A hash in an AI report: proof of a file's identity needs the file
Tie a hash to the file, algorithm and acceptance scope, distinguishing byte integrity from content review and provenance.
Article language: English

In this article
A long string of letters and digits appears at the end of an assistant's report. Beside it is “SHA-256”, followed by “verified version”. It looks more substantial than “done”. But the recipient opens an attachment corrected after the hash was calculated. The report and file are convincing separately, yet they concern different versions. Even a very serious string does not police what someone later attaches to a message.
A hash helps compare data. For accepting AI work, however, it needs neighbours: a specific file, algorithm name, calculation method and a record of what was checked on that version. We organise this evidence using our own working examples. This is not cryptographic analysis or guidance on electronic signatures. It concerns the ordinary situation where someone wants to confirm they are assessing exactly the material handed over.
What does a hash say about a file?
NIST's definition of a hash function describes mapping input data to a fixed-length output and the resistance of cryptographic functions to finding different inputs with the same result. This does not mean collisions are mathematically impossible. In practical acceptance, a hash is a tool for checking byte agreement, not assessing the meaning of stored content.
Ask for the full algorithm name. “Hash” alone does not identify the method, and the appearance of a string does not establish how it was calculated. Two results can be compared only with the same method and input scope. Do not combine a file hash, a selected paragraph hash and a program version identifier in one “evidence” column. Each may refer to a different thing.
If two correctly calculated hashes using the same method differ, the inputs were not identical. The result alone does not show where the change occurred. It might be a corrected number, a line ending or other file data. To establish the cause, compare the relevant versions. A hash indicates a discrepancy; it does not replace reading the difference.
Matching results calculated by both sides usefully confirm agreement under the chosen check. They do not confirm that a file contains true information, error-free code or a current report. Two people can read the same incorrect table. Its integrity fixes none of the calculation. Therefore name the comparison result and the content assessment separately in the protocol.
Start by asking what this check is for. Are you checking a package download, establishing the version tested or connecting a text correction to an accepted document? The choice determines which file to retain and when to calculate the hash again. Without it you can identify with great precision material nobody intended to accept.
Which bytes were actually hashed?
Node.js Hash class documentation describes creating a hash from supplied data, including by reading a file. It also distinguishes text with a specified encoding from bytes. For file acceptance require calculation from the content actually read, rather than a copied description or name in a report.
A name is an organisational hint, not file content. Two documents named “report-final” can have different bytes, while identical bytes can have different names. In a hypothetical example, a copy of a file with the same bytes under another name has the same SHA-256. A version differing by removal of the final newline can have a different hash. The example is not a test of cryptographic strength or client documents.
For text, establish whether someone transforms it before calculation. A tool may change line endings, encoding, spacing or field order on resaving. Such a result may be useful for a separate purpose, but should not pretend to be the original file's hash. Preserve the operation name and version. The recipient must know whether they are comparing received bytes or a representation calculated from them.
Python hashlib documentation describes file_digest for hashing the content of a file opened in binary reading mode. This is a second example of a ready tool working on content. There is no need to change the language used in a project; the agreed check needs an unambiguous input and must be repeatable.
Also ask where the assistant obtained the recorded string. Did it read output from a tool actually run, copy it from an earlier report or merely propose a placeholder for a future result? A correct-looking string establishes none of these. A protocol template with “enter SHA-256 here” is a useful draft but does not confirm calculation. A prepared command is not evidence that it ran either. Acceptance requires a real result from the right input.
If calculation failed because access or the file was missing, or a read failed, record that result. An assistant should not copy an old hash from a similarly named document and present it as a fresh check. Distinguish “provided by the contractor”, “calculated from the file” and “compared with a trusted reference”. Only the last describes an actual comparison.
How can a version be tied to a check?
For each evidence item record its subject: source file, export, archive or material downloaded from a publication. Include the available retrieval location, algorithm, full hash, size and time. Size does not replace a hash, but helps catch an obvious mistake such as receiving a short error message instead of a document. Do not publish private paths or data the recipient does not need.
