AI Transparency Statement

1.Purpose of this statement

Because machine learning is built into several parts of the platform, we describe here — in one place and in plain terms — where it operates, whose systems perform it, what its limits are, and how people stay in control. We update this statement when the platform’s use of AI changes materially.

2.Where AI is used

• Submission assistance. When you paste an announcement into the seminar, conference, workshop, job, or abstract form, a language model extracts the structured fields (title, dates, speakers, location) to pre-fill it. You review and can correct everything before submitting.
• Summaries and descriptions. Summaries of seminars posted on World Wide, and some topic overviews, are written by a model from the underlying material.
• Classification. A model assigns listings, including imported ones, to scientific fields and suggests their topic tags, so that browsing and filtering work across disciplines. Topic tags are shown on listings.
• Related listings. The text of listings is converted into numerical representations (embeddings) used to find related work and to order a seminar calendar feed when you choose to personalise it. Profile text, including the contents of a CV you upload, is also converted; no current feature uses those profile embeddings. Search matches words and does not use embeddings.
• Moderation assistance. Models screen submissions for policy violations. Automated rules can approve or reject content; people can review contested outcomes.
• Decorative imagery. Banner images for scientific fields are AI-generated artwork; they depict no real people or events.
• Checking model output. Every model response is checked by rules and by a second model before it is used; a response that fails the check is discarded.
• Operations. Administrators can have a model summarise a small set of recent log and audit entries while investigating a problem.

3.What AI is not used for

• AI screening is used for content publication, not to determine eligibility for employment, credit, insurance, or public benefits. A publication decision can be contested through our reporting and human-review process.
• No emotion recognition, biometric categorisation, or identification of natural persons.
• No advertising profiles, and no sale of data to AI vendors.
• Uploaded audio and video are not automatically transcribed.

4.Provider and data handling

Text generation, classification, moderation, summaries, and autofill use the same OpenAI models hosted on Amazon Bedrock, billed by AWS. Requests enter through Ireland with global inference profiles, which can process data outside the EU. AWS handles this content; the model provider does not receive it. Every Responses request sets store: false to disable application-state storage. AWS may nevertheless retain flagged traffic for up to 30 days for abuse detection. Embeddings are currently computed with Amazon Titan Text Embeddings on Amazon Bedrock in the EU (Ireland) region; OpenAI’s embedding service remains the alternative. Field artwork is generated through OpenAI’s API (EEA contracting entity: OpenAI Ireland Ltd) from prompts that contain no personal data, when that generation runs. Under its API terms, content is not used for model training unless the customer opts in; standard abuse logs may be retained for up to 30 days. World Wide has no special zero retention agreement. The Privacy Notice describes recipients and international transfers.

We send only the text a task needs: the content being processed, the profile text being indexed, or the log entries an administrator asked about. We do not send credentials, analytics event records, or data unrelated to the task. A document or profile can itself contain an email address or other personal data, so review material before invoking a feature.

5.Machine-generated content and its limits

Language models produce plausible text, not verified fact. Machine contributions on World Wide can therefore be incomplete, outdated, or wrong — a summary can flatten a nuance, an extracted date can be misread, a classification can miss the field. Treat machine-assisted material as a pointer to the original source, and verify details that matter — deadlines, eligibility rules, scientific claims — against that source. Nothing machine-generated on this platform constitutes professional, legal, medical, or career advice.

Public topic descriptions, machine-produced seminar summaries, and AI-generated field artwork carry a visible “AI-generated” label that is also exposed to assistive technology. Fields pre-filled from an announcement carry no label, because the person posting reviews and edits them before publication. Topic tags suggested by a model are shown on listings without a separate label; embeddings, ranking, and moderation signals are not published. If you find machine-generated material that is wrong, misleading, or not identifiable as such, report it (Section 7) and we will correct or label it.

6.Human oversight

Submitters review pre-filled fields before posting. Automated moderation does not mean that every outcome has already been reviewed by a person: you can contest an outcome and request human review. Editorial corrections can replace machine output. AI-dependent extraction, classification, or enrichment may be delayed or unavailable when a provider fails; the existing public material and account controls remain separate from those processing steps.

7.Questions and reports

Questions about AI on World Wide, reports of incorrect or unlabelled machine-generated content, and objections to AI-assisted processing of your data (see the Privacy Notice, Section 10) all reach us at info@world-wide.org.