01 · The Question
Can You Upload Photographs or Other Participant Images to an AI Tool?
AI systems can describe photographs, classify visual material, detect objects, extract text, analyze medical images, organize large image collections, and perform other tasks that may be useful in research. If your study contains hundreds or thousands of images, sending them to an AI system can appear to be an efficient analytical shortcut.
Images involving research participants deserve particular care. A photograph can reveal a face, body, disability, health condition, location, workplace, home, identification card, computer screen, vehicle registration plate, or another person who never participated in the study. Information embedded in or associated with the file may create additional identification risks.
The question is therefore not simply whether an AI system can analyze an image. It is whether that particular image may be disclosed to that particular system for that particular research purpose.
03 · What You Need to Know
An Image Can Reveal Much More Than a Participant's Face
Participant Images Are Data, Not Merely Illustrations
When photographs, screenshots, scans, video frames, medical images, or other visual materials are collected for research, they can constitute research data. If an external AI service receives those files, the researcher is making a data-processing decision rather than simply using an image-editing feature.
The significance depends on the image. A photograph of an experimental apparatus with no people or identifying information presents a very different problem from a photograph of a participant's face, home, clinical condition, classroom, workplace, or personal documents.
The first step is therefore to inspect what information the image actually communicates.
A Face Is an Obvious Identifier, but It Is Not the Only One
Faces can make people immediately recognizable. In some circumstances, technical processing of facial characteristics may also constitute biometric processing under applicable data-protection law.
But a participant can remain identifiable when the face is cropped or blurred. Relevant visual clues may include:
- tattoos, scars, birthmarks, or other distinctive physical characteristics;
- uniforms, name badges, identification cards, or specialized clothing;
- a recognizable home, office, classroom, hospital room, or workplace;
- vehicle registration plates;
- certificates, signs, computer screens, mail, or documents visible in the background;
- other people appearing in the image;
- rare equipment, possessions, or surroundings associated with a particular person; and
- a distinctive event or setting that can be connected with public information.
De-identification therefore requires examining the entire visual scene rather than placing a rectangle over the participant's eyes and declaring victory.
A Photograph Is Not Automatically Biometric Data, but AI Processing Can Change the Analysis
Privacy terminology matters here. Under UK data-protection guidance, for example, an ordinary digital photograph is not automatically biometric data merely because it depicts an identifiable person. Biometric data arise when specific technical processing of physical, physiological, or behavioral characteristics is used to uniquely identify someone, such as creating and matching a facial template.
This distinction is useful because “contains a face” and “is being used for biometric identification” are not synonymous. An AI system used to count people in research photographs presents a different processing purpose from a system used to recognize which participant appears in each photograph.
Researchers should therefore establish what the AI system actually extracts or infers from the image, not merely whether the input file happens to be a photograph.
Some Legal Frameworks Explicitly Treat Faces as Identifying Information
In U.S. health-data contexts governed by HIPAA, the distinction becomes particularly concrete. The HIPAA Safe Harbor de-identification method includes full-face photographic images and comparable images among the identifiers that must be removed. It also includes biometric identifiers such as fingerprints and voice prints.
That does not mean HIPAA's identifier list is a universal anonymization rule for every photograph or jurisdiction. It illustrates why visual data cannot be treated as anonymous simply because textual identifiers have been deleted.
If the image forms part of clinical research, the broader requirements governing clinical research data processed with AI may also apply.
Image Metadata Can Matter Even When It Is Invisible on Screen
What you see when opening an image is not necessarily everything contained in or associated with the file. Depending on how an image was produced, stored, exported, or managed, associated information may include filenames, dates, device-related information, descriptive labels, study identifiers, or other metadata.
Researchers should therefore assess the actual file and workflow rather than assuming that cropping visible information necessarily removes every identifying element.
The safest approach is not to guess which metadata exist. Determine what information will actually be transmitted by the workflow you intend to use and remove unnecessary identifying information before any authorized external processing.
The Background Can Be More Identifying Than the Participant
Imagine a participant photographed from behind so that no face is visible. The wall behind the participant displays the institution's logo, an employee roster, a departmental notice containing names, and a certificate identifying the room's occupant.
Removing the face would accomplish very little.
Visual de-identification therefore requires contextual judgment similar to that required for qualitative text. The relevant question is whether information remaining in the image could reasonably connect it to a person, not merely whether a conventional facial portrait remains visible.
Images Can Reveal Sensitive Characteristics
An image may disclose considerably more than identity. Depending on context, it may reveal health conditions, disability, religious practice, ethnicity, living conditions, political activity, relationships, or other potentially sensitive characteristics.
