AI Transparency Statement

1. You are interacting with artificial intelligence

THIS SERVICE IS AN ARTIFICIAL INTELLIGENCE SYSTEM

EVERY REWRITE, DRAFT, SUGGESTION, NATURALNESS SCORE AND AI-DETECTION SCORE RETURNED BY THE SERVICE IS MACHINE-GENERATED OR MACHINE-CALCULATED. IT HAS NOT BEEN WRITTEN, REVIEWED OR VERIFIED BY A PERSON BEFORE IT REACHES YOU.

When you use a writing form, style-matching feature, AI writer or scoring feature, you are interacting with a generative artificial intelligence system operated by humanizes, the operator of humanizes.com. A person is available through the contact routes below, but a person does not silently take over the interaction or approve each result.

This statement describes the systems behind the humanizes service (the Service) and is part of the Terms of Service. It applies both to text transformed from material you supply and to text drafted from your instructions.

2. Systems that process or evaluate text

Models and providers change over time, so we identify their function rather than a model version. The current processing chain is:

A score does not cause the Service to make a decision about a person. It is information presented to the user, who decides whether and how to use it.

3. What rewriting is intended to change

For a rewrite request, the Service is intended to change expression rather than the underlying proposition: sentence structure, rhythm, vocabulary, transitions and other stylistic features may be altered to make text read more naturally. The AI writer instead generates a new draft from the instructions and context supplied.

That intention is not a guarantee that meaning will be preserved. A rewrite may omit a qualification, strengthen or weaken a claim, change tone, alter a number or name, introduce a contradiction, or add a proposition that was not in the input. The Service does not compare the result against an authoritative source and does not know which details are legally, academically or technically essential.

You should compare a rewrite with the original line by line where fidelity matters. For a generated draft, verify every factual proposition against reliable sources independent of the Service.

4. Known limitations

Output and scores require independent human checking

Factual error and fabrication

Output can confidently state something false, conflate people or events, use stale information, invent a quotation, reference or citation, or give a real-looking source that does not exist. A citation must be opened and checked against the proposition it supposedly supports.

Originality and rights

Generative output can resemble existing text or another person’s work. We do not conduct a plagiarism or rights-clearance search and do not guarantee originality, novelty, ownership or non-infringement. Similar instructions may produce similar output for different users.

Detector uncertainty

AI detectors infer patterns; they do not observe who wrote a passage. Scores from different tools can disagree, can change after a tool is updated, and can produce false positives on genuinely human writing as well as false negatives on generated writing. A score is an estimate, not proof, certification or a promise that text will pass another detector.

Style matching

A style profile approximates statistical features of uploaded samples. It does not reproduce a person perfectly, prove that person wrote the result, transfer authorship, or establish permission to imitate them.

Language and domain limits

Results may degrade for non-English text, mixed-language text, dialect, poetry, code, tables and specialist legal, medical, scientific or technical material. Domain terminology and formatting can be lost or changed.

Non-determinism

Generative systems are probabilistic. Repeating the same request can produce a different result, and a model or pipeline update can change results over time.

The fuller contractual disclaimer is in the Terms of Service. Output is not professional advice and must not be treated as a substitute for a suitably qualified person where consequences are material.

5. Review, fact-checking and disclosure by users

Before publishing, submitting, sending or relying on output, you must use meaningful human review. At a minimum:

• read the entire output and compare a rewrite with the source text;
• verify facts, names, dates, calculations, quotations, links and citations using reliable independent sources;
• check for confidential, personal, biased, discriminatory, harmful or infringing material;
• edit the result so that it accurately expresses what you intend and is suitable for its audience; and
• disclose the use of AI wherever a school, university, employer, client, publisher, platform, professional rule, contract or law requires disclosure.

Do not remove or falsify a disclosure or marking in order to mislead someone about how text was made. Academic and other deceptive uses are governed by the Acceptable Use Policy, including its rules on misrepresentation and assessment work.

