03 · What You Need to Know
Managing Imminent Harm Disclosures During AI-Mediated Research
What Does Imminent Harm Mean in a Research Interview?
Imminent harm generally refers to a serious threat that may occur in the immediate or near future. In research interviews, concerns may involve possible self-harm, harm to another person, or another urgent threat to safety.
There is no single definition that governs every research institution or jurisdiction. The threshold for mandatory action may depend on applicable law, professional obligations, the participant's circumstances, and the approved research protocol.
Researchers should distinguish expressions of ordinary distress from disclosures that suggest an immediate safety concern. A participant saying, "This question makes me uncomfortable," presents a different situation from someone describing a specific intention to cause serious harm in the near future.
Nevertheless, an AI system may not reliably determine the seriousness of ambiguous statements. It should therefore use conservative, predefined escalation criteria rather than independently deciding that a concerning disclosure is harmless.
Why Is an AI Interviewer Different From a Human Interviewer in an Emergency?
A trained human interviewer may recognize concerning statements, ask appropriate safety-related questions within their competence, and contact designated personnel according to the research protocol.
An AI chatbot can recognize certain linguistic patterns and generate responses that resemble supportive conversation. However, conversational fluency does not establish clinical competence, legal judgment, or reliable situational awareness.
The chatbot may misunderstand figurative language, overlook an indirect warning, or respond inappropriately to a serious disclosure. It may also lack the participant's location, identity, or an operational means of contacting assistance.
These limitations make participant distress and safety risks during AI interviews particularly consequential when a conversation moves beyond the intended research topic.
The appropriate response is not to assume that AI can replace emergency judgment, but to establish a reliable route from automated interaction to qualified human assistance.
What Should the AI System Do Immediately?
A defensible safety protocol should specify how the system responds when an interaction meets predefined criteria for possible imminent harm.
1. Stop ordinary research questioning
The chatbot suspends the interview rather than continuing to collect information about the research topic.
2. Provide an approved safety message
The system communicates that the disclosure raises a safety concern and directs the participant toward the assistance specified in the approved protocol.
3. Activate the designated escalation mechanism
Where an operational human-response arrangement exists, the system triggers the relevant notification or handoff procedure.
4. Avoid unsupported assurances
The chatbot does not claim that a person has been notified, that emergency services are responding, or that the participant is safe unless those statements have been verified.
5. Preserve appropriate incident information
The research team retains the minimum information necessary for the approved response, documentation, and any required review.
6. Determine whether research activities may resume
Authorized human personnel assess whether the incident requires further action, changes to the protocol, or suspension of the interviewing procedure.
This sequence is an illustrative design framework, not a universal emergency algorithm. The actual procedure must be developed for the study's setting and tested before participants are exposed to foreseeable risks.
Can an AI Chatbot Determine Whether a Threat Is Credible?
AI systems can identify some explicit statements associated with immediate danger, but detecting concerning language is not the same as assessing actual risk.
Statements may be ambiguous, sarcastic, incomplete, culturally specific, or dependent on context that the chatbot cannot access. A system might incorrectly classify a figurative expression as an emergency or overlook a serious disclosure expressed indirectly.
False positives can create unnecessary distress and inappropriate disclosure of private information. False negatives may delay necessary assistance.
For this reason, automated classification should not be treated as a substitute for qualified assessment. The research team should define when human review is required and how uncertainty will be handled.
Where immediate intervention may be necessary, the system should not depend exclusively on a single generative model's judgment that the situation is safe.
Does Research Confidentiality Prevent Researchers From Acting?
Not necessarily. Confidentiality is an important research obligation, but its limits depend on applicable law, ethical requirements, institutional policies, and the commitments made during informed consent.
Some research settings may require or permit disclosure of limited information when there is a serious and immediate threat to safety. Other settings may impose different thresholds or protections.
Researchers should not assume that every disclosure of possible harm automatically authorizes unrestricted sharing of participant information. Nor should they promise absolute confidentiality when foreseeable circumstances could require an exception.
In the United States, the Common Rule requires informed consent to describe the extent to which confidentiality of records identifying participants will be maintained. This is relevant to explaining foreseeable limits, although specific reporting or intervention duties arise from other applicable legal and professional requirements.