AI may also be used to infer characteristics that the researcher did not originally collect explicitly. SACHRP has warned more broadly that AI and large datasets can generate or infer private information and that research involving AI can create privacy, discrimination, and group-level harms that conventional approaches to identifiability may not fully capture.
An image should therefore be assessed both for what it visibly contains and for what the proposed AI processing is intended to derive from it.
Other People in the Image Matter Too
A photograph collected from one participant can contain several non-participants. A classroom image may show students. A photograph of a participant at home may include children or relatives. A workplace photograph may show colleagues.
The fact that the principal participant consented to the research does not automatically establish permission to disclose identifiable information about everyone visible in the image.
This issue is particularly important when images were collected incidentally rather than through a carefully staged research photography protocol.
Consent to Be Photographed Is Not Necessarily Consent to AI Processing
A participant may have consented to photography for documentation, qualitative analysis, publication, clinical assessment, or another defined purpose. Whether that authorization extends to uploading the image to an external AI system depends on what participants were told and the rules governing the study.
For prospective studies, researchers should consider whether informed consent needs to describe AI processing of participant data. For an existing study, a separate assessment may be necessary if the AI use was introduced only after images had already been collected.
Image Transformation Can Reduce Risk but May Also Change the Research Data
Cropping, masking, blurring, pixelation, generalization, or removing metadata may reduce identification risk. These transformations can also remove information relevant to the research.
If facial expression is the phenomenon being analyzed, for example, blurring the face defeats the purpose. If the research concerns classroom interactions, cropping everyone except one participant may remove the context required for interpretation.
Privacy protection therefore needs to be designed around the research question. Sometimes transformation is sufficient. Sometimes the analytical purpose requires identifiable imagery, in which case the solution is not pretend-anonymization but a properly authorized and protected processing environment.
Watch Out
Do not upload an identifiable image to an unauthorized AI tool merely to ask the tool to anonymize it. The original identifiable image must first be transmitted for the AI system to perform the transformation. An anonymized output does not undo an unauthorized disclosure of the original input.
The AI Provider's Conditions Still Need to Be Verified
If an external AI service will receive participant images, researchers may need to determine how submitted files are retained, accessed, secured, deleted, or used for service improvement or model training, as well as whether subprocessors or cross-border processing are involved.
The answers may differ among consumer products, institutional accounts, enterprise services, APIs, and locally deployed systems. Provider safeguards should be evaluated alongside, rather than substituted for, ethics approval, participant consent, institutional policy, and applicable law.
06 · What This Means for You
Inspect the Whole Visual Record Before AI Processing
Do not reduce visual privacy review to face detection. Examine the participant, the background, other people, visible text, distinctive objects, and associated file information. Then ask which of those elements the AI genuinely needs.
In many cases, researchers can create a purpose-specific derivative image for AI processing while retaining the protected original inside the approved research environment.
A simple decision framework
If the AI task does not require identifiable participant imagery
Crop, mask, generalize, or otherwise remove unnecessary identifying information before any authorized external processing.
If identifying visual characteristics are analytically necessary
Use only an environment specifically authorized for identifiable research data rather than pretending the image is anonymous.
If other people appear in the image
Assess their identifiability and whether their information is necessary and permissible to process.
If AI will perform facial recognition or another biometric identification task
Assess the additional legal, ethical, and institutional requirements applicable to biometric processing.
If you cannot establish what information the AI provider receives or retains
Do not upload participant images until the uncertainty is resolved.
When visual data include synchronized speech or moving images, the problem becomes broader still. The safeguards required for participant audio and video recordings processed with AI need to account for voices, spoken information, movement, and temporal context as well as visible identity.
07 · A Quick Checklist
Before Uploading Participant Images to AI
Before any participant image is transmitted, check:
Determine whether faces or other distinctive physical characteristics identify participants.
Inspect backgrounds for names, signs, locations, screens, documents, registration plates, institutional identifiers, and other contextual clues.
Identify non-participants who appear in the image and assess whether their information needs to be included.
Check associated filenames, metadata, labels, and linked study information for unnecessary identifiers.
Determine whether cropping, masking, generalization, or another transformation can reduce disclosure without undermining the research purpose.
Verify whether the AI performs biometric identification, sensitive-attribute inference, or another processing activity requiring additional scrutiny.
Confirm that consent, ethics approval, institutional policy, and applicable privacy requirements permit the proposed processing.
Verify retention, training, access, security, deletion, subprocessors, and processing conditions for the exact AI service being used.