A reader cannot reliably identify all AI-assisted text by tone, wording or a detector score alone. If provenance matters, ask the author about their process and request source materials or a declaration appropriate to the context rather than treating a detector result as conclusive.

6. Provenance and machine-readable marking

What exists today

We do not currently embed C2PA Content Credentials or another documented provenance signal in output text. We do not currently add an invisible watermark to output text. We do not currently offer a public tool that tests whether text was generated by the Service.

Robust watermarking is particularly difficult for plain text. Ordinary editing, translation, paraphrasing, copying through another application or taking a short extract can weaken or remove statistical signals. A signal strong enough to survive those changes can also distort language, fail across languages or incorrectly implicate human writing. We will not describe a probabilistic detector as certainty.

We are designing the following measures so they can be layered onto generated output as the relevant duties apply and technical standards mature:

1. Machine-readable marking of generated output in an appropriate, interoperable and technically feasible format, designed to remain detectable where reasonably possible.
2. A documented provenance signal that identifies output as generated or materially altered by the Service, with public documentation of what the signal does, its limitations and how common transformations affect it.
3. A free public detection tool through which a person can test whether content contains our provenance signal, with results expressed as signal findings rather than a guarantee of authorship.

The California AI Transparency Act is operative from 2 August 2026. Our understanding is that its provenance-disclosure and public detection-tool duties attach to a covered generative AI provider once the statutory monthly-user threshold is crossed. The Service is below that threshold today. If it crosses the threshold and the Act applies, we will implement the required latent provenance disclosure, provide the required option for a manifest disclosure where applicable, and make the required free public detection tool available within the statutory timetable.

Separately, our understanding of the EU AI Act is that providers of generative AI systems must ensure generated or manipulated content is marked in a machine-readable format and detectable as artificially generated or manipulated, so far as technically feasible, effective, interoperable and robust. Our planned machine-readable marking and documented signal are intended to support that duty when it applies. Until those measures are live, users must not imply that output carries provenance metadata that it does not have.

7. Transparency regimes considered

The following is our operational understanding of duties relevant to the Service. It explains our compliance design; it is not legal advice to a user about their own obligations.

These regimes and their implementing guidance may change. We may adjust the technical form and placement of notices or marking to meet the law without changing the basic fact disclosed here.

8. Automated processing, style profiles and training

On the plan that includes style matching, the Service derives a style profile from writing samples the user chooses to upload and uses that profile to influence later output. This is automated analysis of writing characteristics, not identity verification, psychological assessment or a finding about who authored the samples.

The Service does not make a decision with legal or similarly significant effect about anyone by automated means. It generates text and informational scores for the user. A school, employer, publisher or other third party that chooses to make a decision using text or a score does so independently of us and should not treat the score as conclusive.

We do not use submitted text, writing samples, style profiles or messages to train or fine-tune our own models. Our AI processors receive that material under terms that do not permit them to train on it. The Privacy Policy explains the personal-data basis, retention and rights, and the Subprocessor list identifies recipients.

9. Human oversight and contacting a person

Human oversight happens at two levels. The user reviews each result before use. Our team can investigate a reported output, scoring or account problem, explain the role of the systems described here, and handle a privacy or misuse report. Support cannot certify that output is true, original, lawful, human-authored or acceptable under a third party’s rules.

To reach a person, email support@humanizes.com. For a privacy right, use the privacy request form or email jakemorris@humanizes.com.

10. Reporting a problem with output

Report harmful, unlawful, materially inaccurate or unexpectedly identifying output, a fabricated citation, a scoring problem, or suspected misuse to support@humanizes.com. Include the feature used, the approximate date and time, what went wrong and, where safe, the relevant input and output. Do not email secrets or sensitive personal data that are not needed to investigate.

We may ask for account or request details to locate the event, preserve relevant security records, test whether the problem can be reproduced, and take proportionate corrective action. Reporting does not mean we can recover text that has already been deleted or guarantee a particular outcome.

Operator

humanizes, the operator of humanizes.com.

Website

https://humanizes.com

Support and output reports

support@humanizes.com