Digital Research and Artificial Intelligence

This policy covers online, telephone, and platform research; digital questionnaires and data; and AI-assisted transcription, translation, coding, analysis, summarization, writing, visualization, and programming.

• Legitimate and disclosed purpose.
• Minimum necessary data.
• Tools proportionate to sensitivity.
• Human verification of inputs and outputs.
• Disclosure of material automation.
• Testing for bias, error, and uncertainty.
• Professional responsibility remains with the researcher.

• Identify the organization, purpose, and consent.
• Use proportionate duplicate and bot controls.
• Assess internet, device, and digital-coverage bias.
• Design for mobile use and accessibility.
• Collect device or location data only where necessary.
• Offer alternatives where digital modes exclude important groups.

Public availability does not make every use lawful or ethical. Platform terms, user expectations, sensitivity, identifiability, and harm are assessed. Names, images, or individual posts are not republished merely because they are accessible.

• Identifying or restricted datasets.
• Participant names and contact details.
• Sensitive recordings or images.
• Complaints and safeguarding disclosures.
• Confidential drafts or commissioner documents.
• Access keys and passwords.
• Data or instruments without processing rights.

• Provider retention and duration.
• Whether inputs are used for training.
• Processing locations and subprocessors.
• Deletion and export controls.
• Encryption, permissions, and logs.
• Contract and intellectual-property terms.
• Ability to discontinue or replace the service.

• Match quotations and references to sources.
• Recalculate figures.
• Test code against known cases.
• Review translation and terminology.
• Measure classification accuracy through human validation.
• Assess bias across languages and groups.
• Record material version and settings.

Digital or AI use is disclosed where it materially affects collection, coding, analysis, or published content. Routine spelling assistance need not be listed, but readers receive what they need to evaluate methodological integrity.

Generated text is not presented as participant testimony, a real quotation, or field evidence. Synthetic data are distinguished from real data and are not used to inflate samples or fill values without valid method and disclosure.

An automated tool alone does not determine participant acceptance or exclusion, accuse a researcher, reject a complaint, or expose identity. Human review and an explainable, reviewable decision are required.

Inappropriate upload, leakage, identifying output, or unauthorized tool use is reported, contained, assessed, deleted where possible, and used to improve controls.

Tools and terms change rapidly. Approved-tool lists, contracts, risk, and performance are reviewed; previous approval is not treated as permanent.

• Make a final consequential decision about a participant or worker using a model alone.
• Enter identifiable, restricted, or client-confidential information into an unapproved public tool.
• Create fictional participants, interviews, or quotations and present them as data.
• Infer sensitive attributes a person did not agree to provide.
• Publish automated text as scientific review or an institutional decision without accountable human approval.
• Use a provider with unknown terms for a high-risk task.

Where use affects instruments, data, analysis, or narrative, the service and version or date, purpose, permitted inputs, human review, tests, and outcome are recorded. Simple spelling assistance need not be disclosed unless it has a material effect.

• Storage and processing location and retention.
• Whether inputs are used to train or improve the model.
• Subprocessor access.
• Encryption, access control, and deletion.
• Ability to disable logging or training.
• Intellectual-property and output terms.
• Performance in Arabic and the Palestinian context.

Outputs are tested across groups, languages, and examples to identify higher error for a dialect, name, area, or group. Removing a sensitive variable does not resolve bias where proxy variables still reveal it.

Synthetic data may support testing or training where they are not presented as real observations and do not permit reconstruction of sensitive source data. They are clearly separated from participant data and their role is explained.

Platforms or algorithms should not target people with a sensitive topic without reviewing privacy, consent, and bias. The project must understand who sees the invitation and what the platform infers from interaction.

• Review lawfulness, platform terms, and reasonable privacy expectations.
• Minimize collection of names, identifiers, and irrelevant content.
• Do not treat public posting as consent for every reuse context.
• Assess searchable quotation and reidentification risk.
• Document date, query, collection method, and coverage limitations.

1. Compare output with the original source or data.
2. Test calculations or code on known examples.
3. Look for fabricated quotations and references.
4. Review categorical, causal, and generalized language.
5. Review Arabic errors, direction, and translation.
6. Require accountable human approval of the final version.

Material use of AI in generation, translation, classification, or analysis is disclosed with the function and human review described. A system is not listed as an author because responsibility remains with people and the institution.

Where a tool exposes data, produces erroneous analysis, or changes behavior after an update, the affected function is paused, evidence preserved, prior outputs assessed, necessary parties informed, and material or process corrected.

Tools and terms change rapidly, so one-time approval is insufficient. Services are reviewed after a material update, policy change, use of more sensitive data, an incident, or availability of a less intrusive alternative.