I read this paper Text Compression as a Test for Artificial Intelligence, Mahoney, 1999.

So far I understood the following: Text Compression tests can be used as an alternative to Turing Tests for intelligence. The Bits per character score obtained from compression of a standard benchmark corpus, can be used as a quantitative measure for intelligence

My questions:

  1. Is my understanding of the topic correct?
  2. Does this mean that applications like 7zip/WinRar are intelligent?
  3. How are the ways a human compresses information (as in form of summary) and ways a computer compresses (using Huffman coding or something) are compatible? How can we compare that?

2 Answers 2


The paper suggests an alternative test to the famous Turing test, which tests a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.

In this test, if winRar or 7zip will compress a file similarly to how a human would compress a file (how does a human compress a file?!), then, yes, those programs will pass the test and will be considered intelligent.

... thus compression ratio on a standard benchmark corpus could be used as an objective and quantitative alternative test for AI (Mahoney, 1999).

  • 1
    $\begingroup$ How does the paper think that the way a human compresses some information (in form of summary or something) and the way a computer program does (through Huffman coding or frequency tables) are compatible? $\endgroup$
    – Aether
    Mar 22, 2021 at 14:22
  • $\begingroup$ I have no idea. I think that your question is valid and the paper does not answer it. $\endgroup$
    – Cohensius
    Mar 24, 2021 at 14:24

Matt Mahoney is one of the organizers of the Prize for Compressing Human Knowledge (Hutter prize), with Jim Bowery and Marcus Hutter. The prize is awarded to progress in compressing 1GB of Wikipedia text, so let us try to work out an example from it.

We read in the Wikipedia that Marcus Hutter was born "April 14, 1967". From our prior knowledge of the world (or reading the Wikipedia!), we know there are $12$ months per year, up to $31$ days per month and we are in $2022$ now, luckily close to $2048$. Therefore, we can safely use $4$ bits for the month ($2^4>12$), $5$ for the day ($2^5>31$) and $11$ for the year ($2^{11}>2022$), totalling $20$ bits for a lossless compression of a birth date (CE).

The original string "April 14, 1967" is $14$ characters long, so $112$ bits assuming $8$ bits/character.

Thus some understanding of this tiny bit of information gets a $\frac{112}{20}$ compression factor, about $5.6$. Maybe the comparison makes little sense here, but $5.6$ is similar to the initial $5.46$ compression factor by Matt Mahoney. Of course there are a lot of much more complex information structures in the Wikipedia. Besides, the month length can go up and down some characters, so this sample value is far from a constant for all dates.

Anyway, hopefully this simple example shows some relationship between text compression and understanding.


You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .