AI-Moderated Diary Studies: Build a Living Participant Contract

A seven-day study can become a different study by day three

Picture a fictional seven-day diary study about how small-business owners handle unexpected payment failures. At enrolment, participants agree to one short text entry each evening. The study will examine workflow disruption, not financial circumstances. A researcher will review entries every morning. Participants may skip any prompt and leave the study through a visible control.

On day three, several entries mention screenshots. The study team enables image upload so participants can show error messages. The AI moderator begins asking follow-up questions when it detects account details, unfamiliar terminology, or signs of frustration. On day five, a participant uploads a photograph that includes a colleague, a customer name, and part of a bank notification. Another describes severe personal financial pressure and asks whether the research team can help.

The original information was accurate when the study began. It no longer describes the participant's experience.

The capture mode changed from text to images. The burden changed from one response to a branching conversation. The data risk changed because screenshots and photographs can expose identifiers, bystanders, locations, notifications, and background details. The expectation of researcher involvement changed when an automated prompt responded immediately and the participant interpreted that response as human attention.

None of these shifts is solved by adding a longer consent form on day one. The team needs an operating record that remains visible to the participant and to the people running the study.

That record is the Living Participant Contract. It is an editorial governance model for planning and monitoring a longitudinal study. It does not replace informed-consent review, privacy assessment, safeguarding procedures, accessibility work, legal advice, or the judgement of a qualified researcher. Its job is narrower: keep the current terms of participation legible as the study evolves, and prevent silent protocol drift from becoming normal operations.

Longitudinal research changes the relationship, not only the sample window

A diary study distributes research activity across time and context. Participants record events near the moment they occur, then return to the study repeatedly. That can reveal change, sequence, workarounds, and context that a retrospective interview may miss.

It also creates cumulative demands. A five-minute prompt repeated twice a day is not experienced as a single five-minute task. Reminders arrive during work, care, travel, rest, or stress. A prompt that felt neutral at enrolment can become sensitive after circumstances change. The same image request can be harmless in a private room and risky in a workplace, clinic, school, or shared home.

The 2022 Family Practice methods brief makes the trade-off concrete. Its authors describe diary-interview research as useful for longitudinal depth, while identifying commitment from participants and researchers as a significant requirement. They also note that increased frequency increases burden and can limit study duration. Their example came from 16 Canadian clinicians during the pandemic, so it cannot tell a product team what cadence will work for every population. It does establish that cadence and burden are methodological decisions, not notification settings.

AI moderation adds another layer. A system can react before a researcher reviews the entry. It may ask for clarification, branch into a sensitive topic, interpret an image, or classify a disclosure for escalation. Current vendor material confirms that such capabilities are being marketed in 2026. The claims come from the vendor itself, so teams should test performance, failure modes, accessibility, and participant understanding in their own context instead of treating availability as validation.

The National Academies workshop on longitudinal studies of aging offers a useful principle: consent over time involves an ongoing relationship of information and choice. The setting was aging research, with issues that may differ substantially from UX work. Still, the relational point travels. When a study changes what it asks, collects, infers, or promises, the participant needs information that can support a fresh choice.

Seven fields that should stay current

The Living Participant Contract has seven fields. Each field should be written in participant language, mapped to an accountable owner, and versioned when it changes.

