Quantifying the Impact of Media Strategies on the Public Belief

By Lina Sorg

Journalistic reporting is essential to the formation of individual and collective beliefs about topics that range from scientific research to current events. Journalistic codes of ethics emphasize the importance of fair and objective reporting. Fairness norms are often translated into reporting via a heuristic that implies that different sides of a controversial issue should receive equal coverage. However, sometimes this standard is detrimental. “When we’re talking about matters of fact where scientific evidence should bear on your belief, that can be problematic because it’s often the case that one side of a controversy actually has a much greater weight of evidence and support,” Cailin O'Connor of the University of California, Irvine said.

The media also accentuates novelties and extremities by reacting to reader incentives. “Readers tend to like things that are surprising, novel, shocking, or extreme,” O’Connor continued. “Journalists respond to this pressure by publicizing this sort of content disproportionately.” Because the intent of news outlets is to publish “newsworthy” events, many people do not consider this method to be problematic. However, extremity reporting can subjectively skew information for public consumption towards extreme content that does not accurately represent reality. The advent of social media—which increases incentives for shocking material via clickbait and the prospect of “likes” and “shares”—further encourages this practice.

In short, all reporting is subject to characteristic distortions. “Media is one of a number of important factors that shape what information people see and what beliefs they derive on the basis of that information,” O’Connor said. During a minisymposium presentation at the 2021 SIAM Conference on Applications of Dynamical Systems, which took place virtually this week, she employed computational models to investigate the way in which distortions and individual confirmation bias collectively misconstrue public belief about matters of fact.

O’Connor’s simple model involves the following set of distributions:

An objective distribution of events in the world
A distribution of reported events that is derived from this reality
A distribution that represents an agent’s prior perception of reality
A distribution that represents an agent’s posterior perception of reality after encountering reported events.

She began with a normal distribution of events and considered three reporting practices that distort reality: hyperbole, extremity, and fairness.

Hyperbole exaggerates events to make them sound more extreme. This process moves them away from the “social neutral point,” which shapes consumers’ perceptions of novelty and fairness. “We assume that what seems novel, extreme, or fair to viewers isn’t necessarily going to be indexed to the objective distribution in the world,” O’Connor said. “There might be some social perception that tells them what sounds shocking or extreme, and may or may not track that reality.” Sometimes symmetries or facts about the world anchor the social neutral point; other times, propaganda campaigns or public debate anchor communal perception of neutrality and divert the point away from the distribution of real events.

Extremity bias occurs when journalists only report on real events that are a certain distance away from the social neutral point and fall on the tails of the distribution. Though they do not change or exaggerate the truth, they do selectively choose what truths will be most interesting or surprising to readers. Finally, reporters practice fairness when they record events with equal weight on both sides of the neutral point. This method shapes the distribution into a double-peaked curve, with more weight on one side of the neutral point than would otherwise occur in the real distribution (see Figure 1).

Figure 1. Fairness shapes the distribution into a double-peaked curve, with more weight on one side of the neutral point than would otherwise occur in the real distribution.

O’Connor ran a course of simulated rounds that randomly selected data from the reported distribution, supplied that data to the agent, and allowed the agent to update via Bayesian conditioning. After 1,000 updates, she observed the agent’s posterior distribution and paid particular attention to two specific measures:

Qualitative facts about the posterior distributions
Average agent error across different runs of the model, with respect to both mean and variance of the objective distribution.

All three of these distortions create inaccurate beliefs in agents when the objective distribution is not centered on the social neutral point. The use of hyperbole causes evident misrepresentation in the model. As a result of overestimation, the mean of the distribution is too extreme and the variance is too wide; the agents’ posterior resembles the recorded data and they begin to think that extreme events occur more frequently than in reality. Extremity distortion also overestimates the mean, as one side of the distribution has disproportionately more weight. “Again you see agents who think that what is happening in the world is more extreme than what is actually happening,” O’Connor said. Fair reporting, however, has the opposite effect. Weight on the wrong side of the social neutral point underestimates the mean and pulls beliefs closer to that point. In this scenario, the agents assume that the world is more like the social assumption than in actuality.

Next, O’Connor introduced confirmation bias—a set of biases in which individuals are unresponsive or resistant to evidence that counters their current beliefs—to her model. “If you encounter evidence that disconfirms your beliefs, you tend to reject it more than evidence that confirms your beliefs,” O’Connor said. With this bias in place, agents are less likely to incorporate data in their posterior distribution that does not match their current beliefs. For example, hyperbolic distortion yielded continued confidence in prior beliefs and an error in mean beliefs. “What that means is that people who engage in confirmation bias can find data or evidence that fits with their prior beliefs and confirms what they already thought was true,” O’Connor said. “So they tend to just become more confident in a picture that basically supports what they already believed” (see Figure 2).

Figure 2. Hyperbolic distortion in the presence of confirmation bias yields continued confidence in prior beliefs and an error in mean beliefs.

Confirmation bias increases the already-subjective nature of extremity and fairness, as each distortion produces double-peaked distributions (see Figure 3). Agents therefore move towards the closet peak, which typically corresponds with reported content that fits their prior beliefs. This propensity leads to polarization, path dependency, and a high mean error.

Figure 3. Confirmation bias increases the problematic nature of extremity and fairness. Each distortion produces double-peaked distributions.

Ultimately, hyperbole, extremity, and fairness are especially harmful when combined with confirmation bias because distortions shift weight toward more extreme events, and prior beliefs heavily influence posterior distributions. O’Connor’s model also proves that distortions by omission are sometimes worse than distortions by commission, i.e., hyperbole is less problematic than extremity or fairness. The latter two practices are particularly troublesome because they involve truthful reporting in a selective way, which people often fail to identify as bias.

Lina Sorg is the managing editor of SIAM News.