How to Detect Deepfake Images, Face Swaps and AI Portraits

What You Will Learn

• How face swaps differ from synthetic portraits and ordinary retouching

• Which facial regions can reveal processing problems

• How to verify identity without relying on appearance alone

• Where automated detection and provenance help

• How to investigate without harming the depicted person

Separate the main types of facial media

Face swap

A face swap replaces or reconstructs an identity region while preserving much of a source image or video. Boundaries near the jaw, temples, ears and hair can carry traces, but modern blending may remove obvious seams.

Fully synthetic face

A generative model creates the portrait rather than starting with a photographed person in that pose. The image may still resemble a real person intentionally or by coincidence. Do not assume that finding no matching identity proves the face is synthetic.

Attribute or expression edit

Software can alter age, emotion, gaze, hair, skin or facial shape. These changes may be cosmetic, satirical, malicious or part of an authorised production. “Edited” describes a process, not intent.

Real portrait with false identity

A scammer may steal an authentic headshot and attach a false name. In this case, pixel-level deepfake analysis misses the central deception. Identity and source verification are decisive.

Europol uses “deepfake” broadly for media manipulated or generated using AI and has documented criminal risks including impersonation, fraud and evidence manipulation. For an image review, record the narrower suspected type so the investigation tests the correct claim.

Inspect the portrait systematically

Facial boundaries

Follow the jawline, cheeks, temples and ears. Look for a change in sharpness, colour, noise or compression where the face meets hair and background. A sharp face pasted into a soft source may reveal blending.

The same effects can come from selective focus, skin retouching, portrait segmentation or image upscaling. Compare several boundary regions and the rest of the file.

Skin tone and texture

Compare forehead, cheeks, chin, neck, ears and hands. A swap may leave a facial tone or pore pattern that does not match adjacent skin. Full synthesis may alternate between smooth areas and highly detailed pores.

Makeup, sun exposure, colour grading and reflected light produce legitimate variation. Skin-based judgments also risk bias across complexions, ages and medical conditions. Describe pixels and lighting; do not judge whether a person “looks real.”

Eye reflections and gaze

Catchlights in both eyes usually relate to the same light sources, modified by eye direction and eyelid coverage. Check whether their shapes and positions are compatible. Then assess whether both eyes converge on the stated subject.

Small images, contact lenses, glasses and editing can hide or alter catchlights. Eye reflections are a clue, not a biometric test.

Hairlines and fine strands

Inspect hair crossing the forehead, ears and background. Warning signs include strands that dissolve, a hairline with a different noise pattern, or background colour bleeding into hair. Phone portrait modes routinely mis-segment these same areas.

Teeth and mouth

Teeth may merge, repeat or lose alignment; lips can meet facial skin with an unusual transition. In a face swap, expression and source pose may cause tension around the mouth. Dental work, motion, shallow focus and compression are common innocent explanations.

Earrings, glasses and other accessories

Compare left and right earrings without assuming they must match. Follow spectacle arms toward the ears and check whether frames cast plausible shadows. Chains should pass consistently in front of or behind clothing. Generators may lose continuity, while a local face edit may leave accessories from the source.

Lighting and head geometry

Identify the light direction on the nose, eye sockets and jaw, then compare the neck, hair and clothing. A face lit from one side on a body lit from another can indicate compositing. Also check whether the face orientation fits the skull and neck.

Multiple lights and studio retouching can produce complex results. A single apparent mismatch is not enough.

Background and repeated portrait patterns

Fully synthetic profile pictures sometimes use a centrally aligned face, shallow-focus background and studio-like lighting. Those are aesthetic choices, not proof. Instead, inspect whether background objects merge into shoulders or whether details repeat with mutated forms.

[ORIGINAL EXAMPLE OR SCREENSHOT TO BE ADDED]

Suggested original asset: three rights-cleared portraits—a genuine retouched headshot, an authorised face-swap demonstration and a fully synthetic face—with identical annotated inspection zones.

Suggested alt text: “Genuine retouched portrait, authorised face swap and synthetic face with matching inspection zones.”

