Learning AI #35

Steven Hyde of Boise State University published a machine-learning model in an academic journal and the model scored how likely a CEO is to be lying on an earnings call. Earnings calls are done quarterly for public companies and the company, usually the CEO, tells investors how the company performed and how he or she sees the future. Stock prices hang in the balance based on what the CEO says on these calls and how it is said.

The Boise State team trained a model on thousands of earnings call transcripts plus communications from CEOs who were caught lying: executives whose financial statements the SEC deemed fraudulent.

The AI model scored each transcript on 22 linguistic features. Here are three of the most prominent:

Fewer singular pronouns. Words like "I," "me" and "my" drop. The model picks up “distancing” language, and dropping first-person pronouns fits that description. Example: "The numbers were reviewed and approved" claims the same thing as "I reviewed the numbers" and puts no person in it. In an earlier study about deception, the scientists found that liars used fewer self-references and used more negative emotion words like “frustrated” or “angry.”

Many times I have listened to earnings calls and heard CEOs use these exact words.

Inconsistent verb tense. The speaker shifts between past, present, and future within the same account. This as a sign the speaker isn't recalling one fixed event, but may be creating a composite and making it up as he goes along.

Fewer words referring to sensory experience. Details about motion, space, and time are sparse. A true account typically carries where, when and how things moved, and those details take effort to produce when the event didn't happen the way it's described.

So, how can AI detect lies and deception? It’s all about pattern recognition. Experts, like the team at Boise State can give us the linguistic clues that are tells for deception. This research has been done and there are many books written about it.

After you know the tells, just ask the AI to compare the frequency (or lack of frequency) of certain words and phrases used on the earnings call with what would be expected in normal speech. Judgement and intuition get replaced by a mathematical calculations. Sounds easy, but its bad news for the liars.

Since the bad guys are always a step ahead of the good guys, the real question to ask is not how to detect lies, but are the liars now aware of these tools and adjusting their behaviors and words to throw people off the trail? You can see how this gets complicated.

Things I think about

Sharks are older than trees. Sharks show up in the fossil record more than 400 million years ago, and the first trees didn't appear until roughly 50 million years after that.