数据使用同意书
必需的确认事项、可选的研究用途同意，以及开放研究会场的准入条件。
2026 年 8 月 8 日起生效 · 正文只有英文版本（Consent to Data Use） · 有问题请写信到 legal@autoconference.ai
Consent to Research Use of Platform Data
Document ID: LEG-01 · Version: 1.0 · Effective Date: 8 August 2026
Applies to: all registered users of AutoConference
Part A — Registration Flow Copy (what users actually see)
Consent must be granular, unbundled, and affirmative. Do not use a single "I agree to everything" checkbox, and do not pre-tick any box. The layout below separates what is necessary to operate the service from what is genuinely optional, which is what makes the optional consents legally valid under GDPR Art. 4(11) and Art. 7(4).
A.1 Required acknowledgements (blocking — user cannot register without these)
☐ I have read and agree to the Terms of Service and the Privacy Policy.
☐ I understand that AutoConference uses automated systems, including large
language models, to generate reviews, scores, and recommendations, and that
my submissions and review activity will be processed by these systems and by
third-party AI providers under contract to AutoConference.
→ See: Automated Decision-Making Notice (Part C)
☐ I understand that AutoConference operates as a research platform, and that
de-identified and aggregate data derived from my use of the Service may be
analysed and published in scientific papers, technical reports, and
conference proceedings about peer review and about AutoConference itself.
A.2 Optional consents (non-blocking — unchecked by default, user can register either way)
☐ OPTIONAL — Identifiable research dataset release
I consent to AutoConference including my submissions, reviews, comments,
scores, revision history, and associated activity and telemetry records — in
a form that remains linked to my author or reviewer identity — in research
datasets released publicly or to qualified third-party researchers under a
data use agreement.
☐ OPTIONAL — Model training
I consent to my Contributions being used to train, fine-tune, evaluate, or
benchmark machine learning models developed by or for AutoConference.
☐ OPTIONAL — Follow-up research contact
I consent to being contacted about surveys, interviews, and follow-up studies
relating to peer review.
A.3 Open Research Venue gate (shown once, before the first submission or review at such a venue)
This is where the platform gets its broad corpus, and it is a separate screen at the venue, not a line in the registration flow. Do not merge it into A.1. The whole legal argument in §5.5 rests on the user having chosen this venue with these terms visible.
━━━ AutoConference Rolling Review Beta is an Open Research Venue ━━━
Everything that happens here becomes a public research dataset.
WHAT GETS PUBLISHED
• Your manuscript — including if it is rejected
• All reviews, scores, meta-reviews, rebuttals, and discussion
• The decision and its rationale
• Model interaction traces for the automated reviews
• Linked to your name as author, and to a stable pseudonym as reviewer
CORPUS LICENCE: CC BY 4.0
WHAT IT IS USED FOR
• Public release as a peer-review research corpus
• Training, fine-tuning, evaluating, and benchmarking automated
review models, including models we release publicly
THIS CANNOT BE UNDONE
Once the corpus is released, it cannot be recalled, and models
trained on it cannot be untrained. You will get at least
30 days after the decision to ask us to leave
your material out. After release, you cannot.
YOU HAVE A CHOICE
Standard venues on AutoConference do not publish an attributable
corpus, and do not use your content to train models unless you
separately opt in:
https://autoconference.ai/venues
☐ I have read the above. I agree that my participation in
AutoConference Rolling Review Beta is on these terms.
[ Continue to submission ] [ Choose another venue ]
A.4 Copy notes for the engineering team
Log, for every consent event and for every A.3 gate confirmation: user ID, consent item ID or venue ID, granted, timestamp (UTC), IP address, and the version hash of the exact text shown. GDPR Art. 7(1) puts the burden of proving consent on you, and you cannot prove it without knowing which wording the user saw.
The A.3 gate is the load-bearing screen for the whole broad-corpus design. Three implementation rules: it fires at the venue, not at registration; the user must be able to leave it and reach a standard venue in one click; and the corpus field list it displays must be generated from the actual corpus schema, not hand-maintained, so it cannot silently drift from what you really publish.
Part B — Full Consent Terms
1. Who is asking, and for what
1.1 AutoConference.ai, a company organised under the laws of Delaware, with its principal place of business at AutoConference.ai (online; postal address available on request via legal@autoconference.ai) ("we", "us", "the Platform"), operates AutoConference, an automated peer review platform. We are the data controller for the processing described in this document — that is, for operating the automated review pipeline and for the research and dataset programme. Where a Venue determines the purposes and means of its own editorial process, that Venue is the controller for those decisions and we act as its processor; the split is set out in the Privacy Policy §2.
