Where Deepfakes Still Break in 2026

Face replacement has become extremely good. The failure modes that defined early deepfakes (flickering, obvious blur, mismatched skin tone) have largely been solved. What remains is subtler, but it has not disappeared.

The blend boundary

A face swap must composite a generated face onto an existing head. Somewhere there is a boundary, and the boundary must be blended. Look for:

• A soft transition band around the jawline, hairline and ears where detail level changes.
• Texture discontinuity: pores and fine skin detail present on the forehead but absent on the cheek, or vice versa.
• Colour grading that stops at an invisible line, most visible on the neck.

In ELA these boundaries frequently appear as a closed contour following the shape of a face, which natural scenes essentially never produce. reading the heat map takes some practice.

Ears, teeth and hair

These structures are hard to generate and hard to blend:

• Ears are geometrically complex and often left from the original subject, producing a mismatch with the replaced face’s apparent age or ethnicity.
• Teeth frequently render as an undifferentiated white mass rather than distinct teeth with individual edges and shadows.
• Hair at the boundary of the swap tends to lose strand-level detail, showing a slightly painted quality.

Eyes

Two independent checks:

• Catchlights should reflect the same light sources in both eyes, with consistent shape and position. Deepfakes frequently render generic or mismatched catchlights.
• Gaze geometry: both eyes should converge on a plausible focal point. Slight divergence reads as uncanny before it reads as fake.

Physiological signals

In video, blood flow produces subtle periodic colour changes in the face, the basis of remote photoplethysmography. Synthesised faces typically lack a coherent pulse signal. This is video-only and outside what a single-image tool can measure, but it is one of the most robust research directions.

What our tool can and cannot see

We analyse a single still image. We can surface:

• Compression inconsistency around the blended region, via ELA.
• Noise-statistic differences between the swapped area and the rest of the frame.
• Pixel-consistency anomalies from smoothing at the blend.

We cannot perform face detection, landmark analysis or temporal checks. A deepfake still that has been heavily re-compressed may pass our layers. Treat a clean score on a suspected deepfake as inconclusive rather than exonerating, and escalate to a specialist tool if the stakes justify it. To begin, test the still.

The asymmetry

Generation is improving faster than detection, and that is unlikely to reverse. The durable defence is not a better classifier. It is provenance: signed capture, signed content credentials, and chain of custody. Detection buys time; provenance is the actual answer.