1
$\begingroup$

It is an easy matter to make a paper look old, for example, using any of the techniques explained on this page of WikiHow: https://www.wikihow.com/Make-Paper-Look-Old.

Is current AI sufficient to distinguish a fake old paper from a real one?

$\endgroup$
  • $\begingroup$ How would the AI sense the paper under scrutiny? A photo of someone holding the paper? Output from a scanner? Output from a microscope? Touch sensors? Mechanical stress results? IR Spectrometer readings? C14 radio-carbon dating readings? How do humans typically discover that the paper is real or fake? $\endgroup$ – Neil Slater Nov 1 '18 at 18:37
1
$\begingroup$

Is current AI sufficient to distinguish a fake old paper from a real one?

It depends. Specifically, it depends on how much information you (or any automated system that feeds data to the AI), gathers about the paper you want to test, and how good that information is at distinguishing between real and fake old paper.

A model that distinguishes between cases in an input domain (which can be any consistent data, such as images of a subject) is called a classifier. The current state of the art classifiers working with sensory data, if fed enough training data of sufficient quality, can out-perform humans at certain tasks. This is a success of "narrow AI" over humans. This is not some magical ability of machine learning that always applies though, there are some strong caveats:

  • The task must be feasible given the data collected. Low-resolution photos of samples of the paper might not contain enough information for your example task at all.

  • When comparing with a human performing the same task, they must be given identical data. If a human would gain more data by interacting with the paper in your example, it can give the human an advantage that a narrow AI cannot easily replicate.

  • The quality of training data "ground truth" must be higher than achievable by a single human. In your example, this could be achieved simply by knowing the provenance of all the training data.

  • There often needs to be a large amount of training data to achieve better-than-human ability, although the actual amount required depends heavily on the task.

Expanding on the first point, image classifiers can sometimes be surprisingly good, they have different strengths and weaknesses compared to human vision, so it is possible for them to sometimes be good at making classifications that humans are bad at. However, there is no guarantee of getting that kind of result for any specific task.


Expanding on the second point, if interacting with a subject is allowed in order to classify it, this moves from the domain of "narrow AI" classifiers into active rational agents.

If you gave a piece of paper to a human expert on old paper, and ask them to classify it, they might first look at the paper, then maybe sniff it, touch it . . . still undecided, so get out a magnifying glass to check a couple of features that might be good telltale signs . . . etc.

To build an equivalent artificial agent is harder than a single classifier: It might be possible to purpose-build a collection of narrow AI classifiers, that output a class and confidence levels that decided whether certain tests should be carried out. It might even be possible to give a more general AI the same toolkit plus a very long time to experiment and it could learn for itself which tests to do in which order and how much to rely on the separate results. Or some hybrid of the two approaches. However, any such sophisticated system would still be bound by the limits of individual classifiers.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.