Study: companies are increasingly tailoring their financial statements to cater to algorithms parsing text and speech and avoiding phrases perceived as negative
- Study finds companies alter words to cater to listening algos — Emphasis on positivity as negative phrases get ditched Tweets: @jessefelder , @lisaabramowicz1 , @melinanders , and @rickgreennybiz Tweets: Jesse Felder / @jessefelder : “Because such rules are transparent, observable, or reverse-engineerable to at least some degree, agents who are impacted by the decisions have the incentive to manipulate the inputs to machine learning in order to game at a more desirable outcome.” https://www.bloomberg.com/... Lisa Abramowicz / @lisaabramowicz1 : Corporate leaders have started to adapt their statements and even their delivery to cater to the algorithms parsing text and speech for trading signals: Columbia & Georgia State research https://www.bloomberg.com/... Anders Melin / @melinanders : CEOs now sound even more like robots to charm the robots. https://www.bloomberg.com/... @rickgreennybiz : Great news for #distresseddebt investors, who rely on actual research rather than easily fooled #algorithms. (Also provides some job security for financial journalists.) @gregorhunter explains in plain English. #restructuring #bankruptcy https://www.bloomberg.com/... via @business
Context & Ripple Effects
The study from Columbia and Georgia State researchers lands in a market that has already rebuilt itself around machine-read text: Wall Street has been selling raw data like social media sentiment instead of analysis since research emails stopped getting opened, so algorithms now parse filings at scale. The finding is a textbook Goodhart response — once a metric becomes a target, the measured parties optimize the words themselves.
It also fits a broader pattern of humans managing what machines see: Google's AI unit was separately documented asking scientists to portray its technology positively in papers on sensitive topics. Corporate disclosure is converging on the same behavior — language tuned to the algorithm rather than the reader.
First-order effects
- Investors running text- and speech-analysis models on earnings calls and filings are now consuming inputs that issuers have deliberately scrubbed of negative phrasing, degrading the signal their models were built to extract.
- Issuer investor-relations teams gain a new editing discipline: word choice becomes a compliance-adjacent function aimed at machine scoring rather than human comprehension.
Second-order effects
- Quant shops pushing computer-driven trading into credit — the $40T corporate bond market push led by Blackstone — have to discount textual sentiment features or source harder-to-game data, shifting spend toward alternative datasets.
- Data vendors selling sentiment feeds face pressure to prove their signals survive adversarial wording, or risk the same commoditization that hit sell-side research.
Third-order effects
- If scrubbed language becomes standard practice, textual disclosure loses value as a signal and markets drift further toward hard numbers, alternative data, and primary sources — with regulators eventually forced to decide whether algorithm-gaming of filings constitutes misleading disclosure.
- The deeper shift is that every transparent scoring rule invites input manipulation, so the arms race between model builders and the entities they score becomes a permanent feature of information markets.
The trend: As machine reading replaces human analysis of corporate disclosures, issuers are optimizing language for the algorithm — turning textual sentiment into an adversarially gamed signal.