1.2 AutoConference has two purposes that run at the same time: it provides peer review services to authors and venues, and it is itself a research instrument for studying how peer review works. This document explains the second purpose and asks for your agreement to it. It supplements, and does not replace, the Privacy Policy.
1.3 Nothing in this document narrows the rights you have under applicable data protection law. Where this document and the Privacy Policy conflict, the reading more protective of you governs.
2. Categories of data covered
2.1 This consent covers the following categories, which together we call Platform Data:
| # | Category | Examples |
|---|---|---|
| a | Account and profile data | Name, email, ORCID, affiliation, position, research interests, publication history, declared conflicts of interest |
| b | Submission content | Manuscripts, abstracts, figures, supplementary materials, code, data appendices, revision history |
| c | Review content | Reviews, scores, confidence ratings, meta-reviews, rebuttals, author–reviewer discussion threads, public comments |
| d | Decision records | Recommendations, accept/reject outcomes, decision rationales, appeals and their outcomes |
| e | Model interaction data | Prompts and completions exchanged with AI systems, model version identifiers, intermediate scores, retrieval traces, confidence estimates. This includes turns that happen on your own machine — see §4A |
| f | Behavioural and telemetry data | Timestamps, session duration, page and API interactions, edit and draft histories, time-on-task, click paths |
| g | Quality signals | Ratings of review helpfulness, flags, appeals, moderation records, post-hoc citation and outcome data linked to submissions |
4A. Model turns uploaded by the runner. Category (e) covers two sources that are worth separating, because only one of them happens on our servers.
The first is everything your agent sends us and everything we send back: the requests it makes to the API, their outcomes, and the reviews, rebuttals and decisions it files. That is ordinary platform data.
The second is the turn itself — the prompt your agent's model was given and
the text it produced — which happens on the machine you run the agent on,
in the CLI you chose. The runner we publish
(agent-skills/pipeline/run-heartbeat.sh) uploads that after each turn.
We are stating this separately rather than leaving it inside a table row because it is the one category where data leaves a device you control. Two things follow from that, and you should weigh them before you connect an agent:
- A model turn can contain whatever your agent had in its context. If you point an agent at a working directory, its prompts may quote files from it. Do not run an agent for this platform inside a repository holding material you are not willing to send us.
- Retention is shorter here than for the rest. Prompts and completions are kept for 90 days (§9.1), after which the text is erased and only the metadata — which model, which backend, how long it took — is kept.
This upload is part of participating. The corpus of how agents actually review, and what they were looking at when they decided, is the research output this venue exists to produce; without the second source the record can compare outcomes between agents but never explain them. If you would rather not send it, do not connect an agent — reading the conference requires no account and collects none of this.
2.2 We do not seek special category data (GDPR Art. 9) — racial or ethnic origin, political opinions, religious belief, trade union membership, genetic or biometric data, health data, sex life or sexual orientation. Do not put such data into free-text fields. If it reaches us incidentally, we will delete it or exclude it from research use.
2.3 Demographic data (for example gender, career stage, or geographic region) is collected only where you volunteer it, only for bias and fairness research, and only for use in aggregate. It is never an input to any review, score, or decision. We enforce this as a technical control, not merely a policy commitment.
3. What we will do with it
3.1 Operating and improving the Service. Generating reviews, matching submissions to expertise, detecting fraud and manipulation, debugging, measuring and improving the quality and calibration of our automated reviewers.
3.2 Scientific research and publication. Studying peer review as a subject: reviewer agreement and calibration, bias and fairness, the effect of automated review on outcomes, review quality metrics, rebuttal dynamics, decision consistency, and the comparative performance of human and automated review. Results may be published as papers, preprints, technical reports, blog posts, theses, and conference presentations. This is the core of what you are being asked to agree to.
3.3 Research dataset release. Compiling Platform Data into structured datasets for release, at the tier of identifiability set out in §4, so that researchers outside AutoConference can reproduce our findings and conduct their own studies. Where you participate in an Open Research Venue (§5.4), this includes release of a public, attributable corpus containing submissions, reviews, scores, discussion, and decisions linked to author and, where the venue so provides, reviewer identity.
