What did the AI actually do here? Have the answer ready.
PaperAlly is where you record what you delegated to generative AI, which rule allowed it, and which named human checked it, so that the question has an answer before anyone asks it.
Two audits free to try, and no account needed for them. The working product is licensed, and this page says so here rather than after you have done an hour's work. Your manuscript is stored in the UK, and only a passage you choose to send to an assistant ever leaves. How that works.
Grammarly and Turnitin record what happened while you wrote.
PaperAlly records whether what you delegated was warranted, and who checked it.
How it works
Three moves, repeated once per delegated task. The record is what falls out of doing the work, not a form you fill in afterwards.
Name the task
Say what you are handing over, in one line. It is typed into a task family, and the family is what decides everything after it.
Take the warrant before the run
The family and rules R1 to R7 give the task its warrant, with the reason behind it and the severe error that overrides a good average. You see it before anything runs, not afterwards.
Close it with a named check
A task that needs verification does not close until a person is named and records what they checked and what came of it. There is no button that marks everything verified at once.
What you leave with is the delegation record: every task, its warrant, its six components and its sign-offs. Walk one task through all of it.
Who it is for
The same records underneath, three different readers, and a different name on the artefact for each of them because they are not asking the same question.
Students
your AI use recordYou are allowed to use AI. You are not allowed to be unable to explain it.
I used AI, like everyone does. If someone asks me about it, I need to say exactly what I did, without it sounding like a confession.
Researchers
AI disclosure recordEvery publisher wants the same disclosure in a different place, in different words, at a different threshold.
I have to describe my AI use at submission, months after I did it, from memory. That is how a disclosure becomes a liability.
Reviewers and editors
review method noteSay what you used, or attest that you used nothing, and have the record to stand behind either answer.
I am going to use a model to tighten my prose or to check I have not missed a literature, and I genuinely do not know whether I am allowed to say that.
What makes this different
Including the parts that are a limit rather than a feature, because a product about precision cannot be vague about itself.
Other tools record who typed
Writing-provenance tools record how the text arrived: typed, pasted, assisted. That is a real record and it answers a real question. It does not say which research task was delegated, on what configuration, against what reference standard, or whether anyone verified the output.
Those four are what a warrant record is made of, and the named human check is the one that makes the rest mean anything.
We are not a detector
PaperAlly produces no AI-likelihood score, no probability and no percentage. It records which tasks were delegated, which rule applies, and who checked the output. That is a permanent boundary of the product, not a limitation of this version.
And a clean record never means no AI was used. Nothing in the system could know that, so nothing in the product is allowed to imply it.
Why the refusal, rather than a better detector. The evidence on detection is why this product will not sell one, and the evidence on disclosure is why the record is worth keeping.
Seven detectors flagged 61.22% of 91 TOEFL essays written by non-native
English speakers as AI-generated, against 5.19% for essays by native
writers. Liang, Yuksekgonul, Mao, Wu and Zou, Patterns 4(7):100779,
10 July 2023.
About 70% of 5,114 journals have a generative AI policy. Of 75,000 papers
published since 2023 within a 164,579-paper full-text corpus, 76, roughly
0.1%, disclosed any AI use, with no significant difference between journals
with a policy and journals without one. He and Bu, arXiv:2512.06705.
Free to try, then licensed
Better said now than discovered later. You should be able to judge this properly before anything is asked of you, and then decide.
Try it, with nothing to sign up for
Upload a paper you have already written and read the whole audit: the task families found in it, the warrant each one carries, R1 to R7 and the six components scored per task. Ally, the helper in the corner, answers your questions whether or not you go any further.
The trial is two audits. It limits how many you get and never what an audit tells you: a trial audit is the whole, correct result. A product whose point is a record you can trust cannot hold part of one back to make a sale.
A licence for the working product
Audits without a limit, the paper builder, your delegation record, the record as structured data you can keep, sharing a record with a supervisor or a reviewer, and the institutional surfaces.
Licences are arranged directly at the moment, so tell us what you are working on and we will sort one out: ask about a licence. For a university or a research institute, one agreement can cover everybody: for institutions.
Counting a trial means recognising a repeat visitor, so a trial audit stores a one-way salted hash of your network address and never the address itself. It counts the trial, it matches nothing anywhere else, it is deleted after 30 days, and you can have it removed sooner by asking. What that means in full.
The method, in the open
Published research methodology, operationalised. The warrant states, the six components and the seven rules are the standard itself, and the engine that applies them is deterministic: the same manuscript gives the same reading every time.
Three warrant states, always named in full
The bounded output may enter the workflow after routine provenance and formatting checks.
Useful as a candidate but cannot support a claim until a named check is complete.
The evidence-bearing decision remains human.
The minimum warrant record, six components
One task is warranted when all six can be answered about it. A component marked absent means the signal was not found, which is not the same as the work not having been done.
- Task and outputWhat finite output is delegated?
- ConfigurationWhat shaped the output?
- Evidence accessWhat could the system inspect?
- Reference standardWhat makes the output right or wrong?
- Severe error and repeatabilityWhat error overrides average performance, and does the result repeat?
- Named human checkWho checks what, when, and with what outcome?
The procedure, R1 to R7
- R1Decompose at claim-bearing task level
What exact operation and output are being judged? - R2State the authority at stake
What may the output change if accepted? - R3Bind evidence and source access
What packet, source passage or gold label makes checking possible? - R4Pre-specify metrics, thresholds and overrides
What evidence is sufficient, and what failure is disqualifying? - R5Execute and log a configured run
Which model, prompt, packet, tools and settings produced this output? - R6Verify at the right epistemic level
Is the check about record identity, field extraction, interpretation or proposition support? - R7Issue, retain and revisit the warrant
Who may rely on the output now, and when must the decision be reopened?
Start with a paper you have already written
Upload it and read the audit: the task families found in it, the warrant each one carries, R1 to R7, and the six components scored per task. Nothing is asked of you to see the result on screen, and the first audits are free to try.
Not a detector. No AI-likelihood score. A clean record never means no AI was used.