Separate the material's creation date, hash calculation time and comparison time. Copying an old result into today's document does not mean today's file was read. If an automated process calculates it, retain its actual record and run identifier. If checked manually, identify the person and material used. After a later correction, this establishes event order without treating a cover date as confirmation of every action.

Attach the check's scope to the same record. Full reading, checking a single correction and an automated test are different actions. If one sentence changed after a full read, retain the identifier of the version read and the exact change. A correction can honestly be accepted by reading its context without claiming the whole document was read again. That record says more than mechanically copying a new hash into an old protocol.
In AI-reported code tests (in Polish), this principle links the result to the version actually run. If the test preceded the last change, a new hash does not carry its result into the future. Establish which check remains current and which needs repeating. Merely recording a change is not rerunning a test.
Keep the material the report concerns. A hash without an available file does not permit another reading or identification of the difference. A recipient can see a well-formed label but cannot check its relationship to the handed-over version. The agreed archive therefore needs the artefact itself, meaning the file that is the work's result, and records of the actions concerning it.
For several items, prepare an inventory assigning a hash to each. Do not describe an entire package with a result calculated for one file alone. If checking the archive as a whole, name that scope and retain the exact archive. Checking individual materials after extraction is a separate stage. Different packaging methods can produce different archive bytes with the same files inside.
Does a hash confirm author and origin?
Someone can change a document, calculate a new hash and attach both to a new report. Everything will agree within that set. That is why the reference point matters: who provided it, through which process and which approval it concerns. Do not assume a long string itself establishes sender credibility or the material's creation date.
Do not equate a simple hash with a signature or author attestation. Those require separate mechanisms and appropriate verification. In ordinary acceptance state the boundary honestly: the received file was checked against a result recorded in a specified protocol. If stronger assurance of origin or legal effects is needed, separately choose the scope and expertise instead of expanding the meaning of this one check.
The article on AI-image provenance in company materials (in Polish) develops a similar distinction. A file's history and the truth of the scene depicted answer different questions. For text, a hash will not check cited sources; for a spreadsheet it will not verify the chosen denominator. These remain content-review tasks.
Set access to evidence according to its role. Possession of a hash does not authorise access to a confidential source. Acceptance may require working in an agreed place with a restricted group. Do not upload private files to a random online calculator merely to confirm a number in a report. Choose a tool and environment permitted for the material.
In project handover this check is part of broader acceptance of code, access and confidentiality after AI-agent work (in Polish). Identifying files does not decide library rights, control of accounts or termination of a contractor's access. It does help both sides agree exactly which package they are discussing and what the open comments concern.
What should happen after a mismatch?
First stop unjustified transfer of acceptance to new material. Retain both received versions and check the algorithm and input scope. Establish whether someone compared an archive with an extracted file, source with export or transformed text with an original. Do not call every mismatch manipulation. First identify the actual difference and its significance for the work.
Trace deliberate transformations too. In a working process a source image may be reduced for publication or a text document exported to PDF. The hashes at those stages concern different files. Record the link “this export came from this source” and identify which was reviewed for content and which checked after publication. Do not obtain apparent agreement by replacing the old hash in history. History should explain change rather than make it appear nothing changed.
If a change is justified, record the new version's hash and appropriate scope of renewed acceptance. For a small programming rule, regex trials on valid and invalid data (in Polish) help. For a visibility report, return to Search Console query, page and position scope (in Polish). The content of a change determines the work needed to approve it, not the number of characters in the new hash.
The person taking over should be able to check the material received without reconstructing the author's private computer. Provide a description of the tool and read, an understandable mismatch result and a way to report a missing file. The protocol ends with a decision to accept a particular version or a concrete unknown to resolve. The hash helps keep that decision attached to the right document.
The scope of reviewing AI-work evidence can be agreed with Aeneia Audit. It includes operational material flows, acceptance criteria and limits on claims, without an automatic promise of cryptographic analysis.
This article does not cover verification of signatures, authors' rights or cryptographic algorithm strength.










