The existence of these sentiment analysis tools is no secret, though. Which means, Balaji said, that companies are now learning how to influence the bots reading their reports. “The stakes are really high here. So the minute they can game the software, they will try to game it.”

These efforts take various shapes. Institutionally, companies are centralizing all corporate communication “to tightly manage the specific words used…since they are being stored to develop a trackable lexicon to feed AI automated trading algorithms,” a report from the National Investor Relations Institute said last year. At T-Mobile, Paellmann said, the public relations team knew that media companies use bots for routine stories to save their reporters time, and that algorithms write many of the reports that summarize earnings releases. “So they’d draft the releases thinking about what information to present first, so that the first headlines coming out on company earnings would sound positive.” For other documents, such as quarterly reports or SEC filings, Paellmann’s team tried to alter as few of the words as possible from quarter to the next. “Some of these bots could look at materials and see where changes have been made, as an indication of where the tone of the company has changed. We were aware of that, so we were very careful in changing the language.”

When changes are made, the most sophisticated companies pay Johnsonian attention to their lexicons. Ashwell, the editor of IR Magazine, recalled being told, by the investor relations executives of a Canadian insurance company, that they used IBM’s Watson to scrutinize their language for sentiment. Words cataloged as “negative” in the Loughran-McDonald dictionary are swapped out for synonyms it doesn’t list. Yang and his co-authors found that, in the pool of corporate filings they studied, the average document used 1.63 negative Loughran-McDonald words per 100 words. But there were four other words per 100 that another taxonomy, the Harvard General Inquirer psychological dictionary, classed as “negative” that Loughran-McDonald doesn’t include. Admittedly, not all of the Harvard words will be necessarily have negative implications in a financial context, Loughran pointed out. “The Harvard dictionary considers ‘liability’ a negative word, but no reader of corporate reports will read one and go: ‘Oh my gosh, they said liabilities!’” Even so, the deliberate bypassing of Loughran-McDonald’s “negative” words is statistically clear: The frequencies of those terms shrank significantly after its publication. The five most-avoided words, ranked by the drop in their usage, were: “restatement,” “declined,” “misstatement,” “closure,” and “late.”

This tug-of-war between companies and their automated analysts is likely to result in a stalemate. Companies still have to make their disclosures, and they still have to do so in language that meets regulatory standards. As Ashwell pointed out: “I don’t think we’ll get to the point where people say: ‘We’re bidding adieu to our dividend payments this quarter.’” And in large, complex companies, the sentiment behind language often gets overshadowed by more pressing and immediate concerns. Since the fund of words that can be permissibly used is limited, the bots will eventually learn them all.

Eventually too, though, these improved bots will be ubiquitous. Everyone will use them—which is, in the quest for an edge, the same as no one using them. Then the game will move on to the next tactic. “What I’m wondering now is whether there’s a tool being developed that reads the body language of executives,” Ashwell said. “Do they have physical tells? Will a company find its CFO is more relaxed if he’s in a home office? Now that we’re all on video all the time, that’s what I’m interested in.”

📬 Sign up for the Daily Brief

Our free, fast, and fun briefing on the global economy, delivered every weekday morning.