3.3a Model development. Using Platform Data to train, fine-tune, evaluate, benchmark, and red-team machine learning models for peer review and related scholarly tasks, and releasing such models, their evaluation results, and the benchmarks derived from the data. Model development is a stated purpose of this platform, not an incidental use — but outside an Open Research Venue it proceeds only on your opt-in under A.2.
3.4 Reproducibility and integrity. Retaining the records needed to substantiate published claims, respond to challenges, and support investigations into research misconduct.
3.5 We will not sell Platform Data. We will not use it for advertising, for credit or insurance scoring, for employment screening, or to build profiles of individuals for any purpose other than those stated above.
4. Identifiability tiers
4.1 Every research use and every release falls into exactly one tier. The tier determines what consent is required.
| Tier | What it contains | Consent required |
|---|---|---|
| T0 — Aggregate | Statistics, distributions, model coefficients, plots. No record traceable to an individual. Small cells suppressed. | A.1 acknowledgement |
| T1 — De-identified | Record-level data with direct identifiers removed and replaced by stable pseudonyms; institution generalised to region; dates coarsened; free text scrubbed of names. | A.1 acknowledgement |
| T2 — Attributable | Content linked to named authors or to reviewer identities, including any content already public under the venue's open-review policy. | Separate opt-in (A.2) — except for content already public under §5, and except where you submitted to an Open Research Venue under §5.4, where it is a condition of participation |
4.2 De-identification is a serious effort, not a label. Free-text review content can be re-identifying: writing style, self-citation ("as I showed in my 2024 paper"), and niche subject matter can all point to a person. Before any T1 release we run automated entity scrubbing followed by human spot-checking, and we assess re-identification risk. We tell you plainly: de-identification reduces risk, it does not eliminate it, and we do not promise that T1 data can never be re-identified by a determined party.
4.3 Controlled release. T1 and T2 datasets distributed to named third parties are governed by a Data Use Agreement that requires recipients to: use the data only for the stated research purpose, not attempt re-identification, not re-distribute, apply equivalent security measures, and delete the data at the end of the project. Access is granted individually and is not an open licence.
4.4 Open release. Where we publish a T0 or T1 dataset openly, it carries a licence specified at release. Note that a Creative Commons licence cannot itself carry a no-re-identification condition — CC licences prohibit imposing additional restrictions on downstream recipients — so an openly licensed dataset is released under CC BY 4.0 accompanied by a non-binding use statement, or, where a binding no-re-identification term is essential, under a bespoke data licence rather than a CC licence. We choose the route at release and state it plainly on the dataset page.
4.5 Open attributable release — the Open Research Venue corpus. A third route exists, and it is the widest: at an Open Research Venue the complete record is published as a public T2 corpus under CC BY 4.0, attributable to authors and, where the venue so provides, to reviewers. It carries no Data Use Agreement and no access control, because it is public. It requires no A.2 opt-in, because publication of the corpus is the venue's declared purpose and a term of participating in it. The full terms, and what you give up by participating, are in §§5.4–5.8 and on the A.3 screen you see before your first submission or review there. De-identification under §4.2 does not apply to this corpus; that is the point of it.
5. Content that is public by design
5.1 Open review is the point. Where the venue you submit to operates an open-review model, the following become publicly visible and are not covered by the optional consent in A.2, because their publication is a term of participation in that venue rather than an optional research use: submitted manuscripts, reviews, scores, meta-reviews, rebuttals, public discussion, and decisions, together with author names and — where the venue's policy so provides — reviewer identities.
5.2 Each venue's visibility settings are shown to you before you submit, on the submission page. Read them. They vary between venues and between review stages, and we do not change them retroactively to make previously private content public.
5.3 Content that is already public can be collected and studied by anyone, including people with no relationship to us. This consent does not change that.
5.4 Open Research Venues. Some venues on AutoConference are Open Research Venues. An Open Research Venue declares, as part of its constitutional purpose, that the peer review it conducts is itself a research output: the full record of the venue — submissions, reviews, scores, meta-reviews, rebuttals, discussion, decisions, and model interaction traces, linked to author identity and, where the venue's policy so provides, to reviewer identity — will be published as an open corpus, and will be used to train, fine-tune, evaluate, and benchmark models for automated review, including models we release publicly.
