Limits of Detecting Text Generated by Large-Scale Language Models
Preprint
Abstract
Some consider large-scale language models that can generate long and coherent pieces of text as dangerous, since they may be used in misinformation campaigns. Here we formulate large-scale language model output detection as a hypothesis testing problem to classify text as genuine or generated. We show that error exponents for particular language models are bounded in terms of their perplexity, a standard measure of language generation performance. Under the assumption that human language is stationary and ergodic, the formulation is extended from considering specific language models to considering maximum likelihood language models, among the class of k-order Markov approximations; error probabilities are characterized. Some discussion of incorporating semantic side information is also given.Keywords
All Related Versions
This publication has 13 references indexed in Scilit:
- A Note on Reverse Pinsker InequalitiesIEEE Transactions on Information Theory, 2019
- $f$ -Divergence InequalitiesIEEE Transactions on Information Theory, 2016
- On Rate of Convergence of Statistical Estimation of Stationary Ergodic ProcessesIEEE Transactions on Information Theory, 2010
- Neyman-pearson detection of gauss-Markov signals in noise: closed-form error exponentand propertiesIEEE Transactions on Information Theory, 2006
- Authentication theory and hypothesis testingIEEE Transactions on Information Theory, 2000
- A convergent gambling estimate of the entropy of EnglishIEEE Transactions on Information Theory, 1978
- Codes based on inaccurate source probabilitiesIEEE Transactions on Information Theory, 1971
- On estimating the entropy of random fieldsInformation and Control, 1970
- A note on the entropy of partially-known languagesInformation and Control, 1968
- The reconstruction of abbreviated printed messages.Journal of Experimental Psychology, 1954