For studies conducted in the Philippines, researchers should consider applicable institutional research ethics requirements, the Data Privacy Act of 2012, and relevant legal or professional obligations. The Data Privacy Act should not be interpreted as creating a blanket authorization to disclose research information whenever an AI system identifies a possible risk.
Where the appropriate legal response is uncertain, the research team should establish the procedure with its ethics committee, data protection personnel, and qualified legal or clinical advisers before deployment.
What If the AI Interview Is Anonymous?
Anonymous participation can reduce certain privacy risks, but it may also limit the research team's ability to intervene when an urgent disclosure occurs.
For example, a chatbot may receive a statement indicating immediate danger without knowing the participant's name, physical location, or reliable contact information.
Researchers should distinguish anonymous data collection from confidential data collection. In anonymous research, the team may have no practical way to identify or contact the participant. In confidential research, identifying information may be available but subject to access restrictions.
Anonymous AI interview
The research team does not collect information that reasonably permits participant identification. Individual follow-up or emergency intervention may be impossible.
Confidential AI interview
The research team may hold identifiable information under controlled access. Human follow-up may be possible, subject to the approved protocol and applicable requirements.
Anonymity should not be represented as an absolute technical guarantee if the platform collects identifiers, IP addresses, or other information that may permit identification.
Researchers must decide in advance whether the study's foreseeable risks are compatible with the available contact and escalation arrangements. Collecting additional identifying information solely for hypothetical emergencies also requires a defensible justification.
What If Nobody Is Monitoring the Chatbot in Real Time?
This is a particularly important problem for asynchronous AI interviewing.
A participant may complete an interview at night or during a period when no research personnel are available. If the system flags a serious disclosure but no one receives or acts on the notification promptly, the existence of an escalation feature may create a misleading impression of protection.
Researchers should distinguish between automated detection, notification delivery, and actual human response. These are separate operational functions.
A system that sends an email to an unmonitored inbox does not provide immediate intervention. Similarly, a chatbot that displays emergency contact information cannot guarantee that the participant receives assistance.
Watch Out
Do not tell participants that disclosures will be monitored or that urgent assistance will be arranged unless the research team has an operational mechanism capable of providing the stated service. A notification feature is not equivalent to a staffed emergency response.
When reliable human escalation cannot be provided, researchers should reconsider whether autonomous interviewing is appropriate for a population or topic where imminent-risk disclosures are reasonably foreseeable.
Should the Chatbot Ask Additional Questions About the Threat?
Ordinary research probing should stop once a disclosure meets the protocol's criteria for possible imminent harm.
Any subsequent questions should be limited to what the approved safety procedure permits and should not become an improvised clinical assessment conducted by a general-purpose language model.
For example, a chatbot should not continue asking detailed questions about traumatic experiences merely because those responses could be analytically useful.
If the study requires immediate risk assessment, the protocol should specify who is qualified to perform it and how the participant can reach that person.
This distinction is essential because the objectives of research interviewing and emergency response are different. A procedure optimized for collecting detailed narratives may be unsuitable for managing an urgent safety concern.
What Information Should Be Recorded After an Incident?
Documentation may be necessary to determine what occurred, whether the protocol functioned as intended, and whether institutional reporting obligations apply.
Depending on the approved procedure, relevant information may include the triggering statement, the chatbot's response, the time of the interaction, the escalation action attempted, and any confirmed human follow-up.
However, researchers should avoid retaining unnecessary sensitive information. Access should be restricted to authorized personnel, and retention should follow the applicable protocol and data protection requirements.
Appropriate AI-participant interaction records can support accountability, but comprehensive logging is not automatically preferable to proportionate documentation.
Does Every Disclosure Require Reporting to an Ethics Committee?
No universal rule requires every concerning statement to be reported in the same manner.
The U.S. Office for Human Research Protections distinguishes adverse events from unanticipated problems involving risks to participants or others. Whether an incident meets reporting criteria depends on factors such as its relationship to the research, whether it was anticipated, and whether it indicates increased risk.
Other jurisdictions and institutions may use different reporting frameworks or impose additional requirements.
Researchers should follow their approved protocol and institutional procedures rather than allowing the chatbot to decide independently whether an event is reportable.
Serious failures in the automated response mechanism may themselves warrant review, even when the research team cannot determine whether actual harm occurred.
