Corporate scandals often follow a pattern: Whether it’s Theranos and its fraudulent blood testing technology, Wells Fargo and its fake financial accounts, or Volkswagen and its bogus emissions data, a whistleblower eventually comes forward to expose the behavior, and executives are held accountable.
“But what you start to realize is that the problems that have been uncovered have been going on for a very long time,” says Dennis Campbell, a professor of business administration at Harvard Business School. Far from being just a few bad apples, most business improprieties occur within a widespread culture of bad behavior—or at least, a lot of people looking the other way as misconduct is taking place, he says.
“You think about Wells Fargo, and there were thousands of employees engaging in these practices over many, many years,” says Campbell, who specializes in accounting and control. “But even though this stuff can be pretty widespread inside of companies when it occurs, it’s not getting out in a timely way.”
In an attempt to discover whether these problems could be exposed earlier, Campbell conducted an experiment with Ruidi Shang of Tilburg University in the Netherlands. To find upstream indicators of culture gone sour, they scraped employee reviews from Glassdoor.com, a website where employees can leave subjective anonymous reviews about their employer.
Harnessing machine learning to analyze text, they found that the reviews could serve as canaries in the coal mine of corporate misconduct, pointing to cultural factors that might eventually result in scandal earlier than they would otherwise attract attention. In a business environment where ethics are paramount, the findings may offer managers and compliance departments a new way to stop fraud before it starts.
“We show that this aggregation of information is not only a leading indicator of violations of rules and regulations, but it’s even a leading indicator of the whistleblower complaints themselves,” Campbell says of the study that resulted, which is forthcoming in the journal Management Science.
How bad behavior bubbles below the radar
An employee may not come forward right away to expose wrongdoing at a corporation for many reasons.
In the absence of directly observing egregious behavior by a particular individual, an employee may not think that a general breakdown of values and standards rises to the level where it should be reported to outside regulators or the press. Would-be whistleblowers might also fear possible repercussions against them, or worry they don’t have enough evidence to prove their allegations.
“Research shows there are big risks to being a whistleblower, and people are unlikely to stick their necks out unless they feel very strongly that misconduct is occurring and that there will be some reward for coming forward,” Campbell says.
On the other hand, sites such as Glassdoor allow employees to freely vent about problems within companies that may not amount to criminal malfeasance but can cause problems nonetheless.
“They are not even designed to detect misconduct, which is part of what makes this work—employees feel much more comfortable talking than they might in a narrower platform,” says Campbell.
Searching for red flags
To pick up on the indicators of trouble, Campbell and Shang constructed a dataset of reviews from more than 4,000 publicly traded firms between 2008 and 2016, extracting a vocabulary of 11,772 unique words used in the reviews. At the same time, they obtained a database assembled by the nonprofit Good Jobs First of almost 27,000 violations by the same firms from 2008 and 2017.
They then trained a machine-learning algorithm on the text, letting it independently decide which words were most highly associated with eventual violations. The words that rose to the top ranged from descriptive words, such as "pay" and "promotion," to clearly negative evaluations, such as "discrimination," "trouble," "favoritism," and "unethical."
Within the results, the researchers found they could clearly distinguish between firms with high and low numbers of violations by the weighted proportion of those words, even independent of other risk factors such as violation history or financial leverage.
In fact, those indicators rose and fell with remarkable consistency as violations occurred. “When we looked at Wells Fargo relative to the average in the banking industry, we could see it rise and then spike from 2009 to 2013 when their misconduct was happening,” Campbell says.
Even more important, the measure was able to predict violations before they were exposed by whistleblowers or the press. “Where the rubber meets the road is whether this measure says something about future misconduct,” Campbell says. “And [we found that] yes, it does, at least a year ahead of time.”
How firms can spot future misconduct
Glassdoor isn’t the only site that may uncover wrongdoing, Campbell says. The same analysis could be applied to any social platform in which employees are discussing company dynamics. Other sites where employees can review their employers or share feedback about their workplaces are Indeed and Great Place to Work.
These kinds of machine-learning techniques could be used by regulators in deciding where to target scarce resources on which firms to examine, as well as by investors in making decisions about risk.
It could also be used by companies themselves, Campbell says. If executives or board members are concerned about company culture, external or internal employee communications may expose potential issues.
“There could be all kinds of internal platforms where employees are talking and sharing information, and that you could apply this high-level machine learning to in an aggregated way,” says Campbell. “It could help you understand pockets in your organization where these cultural issues are popping up so you can get a handle on them.”
In that way, the most valuable benefits of this approach may not be in identifying future scandals, he says—but in preventing them before they occur.
Related reading from the Working Knowledge Archives
Reviews, Reputation, and Revenue: The Case of Yelp.com
Feedback or ideas to share? Email the Working Knowledge team at workingknowledge@hbs.edu.
[iStockphoto/EnvisionSrini]