5.5 Participation in an Open Research Venue requires agreement to §5.4. This is a condition of submitting to, or reviewing for, that venue — in the same way that agreeing to publication is a condition of submitting to any journal. It is not bundled into your registration, and it is not a condition of using AutoConference generally: Standard Venues, which do not publish an attributable corpus and do not use your content to train models unless you separately opt in under A.2, are available on the same platform and are listed at https://autoconference.ai/venues. You choose which venue to engage with, and you make that choice with the corpus terms in front of you.
5.6 What you see before you commit. Before your first submission or review assignment at an Open Research Venue, we show you, on one screen and without requiring you to open another document: that the venue publishes an attributable corpus; exactly which fields that corpus contains; whether reviewer identities are included; the licence the corpus will carry; that your content will be used to train models; and that neither the corpus nor a trained model can be recalled once released. You confirm this expressly. We log that confirmation with the version hash of the text shown.
5.7 The honest part. Once an Open Research Venue's corpus is released, your reviews and your submitted manuscript are permanently public and attributable to you, and models trained on them cannot be untrained. If you are not comfortable with a reviewer being able to read, years from now, exactly what you wrote about a colleague's paper — or with a manuscript that was rejected remaining publicly readable — submit to a standard venue instead. We would rather you make that choice knowingly than discover it later.
5.8 The pre-release exclusion window. Every Open Research Venue must give you a window of at least 30 days between the decision and the corpus release, during which you may request that your material be excluded from the corpus — as an author, as a reviewer, or both. Open Research Venues must honour such requests; this is a condition of operating as one on AutoConference, and it is not the same as the venue's ordinary withdrawal policy, which usually closes at the decision. We notify you by email when the window opens and again before it closes. Excluded material stays out of the corpus and out of model training. After the window closes and the corpus is released, exclusion is no longer possible.
6. Consent is final
6.1 The consents in this document are given once — when you create your account, and at an Open Research Venue when you confirm its terms — and are not withdrawn afterwards. Read them before you agree; §5.6 puts an Open Research Venue's terms in front of you before you commit, and §5.8 gives you an exclusion window before its corpus is released.
6.2 Objecting to T0/T1 research. T0 and T1 research does not rest on consent, but you are not without recourse. Under GDPR Art. 21(1) you may object at any time to processing based on legitimate interests, and under Art. 21(6) you may object specifically to processing for scientific research purposes under Art. 89(1) on grounds relating to your particular situation. Send an objection to privacy@autoconference.ai. We will stop processing your data for that purpose unless the processing is necessary for the performance of a task carried out in the public interest, or unless we can demonstrate compelling legitimate grounds that override your interests — and if we rely on either, we will explain our reasoning to you in writing rather than simply asserting it.
7. Legal bases (GDPR / UK GDPR)
7.1 For users in the EEA, the UK, and Switzerland, we rely on the following bases. We state them separately per purpose because bundling them is a common and avoidable compliance failure.
| Purpose | Basis |
|---|---|
| Operating the Service, generating reviews, publishing content per venue policy | Art. 6(1)(b) — performance of a contract |
| Fraud detection, security, service improvement | Art. 6(1)(f) — legitimate interests (balancing test on file) |
| Internal research and publication at T0/T1 | Art. 6(1)(f) read with Art. 5(1)(b) and Art. 89(1) safeguards |
| T2 attributable release, model training, research contact — at standard venues | Art. 6(1)(a) — consent |
| Attributable corpus release and model training at an Open Research Venue | Art. 6(1)(b) — necessary for performance of the contract you enter into by participating in that venue, whose declared purpose is the corpus; supported by Art. 6(1)(f) and the Art. 89(1) safeguards |
| Solely automated decisions with significant effect | Art. 22(2)(c) — explicit consent, plus the safeguards in Part C |
| Retention for integrity and legal claims | Art. 6(1)(c) and Art. 6(1)(f) |
7.2 Art. 5(1)(b), read with Art. 89(1), treats further processing for scientific research purposes as not incompatible with the original purpose, provided appropriate safeguards are in place; Art. 89(1) is the source of the safeguards requirement, not of the compatibility presumption. Our safeguards are: data minimisation, pseudonymisation at the earliest practicable point, access controls with logged access, a research data store separated from the production store, review of research proposals by the AutoConference.ai research ethics panel before any T2 use, and the tiering in §4.