1. Purpose. State the decision the study is intended to inform and what falls outside that purpose. “Understand payment-failure workflows to decide which recovery guidance to redesign” is more useful than “improve the experience.” If the team later wants to study personal financial resilience, that is a new purpose requiring review.
2. Cadence and burden. Show expected frequency, response window, approximate effort, study duration, quiet periods, reminder behaviour, and incentive terms. Track actual effort as the study runs. A nominal five-minute entry that regularly triggers fifteen minutes of follow-up has exceeded its stated burden.
3. Modality. List text, audio, image, video, screen recording, location, sensor, or other inputs separately. Explain what each format may capture accidentally. Give participants a less intrusive alternative where the research question permits it.
4. Adaptation rules. Explain what automation may change. Define allowed follow-ups, prohibited topics, maximum depth, quiet hours, stop conditions, and changes that require researcher approval. “The moderator may ask one clarification about the workflow” is inspectable. “Questions may adapt to your answers” leaves the boundary invisible.
5. Data use and access. Describe what is collected, why, who can access raw and derived material, which AI services are involved, where processing occurs, how long material is retained, and which outputs may be shared. Verify these statements against the actual system and contracts.
6. Escalation and safety. Tell participants when automation stops, when a researcher may intervene, what response time is realistic, and what support is unavailable. A research channel should not imply crisis, legal, clinical, financial, or employment support that the team cannot provide.
7. Withdrawal and repair. Make skipping, pausing, correcting, revoking a modality, leaving the study, and requesting deletion accessible. State any lawful or technical limits before participation. Do not wait for a withdrawal request to discover that identifiers have been copied into local notes, exports, or stakeholder decks.

These fields turn a broad promise into observable conditions. They also expose ownership. Research may own cadence and prompts. Privacy may own approved purposes and data handling. Security may own access and transfer controls. Product may own the decision. A safeguarding or domain specialist may own escalation. One person can hold several roles in a small team, but each judgement still needs a name and a response path.

Three kinds of change need three different responses

Repeated consent prompts can create noise without improving agency. Silent changes create the opposite risk. The team needs a practical distinction among expected variation, a material protocol change, and a safety escalation.

The categories depend on context. Switching from text to audio may be a minor convenience in one low-risk study and a material privacy or accessibility change in another. Extending a response window may reduce burden, but changing the timing could affect incentive eligibility or comparability. The research and governance owners should classify foreseeable changes before fieldwork, then review unexpected ones rather than letting the tool decide its own authority.

Show the participant what changed

Return to the payment-failure study. The team wants to enable screenshots on day three. A vague message would say, “You can now add images.” A useful change notice answers six questions:

• What changed: image upload is now available for error messages.
• Why: the team needs to understand which message and screen state participants encountered.
• What an image may reveal: account names, balances, notifications, people, locations, device details, or unrelated applications.
• What happens to the image: approved access, analysis, storage, retention, and sharing conditions.
• What choice exists: continue with text, upload a cropped or redacted image, pause, or leave according to the study terms.
• When the change applies: only after the participant chooses.

The participant should not have to compare two policy documents. Present a short diff, preserve access to the full current terms, and record the version accepted. If image submission is necessary for the research question and no lower-data option works, say so. Do not present a non-existent alternative to make the interface look considerate.

The same discipline applies to incentives. If the team adds a bonus for completing every day, the reward may alter pressure to continue or discourage skipping a sensitive prompt. The change notice should explain eligibility and preserve the stated right to pause or withdraw. Reviewers should assess whether the new incentive changes voluntariness for the population involved.

Bound adaptation before the AI meets a participant

An adaptive moderator needs a defined action space. Start with the approved topic, allowed evidence types, maximum follow-up depth, prohibited inferences, sensitive-topic triggers, stop conditions, response-time promises, and handoff owner.

For the fictional study, the AI may ask one clarification about the sequence of actions, the error message, or the participant's next step. It may not ask for account balances, debt, health, immigration status, family circumstances, credentials, or names. It must not infer a participant's emotional state from an image or voice and present that inference as fact. It stops if an entry contains an unexpected person, a credential, a request for professional help, or content covered by the escalation policy.

Test the boundary with adversarial and ordinary cases before fieldwork. Include a short answer, a long narrative, sarcasm, a second language, assistive-technology output, a blurred screenshot, a photo with bystanders, a refusal, and an off-topic disclosure. Review both the prompt selected and the prompt withheld. A safe-looking transcript can hide a failure to ask a necessary clarification, while a polished follow-up can cross the approved boundary.

Evaluate every AI-assisted transformation separately. Fred's guide to AI research quality describes source integrity, capture fidelity, meaning preservation, calibration, and decision traceability as distinct checkpoints. In a diary study, that means reviewing moderation, transcription, image interpretation, classification, synthesis, and recommendation as different components with different error paths.