Verify identity outside the image

Start with the claim

Write down the name, role, organization and reason the image was presented. A supposed executive contacting staff requires different evidence from an obviously fictional avatar.

Search the full frame and face crop

Run reverse-image searches on the complete image, then on a crop containing the face and distinctive clothing or background. Search results may reveal:

• The original person under another name

• An older portrait used in unrelated profiles

• A stock photograph or generated-image gallery

• The unaltered source frame

• Earlier, higher-quality copies

Absence from search results is not evidence of synthesis. Private photos, new uploads and unindexed pages may have no match.

Use trusted identity sources

For a public-facing employee, compare the organization’s official staff page, verified communication channels, reputable coverage and prior appearances. Do not trust a link supplied by the suspicious account; navigate independently to the organization.

For a private person, ask for verification that fits the risk without collecting unnecessary sensitive material. A live call can also be manipulated and should not be treated as infallible. Confirm through a known phone number, established contact or second channel.

Compare stable features cautiously

Ear shape, moles, tooth pattern and facial proportions can assist trained examiners, but pose, age, expression and lens perspective change appearance. Consumer side-by-side comparison is not a formal facial-identification result.

Inspect context and chronology

Check whether clothing, weather, event schedules and surroundings fit the alleged time and place. A real face in an old photograph can support a false current claim.

The FBI advises people facing AI-enabled impersonation to verify requests through independently known channels and to resist pressure to act quickly. Identity research often provides safer evidence than trying to infer a generator from pores.

Use detection and provenance carefully

Automated deepfake detection

Deepfake detectors may analyse facial texture, blending boundaries, frequency patterns or learned representations. NIST’s 2025 evaluation work and 2026 challenge focus on operational generalization, adversarial conditions and confidence-scored systems—evidence that laboratory success alone is not enough.

Compression, screenshots, beauty filters, low resolution and new manipulation methods can produce false positives or false negatives. A face-specific model may not be valid for a group photo or illustration.

Submit a non-sensitive copy to the AI Image Detector if its privacy terms and scope fit your use. Record the exact file and regard the output as one measured signal. For a technical explanation, see how AI image detectors work and the safety guide to AI images used in scams.

Metadata

Camera and software fields may indicate a capture or editing path. They may also be stripped, copied or changed. Compare internal consistency rather than searching for one generator name.

Content Credentials

Signed provenance can show the signer and recorded actions, including AI-related source types when declared. It does not prove that a portrait belongs to the claimed person or that the surrounding story is true. C2PA also cautions that provenance may be incomplete and that missing credentials should not automatically reduce trust.

Respect ethical and evidential limits

Minimise harm

Do not repost a humiliating or intimate suspected deepfake merely to ask whether it is real. Preserve a lawful evidential copy, restrict access and use platform reporting or specialist support. Avoid searchable publication of a private person’s face without necessity.

Avoid biased appearance judgments

Claims such as “unnatural skin” can encode assumptions about ethnicity, disability, age, makeup or gender presentation. Use reproducible observations—an edge discontinuity at coordinates or conflicting source history—rather than subjective normality.

Do not confuse authenticity with consent

An authentic photograph may be shared without consent. A consensual synthetic avatar may be legitimate. Pixel origin does not answer privacy, copyright, harassment or impersonation questions.

Escalate high-stakes cases

If the image affects a legal case, financial transfer, hiring decision, access control or public accusation, preserve the original and obtain specialist review. Read the site’s privacy policy before uploading faces and document every processing step.

Final Summary

Deepfake-image detection is strongest when it separates the suspected process—face swap, synthetic portrait, attribute edit or false identity—and tests that specific claim. Facial boundaries, texture, eye highlights, hair, teeth, accessories and lighting can guide inspection, but each has ordinary alternative explanations.

Move beyond the face. Reverse-search the image, verify the identity through independent channels, compare trusted sources, inspect metadata and signed provenance, and add a detector score only as supporting evidence. A careful finding may be “identity unverified” even when the pixels cannot be classified. Use the AI image detection guides and seek expert assistance before a conclusion could affect someone’s safety, reputation or rights.