation of AI text detectors found that performance could deteriorate sharply in unfamiliar settings and under modest evasion strategies. That is a poor foundation for automatically rejecting a public submission, accusing an employee of misconduct, or removing a record from an archive.
A defensible government workflow
1. Preserve the submission before analyzing it
Store the original file, message headers where available, upload time, sender information, file hash, and relevant access logs. Create a working copy for scanning. If the document is edited, preserve each material version rather than replacing the first upload with a cleaned copy.
This is basic chain-of-custody practice, but it is often the difference between “the document was suspicious” and “the agency can show exactly what it received.” It also prevents the detector itself from becoming the only surviving record of an incident.
2. Run multiple, low-consequence checks
A first pass can combine an AI-text detector with plagiarism or duplicate checking, metadata inspection, virus scanning, and comparison against known agency templates. For images, audio, and video, use media-forensics or provenance checks rather than assuming a text detector applies.
Keep the result private while it is being evaluated. Do not label a person or organization publicly as deceptive because one automated tool produced a high score.
3. Verify the substance against primary records
Ask what the document claims and what evidence would exist if the claim were true. A supposed council decision can be checked against the agenda, minutes, recording, vote record, and publication log. A purported inspection notice can be compared with the case-management system and the issuing officer’s authenticated copy.
This is where domain knowledge beats detector confidence. The strongest response to misinformation is often not “the prose looks synthetic,” but “the cited meeting did not occur, the resolution number belongs to another year, and the signed source file has a different timestamp.”
4. Escalate high-impact cases to a human review panel
Set a threshold for review, not a threshold for punishment. The panel might include a records officer, subject-matter expert, information-security representative, and legal or privacy adviser when appropriate. It should record the evidence considered, the detector version or service used, the conclusion, and the reason for any correction, restriction, or referral.
For politically sensitive or personally consequential records, add an appeal route. A resident should be able to explain that a document was translated, grammar-corrected, or drafted with assistance without being treated as a bad actor solely because a detector flagged it.
5. Publish the correction trail
When a false or manipulated record has reached the public, silently replacing it creates a second trust problem. Keep the original according to the applicable retention schedule, mark its status, publish a clear correction where appropriate, and explain which evidence established the correction.
The correction should distinguish among three different findings:
- The document is confirmed authentic.
- The document is authentic but contains factual errors or misleading context.
- The document’s origin or integrity cannot be established.
“Unverified” is sometimes the most accurate label. Government credibility suffers when uncertainty is converted into a confident accusation.
Provenance closes a gap detection cannot
Detection asks, “Does this content resemble machine-generated material?” Provenance asks, “What is the documented history of this asset?” Those are different questions.
The C2PA Content Credentials explainer describes a cryptographically bound record of an asset’s origin, edits, and use of AI. A signed provenance record can make unauthorized alteration easier to detect and can show how a file moved through approved systems. It does not prove that the underlying statement is true, and it does not solve a context problem such as a genuine photograph being presented as evidence of the wrong event.
Agencies can apply the idea without waiting for every public document to support a new standard. At minimum, an authoritative publication system should record the source system, responsible office, creation and approval times, version history, and a tamper-evident hash. For sensitive media, cryptographically signed manifests or equivalent controls can add stronger evidence.
NIST’s report on synthetic-content transparency makes the limitation explicit: provenance, labeling, and detection can improve trustworthiness, but they do not guarantee it. That is why a detector score should sit beside access logs, source documents, signatures, and corroborating evidence rather than replacing them.
The procurement questions agencies should ask
Buying an AI detector is easier than governing one. Before sending public or employee content to a vendor, an agency should ask:
- What data is retained, for how long, and in which jurisdiction?
- Is submitted text used to train or improve the service?
- Can the agency export reports, model identifiers, timestamps, and audit logs?
- How does the vendor measure false positives and false negatives on government-style writing?
- Does the tool support the languages and document types the agency actually receives?
- What happens when a document is translated, edited, quoted, or partly generated?
- Can the agency test the system against a local benchmark of known human and synthetic documents?
That last question is especially important. A detector’s marketing accuracy claim is not the same as performance on procurement notices, zoning filings, public comments, police reports, or multilingual benefits applications. Agencies should build a representative test set, restrict access to it, and re-evaluate the tool after major model or service changes.
What governments should not do
Do not make an AI score the sole basis for denying a records request, disciplining an employee, rejecting a public comment, or declaring a document fraudulent. Do not publish raw scores without explaining their uncertainty. Do not upload confidential case files to a consumer service merely because it is free. And do not confuse “AI-generated” with “false”: a machine-assisted summary can be accurate, while a human-authored document can contain misinformation.
My view is straightforward: governments should invest first in authenticated source systems, retention rules, and trained reviewers. Detection tools are valuable after those foundations exist. Without them, the detector becomes an impressive-looking alarm attached to a records process that still cannot explain what happened.
Five questions records officers will hear
Can an AI detector prove that a document is fake?
No. It can indicate that a document resembles AI-generated text or media, but authenticity requires corroboration, provenance, and review of the underlying transaction.
Should agencies scan every public submission?
Not automatically. Agencies should conduct a privacy and proportionality assessment, define a legitimate purpose, and focus scanning on risk-based workflows where the result will lead to meaningful human review.
What should be retained when a detector flags a file?
Retain the original file, relevant metadata, the detector output, the tool and configuration used, reviewer notes, corroborating evidence, and the final disposition under the applicable records schedule.
Are AI-assisted government documents automatically improper?
No. Their status depends on how they were created and used. If an AI-assisted output documents official business or supports a decision, it may need to be managed as a record even when a human reviewed and edited it.
What is the best defense against AI-generated misinformation?
A layered system: authenticated publication channels, preserved originals, provenance metadata, automated triage, independent fact-checking, human approval, access controls, and a visible correction process.
Trust is built by the trail, not the score
AI detection can give a records team an early warning. It cannot give the public a complete account of reality. That account comes from a trail that another person can inspect: the original submission, the source system, the approvals, the corroborating records, the analysis, and the correction history.
Governments that build that trail will be better prepared for synthetic documents, altered media, and ordinary human error alike. The goal is not to make every record look human. It is to make every consequential record explainable, traceable, and difficult to alter without leaving evidence.
