Your DMV photos would be ideal. She would likely know the exact type of camera used, the color temperature of the lights installed in the building, and the date to know the ambient conditions (and god knows what else).
Distinguishing Identical Twins by Face Recognition
The best performance was observed when all images of a pair of identical twins were taken on the same day in the studio environment and the twins had a neutral expression. For this case, the best performing algorithm had a 90% confidence interval for the EER from 0.01 to 0.04.
The corresponding confidence interval for the same day with ambient light and neutral expression was 0.12 to 0.21. For cross-year recognition, the best 90% confidence interval for the EER was from 0.15 to 0.27.
Performance also showed that gender does not effect performance and that there is an age effect. The results showed that it is easier to distinguish twins over 40 year old than twins under 40.
–P. Jonathon Phillips, Patrick J. Flynn, Kevin W. Bowyer, Richard W. Vorder Bruegge, Patrick J. Grother, George W. Quinn, Matthew Pruitt. www3.nd.edu pdf
Abbreviated as EER. A biometric security system predetermines the threshold values for its false acceptance rate and its false rejection rate, and when the rates are equal, the common value is referred to as the equal error rate. The value indicates that the proportion of false acceptances is equal to the proportion of false rejections. The lower the equal error rate value, the higher the accuracy of the biometric system. –webopedia.com
I don't exactly understand how to interpret an EER (as a percentage?) but The Machine would compare pictures taken under the closest matching conditions: with the same expression (ideally, none) and indoor lighting, to reduce the EER as much as possible.
Note that the ERR increases by an order of magnitude when using ambient lighting, as opposed to a studio environment. Whereas a lapse of one year between photos was only a small increase.
Don't forget that this study had a control, in that the same camera was used, but She can probably compensate if it's not. I assume She is quite capable of reverse-engineering what a particular piece of hardware was used from its result, e.g., a photo.
Besides, we're talking about an A.I. that was built to deduce. She probably has an EER of .001 at a 99% confidence interval.