The Short Answer
Yes, face recognition keeps working as you get older, and modern systems are far better at it than older ones. A photo of you at 30 and a photo of you at 40 will usually still be matched to each other. The longer the gap, the harder it gets, and two kinds of photo are much harder than the rest: childhood photos and photos of very old faces.
To see why, it helps to know what actually changes in a face over time, and which parts of a face a recognition network leans on.
What Changes in an Ageing Face
The common line is that "bone stays the same and only skin changes". That is too simple. Every layer of the face changes, at different speeds.
| Layer | What happens with age | How much it matters for matching |
|---|---|---|
| Skin | Loses collagen and elasticity, gains fine lines, spots and uneven tone | Small. Networks are trained to ignore texture like this |
| Fat | Facial fat pads shrink and slide down, so cheeks flatten and folds deepen | Medium. It changes the outline of the cheeks and jaw |
| Muscle | Tone drops; habitual expressions leave lasting creases | Small to medium |
| Bone | Slow remodelling: the eye sockets widen, the upper jaw recedes and the lower jaw loses some height | Medium over several decades |
| Cartilage | The nose and ears keep lengthening slightly through adult life | Small |
The bone changes are real and well documented. A review by Mendelson and Wong in 2012 described how the eye sockets get larger and the midface and jaw lose volume with age, which is part of why older faces look "hollower". But these changes are slow and they mostly keep the proportions between features in place. Your eyes stay the same distance apart. Your nose stays where it is relative to your mouth. That layout is what recognition networks rely on most.
Read more: Do You Look Like a Celebrity from a Different Era? Here's Why AI Thinks So
Why Modern AI Copes With Age
Older face recognition methods compared surface patterns, so new wrinkles or sagging cheeks could throw them off. Deep networks work differently. They are trained on millions of photos of the same people taken years apart, and the training rewards them for producing the same output for the same person no matter what. Over time the network learns that wrinkles, grey hair and a softer jaw are not reasons to call someone a different person.
Training sets help here. The dataset Ollie's own model learned from has about 5.8 million photos of 85,742 people, and many of those people are public figures photographed across long careers. A network that sees an actor at 25 and again at 60, labelled as the same person, learns which parts of the face survive ageing.
The result is a face embedding: a list of 512 numbers that describes the face. Two photos of the same person, even years apart, produce embeddings that sit close together. You can read more about how this works in what a facial embedding is.
What the Benchmarks Show
Researchers measure cross-age accuracy with datasets built from photos taken years apart. The best known are:
- AgeDB: 16,488 photos of 568 people with exact ages. Its hardest test, AgeDB-30, compares photos taken about 30 years apart.
- CALFW (Cross-Age LFW): the classic LFW test rebuilt with same-person pairs chosen to have large age gaps.
- CACD: about 163,000 photos of 2,000 celebrities taken between 2004 and 2013.
- FG-NET: 1,002 photos of 82 people from childhood to old age. Small, but brutal.
On these tests the gap between easy and hard is clear. The ArcFace paper from 2019 reported 99.83% on LFW, about 98% on AgeDB-30 and about 95% on CALFW for the same model. So a 30-year gap costs a top system a couple of percentage points, not a collapse. Ollie's own model shows the same pattern: 99.5% on LFW, 96.5% on AgeDB-30 and 95.2% on CALFW. FG-NET, with its baby and child photos, is much harder still.
Independent testing agrees. The US National Institute of Standards and Technology (NIST) found in its 2019 demographic study of face recognition that error rates were higher for children and for the elderly than for adults in the middle of life.
Read more: What Is Facial Embedding? The Technology Behind Celebrity Matching
Get new guides by email
Where Matching Breaks Down
Childhood to adulthood
This is the hardest case by far. A child's face is not a small adult face. The skull grows, the lower face lengthens, the nose and jaw develop, and proportions shift a lot through puberty. A photo of you at 8 and a photo at 28 can look like two different people to a network, and sometimes to your relatives too.
Very large weight change
Gaining or losing a lot of weight changes the outline of the cheeks and jaw. Combined with ageing, it can move your embedding further than age alone would. There's more on this in how weight change affects face matching.
Old photo quality
A scanned print from the 1990s is often soft, faded and slightly tinted. Some of what looks like an "age problem" is really a photo problem: low resolution and colour casts hurt matching on their own.
What This Means for Your Celebrity Match
If you are an adult, your celebrity lookalike results should stay fairly stable from year to year. The same names tend to come back because the network is reading the same underlying layout. Over a decade or two, your matches can drift, often toward celebrities who are photographed at a similar age to you now.
You can test this yourself. Here's a simple way to do it:
- Find a clear, front-facing photo of yourself from five or more years ago, ideally the original file rather than a scan.
- Take a new photo in similar conditions: rear camera, even light, neutral face.
- Run both through Ollie and compare the top five.
- Names that appear in both lists reflect your stable structure. Names that only appear once are more likely tied to that photo's age, weight or lighting.
If the old photo is from your teens, expect bigger differences. That is your face finishing its growth, not a fault in the model.