Protocol adaptation can damage comparability

Adaptive prompts may improve an individual entry while making the study harder to interpret across participants. If one person receives a single neutral clarification and another receives ten leading probes, their responses were produced under different conditions. A synthesis system may merge them into one theme and erase the difference.

Keep the protocol version attached to every entry and derived claim. Record the original prompt, automated follow-up, reason or trigger, participant response, researcher intervention, modality, and contract state. During analysis, compare like with like before combining evidence. If a protocol change splits the study into meaningful phases, report the phases and their sample limits explicitly.

Do not treat comparability as a reason to suppress useful adaptation. Treat it as a reason to expose adaptation. The decision owner should be able to see whether a finding appeared across the common protocol, only after an image prompt was introduced, or only among participants who accepted a higher-burden branch.

Multimodal capture brings context and collateral data

Photos, screen recordings, audio, and video can reveal context that text misses. They can also collect more than the research question needs. Outset's March 2026 vendor announcement claims that its system can analyse screens, facial expressions, photos, and live video. Whether or not a team uses that product, the marketed capability illustrates the governance issue: the analysis surface can expand beyond what the participant consciously meant to submit.

Before enabling a modality, list likely collateral data. A screenshot may contain names, balances, tabs, notifications, or internal URLs. Audio may capture colleagues or family members. A photograph may reveal an address, location, health information, religion, political material, or a child's image. Video may show a workspace, access badge, customer record, or screen reflected in glass.

Then reduce the collection. Offer crop and preview, explain how to remove unrelated details, block known credential patterns where feasible, and let the participant replace the submission. Prefer a description when it can answer the question. Restrict raw-media access and decide whether summaries can retain the insight without retaining the file.

For teams within GDPR scope, the official regulation includes purpose limitation and data minimisation among its core principles. Application to a particular study requires qualified review. Fred's GDPR-compliant user research guide provides an operational checklist for consent records, access, retention, deletion, and data flows, but an article cannot establish the lawful basis or obligations for a specific organisation.

Burden includes interpretation, recovery, and accessibility

Time-on-task is only one part of participant burden. People also spend effort deciding what a prompt means, whether a response is safe, how much context to remove, whether silence affects payment, and how to recover from an accidental upload.

Monitor actual response time, follow-up count, skipped prompts, delayed entries, pauses, corrections, help requests, and withdrawals. Treat sudden compliance as a possible burden signal, not automatically as engagement. A participant who gives shorter answers after repeated probing may be tired, uncertain, protecting privacy, or simply finished with the topic.

Accessibility choices must remain active throughout the study. A participant may need text instead of audio, more time, fewer reminders, keyboard operation, screen-reader compatibility, captions, plain language, or help from an assistive tool. If the AI treats these patterns as low quality, fraud, or noncompliance, the protocol is excluding valid evidence.

Ask what adaptation makes participation possible without changing the research claim beyond recognition. Record the accommodation separately from the participant's substantive response and restrict access to any sensitive explanation. When the method cannot support a participant fairly, report that exclusion as a study limitation rather than allowing the missing population to disappear from the decision.

Withdrawal needs an operational map

“You can withdraw at any time” is incomplete unless the team knows what withdrawal does.

Map every place participant material can travel: recruitment system, consent record, diary platform, raw-media storage, transcript, AI service, analysis workspace, local export, repository, report, presentation, and decision record. For each location, define the identifier, owner, access, retention, correction path, deletion path, downstream effect, and any justified limit.

Participants should be able to skip an entry without negotiating with the moderator. Pausing should stop reminders and automated follow-up. Correcting a diary entry should preserve an appropriate audit record without continuing to use the superseded content as if it were current. Revoking one modality should not silently remove the person from every part of the study unless that consequence was explained and is necessary.

Do not promise deletion that the system cannot deliver, and do not use technical complexity as a surprise exception. Resolve the workflow before recruitment with the appropriate privacy, legal, security, and ethics owners.

Keep the participant choice connected to the evidence

A final report often contains a polished theme without the protocol history that shaped it. That makes review difficult. A decision owner cannot tell whether the finding came from the original text-only study, a later image branch, a participant who declined the new modality, or an entry collected after the burden exceeded the stated range.

