"Human Reviewed" AI Content Needs Evidence, Not Just a Promise

Many small teams now use AI to draft, summarize, research, edit, and prepare written work.

That is not the problem.

The problem appears later, when the work is handed to someone else and described with a sentence like:

This was human reviewed.

That sentence is becoming important. Clients hear it. Editors hear it. Reviewers hear it. Customers hear it. But in most workflows, it is still only a promise.

There is usually no small, portable record showing:

• what final file was reviewed;
• what AI contributed;
• whether review was complete, partial, or absent;
• who accepted responsibility;
• whether the record still verifies later.

I built Pramaan as an open-source experiment for that narrow gap.

GitHub: https://github.com/sushilsoni-tech/pramaan

What Pramaan Is

Pramaan is a local tool for creating and verifying signed responsibility records for AI-assisted written work.

It does not try to detect AI writing.

It does not judge whether the content is true.

It does not certify legal compliance.

It does not prove every real-world action was recorded.

Instead, it creates a signed record for one finished file.

That record can say:

• this is the exact final file hash;
• this is what AI helped with;
• this is whether human review was complete, partial, or absent;
• this is who accepted responsibility;
• this is the public key that signed the record;
• this is whether the signed bundle still verifies later.

The recipient can verify the bundle on their own machine.

For identity trust, they still need to compare the bundle's signer fingerprint against a fingerprint received through a separate trusted channel.

Why This Might Matter For Small Teams

Large organizations may eventually buy governance platforms, connect identity systems, and enforce audit workflows across every tool.

Small teams do not start there.

A small content studio, research writer, AI-assisted proposal shop, grant writer, technical writer, or responsible AI consultant often needs something much lighter:

I used AI, a human reviewed the final work, and here is the record for this specific deliverable.

That is the first use case I am testing.

The goal is not bureaucracy.

The goal is to make a human-review claim inspectable.

What It Does Not Prove

This boundary matters.

Pramaan does not prove that the underlying content is true.

It does not prove that the human review was good.

It does not prove the named reviewer is who they say they are. In the current editorial profile, reviewer identity is a producer-supplied assertion, and the signature binds the record to the producer's key, not to a separate reviewer-held key.

It does not provide trusted timestamping by itself.

It does not prove legal, regulatory, editorial, or academic compliance.

It proves something smaller:

This signed record has not changed, it refers to this final file, and it satisfies or fails the declared checks.

That smaller claim is the product.

Try The Samples

The repo includes one PASS sample and one FAIL sample:

git clone https://github.com/sushilsoni-tech/pramaan.git
cd pramaan
python -m pip install .
pramaan verify samples/editorial-pass
pramaan verify samples/editorial-fail-missing-reviewer

The PASS sample contains a complete declared review record.

The FAIL sample has intact signed files, but fails overall because substantive human review and a responsible person are missing.

That distinction is important. Pramaan is not a badge generator. It is allowed to say that a signed record is intact but the declared review is not satisfied.

Who I Am Looking To Learn From

This may be useful if you already make a public or client-facing human-review promise:

• AI-assisted content studios;
• SEO and editorial agencies;
• freelance researchers or writers;
• grant and proposal writers;
• technical writers;
• responsible AI consultants;
• small teams sending AI-assisted written deliverables to someone else.

If nobody ever asks you to explain or evidence human review, Pramaan may be unnecessary.

If your clients already ask what AI contributed, who reviewed the work, and who is accountable, Pramaan may be a useful primitive.

The Open Question

The question is not whether AI-assisted work will continue.

It will.

The question is whether "human reviewed" remains an unverifiable phrase, or becomes something a second person can inspect.

Pramaan is a small open-source attempt at the second path.

I am especially interested in blunt feedback on three questions:

1. Is this useful, or is it solving the wrong trust problem?
2. Is the boundary clear enough, or does it still sound like overclaiming?
3. What would make this understandable to a nontechnical client or reviewer?

GitHub issue for feedback:
https://github.com/sushilsoni-tech/pramaan/issues/3