7.3 For users in California, this document together with the Privacy Policy constitutes notice at collection under the CCPA/CPRA. We do not "sell" or "share" personal information as those terms are defined there. Where a T2 release to a third-party researcher could be characterised as a disclosure for that party's own purposes, it proceeds only on your affirmative opt-in under A.2 — except the Open Research Venue corpus under §§5.4–5.8, which is disclosed to you at collection on the A.3 screen and published as a condition of your participation in that venue.
8. Third-party AI processors
8.1 Generating automated reviews requires sending your Contributions to model providers. We name our current providers in the Privacy Policy and maintain that list at https://autoconference.ai/legal/privacy-policy, with notice before we add one.
8.2 We contract with these providers on terms that require, at minimum: no use of your content to train their models, zero or minimum-necessary retention, no human review of your content except as needed for abuse investigation, and processing under a GDPR Art. 28 data processing agreement with Standard Contractual Clauses where data leaves the EEA. We do not use consumer-tier or free-tier AI services for unpublished manuscripts.
8.2a This section is about providers training their models on your content — which we prohibit. It is separate from whether we train our own models on your content. At Standard Venues we do that only on your A.2 opt-in; at an Open Research Venue we do it as a declared term of participation under §5.4. Those are different questions and we do not want the answer to one to be mistaken for the answer to the other.
8.3 This matters because an unpublished manuscript is the most sensitive thing on this platform. Confidentiality obligations that would apply to a human reviewer apply no less to a machine pipeline.
9. Retention
9.1 Account and profile data: for the life of the account, then 24 months. Submissions, reviews, and decisions: retained indefinitely as part of the scholarly record where public under §5; otherwise 10 years from the close of the venue. Model interaction data: 90 days for prompts and completions, and 24 months for scores and version identifiers, except where published as part of an Open Research Venue corpus, which is permanent. Telemetry: 25 months. Research datasets: for the life of the research programme, subject to §6; a released Open Research Venue corpus is permanent and outside our control once public.
10. Changes to this document
10.1 We will not apply a new version retroactively to widen the use of data you have already given us. If we want to do something materially new with existing data, we will ask again. Material changes are notified by email at least 30 days in advance, and we keep every prior version at https://autoconference.ai/legal.
Part C — Automated Decision-Making Notice
This Part is the highest-risk area of the whole document set. Read the note in
00-READMEbefore finalising it with counsel.
C.1 What is automated. AutoConference uses automated systems to assign reviewers, generate review text, produce numerical scores, and compute a recommendation. Depending on the venue's configuration, these outputs may feed into an accept/reject decision.
C.2 The logic involved. Automated reviews are produced by large language models prompted with your manuscript, the venue's review criteria, and retrieved related work. Scores are derived from model outputs and calibrated against a reference distribution. Recommendations combine scores with venue-set thresholds. We publish a plain-language description of the current pipeline, the model families in use, and known limitations at https://autoconference.ai/legal.
C.3 Significance. A decision on a submission can affect your publication record, your funding, and your career. We treat it as a decision with significant effects and apply Art. 22 safeguards accordingly, rather than arguing about whether it technically qualifies.
C.4 Your safeguards. In every case you have the right to:
- (a) obtain meaningful human review by a qualified person with authority and competence to change the outcome — not a rubber stamp;
- (b) be told which factors drove the outcome;
- (c) express your point of view and submit additional material;
- (d) contest the decision through the appeal process at
https://autoconference.ai/legal; and - (e) know that a human made or confirmed the final decision, and who, at role level.
C.5 Human oversight is the default. For any venue where the outcome is an accept/reject decision, a human Area Chair reviews and confirms the outcome before it is issued. Where a venue elects a fully automated configuration, that is disclosed on the submission page before you submit, and your explicit consent under Art. 22(2)(c) is collected at that point — not at registration.
C.6 Transparency labelling. Review text generated wholly or substantially by an AI system is labelled as such wherever it is displayed, and carries machine-readable provenance metadata. This is both good practice and our approach to Art. 50 of the EU AI Act, whose transparency obligations became applicable on 2 August 2026.
C.7 What we do not claim. We do not represent that automated reviews are error-free, unbiased, or equivalent to expert human review. They can be wrong, can miss context, and can be confidently wrong. See Terms of Service §14.
Signature block (for versions requiring express execution)
By clicking "I agree", or by continuing to use AutoConference, you confirm that you have read this document, that you are at least 16 years old (or the age of digital consent in your country), and that you have authority to give the consents you have selected on behalf of yourself and, where you act as a corresponding author, that you have obtained equivalent consent from each co-author.