For each decision-relevant claim, retain the study purpose, protocol version, participant contract state, modality, prompt path, source entry, transformation history, reviewer corrections, contradictions, population limits, and permitted decision use. Separate raw sensitive media from the smaller set of derived material needed for the decision.

A searchable research repository can keep studies, findings, source context, and report history available for later review. Reuse still needs a purpose, permission, freshness, and population check. A prior participant's agreement to one diary study does not automatically authorize every future use of the material.

A preflight and daily monitoring routine

Before launch, the study owner should be able to answer:

• Which product or business decision will the diary evidence inform?
• Why does the question require repeated in-context entries?
• What are the seven current contract fields, and who owns each one?
• Which protocol changes are foreseeable, material, prohibited, or safety-related?
• What can the AI ask, infer, classify, and escalate?
• Which modalities are necessary, and what collateral data could they capture?
• What is the lower-data or accessible alternative?
• How are actual burden and protocol divergence monitored?
• How do skipping, pausing, correction, withdrawal, and deletion work end to end?
• Which privacy, security, accessibility, domain, legal, or ethics reviews are required?

During fieldwork, review four things each day: participant burden, contract changes, automated exceptions, and evidence comparability. Record the decision and owner when action is required. Stop a prompt or modality when the team cannot verify that it remains within the approved boundary.

At synthesis, do not collapse protocol versions into a single clean narrative. Show which evidence supports the claim, which participants encountered a different study path, which contradictions remain, and which decision the evidence can reasonably inform.

Then bring the decision into Fred's decision-intelligence workflow. Connect the question, evidence, method, limitations, contradictions, synthesis, and recommendation so the team can inspect the case before acting. The AI moderator may help the study respond in the moment. Accountability remains with the people who design the relationship, review its changes, and decide what the evidence is allowed to change.

Source notes

• Outset, “Top Diary Study Software Tools for Research in 2026”, published 3 June 2026 and opened 17 August 2026. Used only as a current vendor signal that AI follow-up, multimodal capture, continuous synthesis, and researcher configuration are being marketed for diary studies. Capability, ranking, scale, compliance, and effectiveness claims remain vendor claims and are not independently validated here.
• Aaron Cannon, Outset, “Introducing: Visual Intelligence”, published 31 March 2026, updated August 2026, and opened 17 August 2026. Used as a vendor signal for claimed analysis of screens, photos, facial expressions, and live video. The article does not establish accuracy, ethical sufficiency, participant understanding, or generalisability.
• Patricia Thille, Louise Chartrand, and Cara Brown, “Diary-interview studies: longitudinal, flexible qualitative research design”, published 26 April 2022 in Family Practice and opened 17 August 2026. Used for peer-reviewed methodological context on longitudinal depth, timing, participant commitment, burden, prompt design, and submitted diaries. Its worked study involved 16 Canadian primary-care clinicians during the pandemic, so lessons are transferred cautiously to UX research.
• National Academies of Sciences, Engineering, and Medicine, “Informed Consent”, in Improving Consent and Response in Longitudinal Studies of Aging: Proceedings of a Workshop, published 25 March 2022 and opened 17 August 2026. Used for expert discussion of consent as an ongoing relationship involving information, choice, burden, incentives, withdrawal, and changing conditions. Workshop proceedings focused on longitudinal studies of aging and do not validate this article's interface or governance recommendations.
• European Union, Regulation (EU) 2016/679, Article 5, consolidated text opened 17 August 2026. Used for the official principles of purpose limitation and data minimisation. This article does not determine lawful basis, controller or processor roles, consent validity, data-subject rights, transfer requirements, or the legal obligations of a specific study.
• The Living Participant Contract, its seven fields, change taxonomy, participant-facing diff, protocol-to-evidence provenance, fictional payment-failure study, preflight questions, and daily monitoring routine are original Fred editorial contributions. They are operational decision aids, not validated scientific instruments, legal advice, ethics approval, safeguarding procedures, or claims about an autonomous Fred feature.