Can a speaker fool the deception model?

If deception can be detected in text, and the properties that signal deception are known (more or less), can a clever speaker or author use this knowledge to come across as truthful?

The short answer is No. The reason is that language production is a deeply unconscious process. Although we can decide consciously what we would like to say, we have much less control over how we say — much less even than we think we do.

A speaker, with some practice, could start to insert more first-person singular pronouns and exclusive words into their speaking, but only by concentrating. But concentration creates other problems: it tends to make the whole delivery sound more stilted, and it consumes processing resources spent thinking about the form of the speech that come at the expense of the content. In other words, to make the speech sound less deceptive, it’s almost necessary to make it more bland, and therefore less effective for whatever reason it is being made in the first place.

In written text, there is more opportunity to work on the signals in the text to make it seem less deceptive. For example, an author could use our deception detection software to measure the deceptiveness of the text, and change a few words to improve it.

There are two problems with this. First, it’s always better to think of deceptiveness in the context of a set of documents from the same domain. It may not be obvious what the norms for a domain are, and so how much the particular document needs to be adjusted. Second, explicitly manipulating word frequencies tends to create unusual documents because that’s not how documents get edited. These kinds of tinkering run the risk of creating an alternative signature which may also be detected by a different kind of analysis.

The bottom line is that language carries all sorts of information at different levels of abstraction, and it’s all consistent when the language was generated in an ordinary way. Messing around with pieces of language as if they were independent quickly breaks some or all of this global consistency.


%d bloggers like this: