“My notes and other stuff”

2023/01/14

Paper: Agentive Language in Accident Investigations

It is time to cover a very good paper by Crista Vesel, a shorter version of her thesis, titled Agentive Language in Accident Investigation: Why Language Matters in Learning from Events.

In short, the paper states that inadvertent ways to structure your sentences in a text or a report may carry implications of blame and convey more deliberate actions from participants than they actually intended, and harm your ability to learn from events:

Humans naturally want to know who or what was responsible for an action, especially if it led to an undesirable event. This assignment of action is called agency. An agent is “A person or thing that takes an active role or produces a specified effect”. In the example, “Bob spilled the chemical”, Bob is the agent of the action. However, the simplicity of this sentence does not tell us whether Bob spilled the chemical intentionally, by accident, or was just near the chemical when the event occurred. We likely assume that the agent of the action acted independently and made a free will choice to act. This assumption can make all the difference when we are attempting to learn from the event and influence how we create safety in our work environment.

She starts by describing three biases, which are likely more frequent in Western societies with an individualistic slant than others with a more collective view:

Put another way, these tend to make it very damn intuitive to assign blame to people during accidents. Subtleties in the language chosen can cater to these tendencies, unintentionally.

Research points out that small changes in language can have an impact on how we assign agency. For example, the author mentions how we use the word "accident" interchangeably for both acts that were intentional and unintentional, despite most people thinking there is a serious difference between both. In a culture where agentive language is the norm, mentioning the term "car accident" brings up images of a driver who is at fault, even before any detail is given and if the driver were actually not to blame.

A rich context and description is required to properly convey the nature of events. An author assuming that the audience will ascribe the same meaning to words as the one they intended can be enough to leave room for interpretation with unintended consequences. The fewer words you use, the more room you may leave for interpretation, and the greater the risk.

The use of agentive language tends to be more present in English than many other languages:

In one study, English and Spanish speaking participants viewed videos of actors in an event that could be interpreted as either nonintentional or intentional and then provided verbal descriptions of the events. For example, an actor would pop a balloon using a tack (intentional). Alternatively, the actor would reach to put a tack in a container and the balloon would pop during the reach (nonintentional). The participant descriptions were coded as being either agentive or nonagentive. An agentive description would be something like, “He popped the balloon.” A nonagentive description could be, “The balloon popped.” The study concluded that English, Spanish, and bilingual speakers described intentional events agentively, but English speakers were more likely than the other groups to use agentive descriptions for nonintentional events. Another study showed similar results between English and Japanese speakers.

(Personal note: I've seen this happen a lot through the pervasive "avoid passive voice" writers' tips, which I've never found to have such universal support in my native French. That tip is everywhere in English as far as I can tell, and always felt weird specifically because it added intention where I didn't want it! The one place I noticed passive voice winning consistently in English is news articles about people dying in the presence of police officers...)

Another concept is language priming, which was studied by exposing participants to agentive language (“He crashed the car”) or nonagentive language (“The car crashed”). They found out that English speakers in this experiment remembered people involved better with agentive language, but may have compromised their ability to focus or remember other details of the situation. These influences are everywhere, subtle, and encoded in all sorts of cultural artifacts.

The author then circles back to incidents, tying language more directly to blame. In the context of incidents, the earliest models (see previous post on incident models) were focused on human actions as principal causes. Despite decades and many newer models trying to break away from this, human error still has a prominent role in investigations to this day.

Simplistic reductions from a rich spectrum of acts into "success" and "failure", matching these categories, limit possibilities and distort problems into having equally simple solutions:

Humans are different and inherently complex, with dynamic and emergent cognition and a unique ability to learn. The social nature of human interaction also leads to uncertainty and unpredictability. We can really only say that a human “failed” or “succeeded” in hindsight, once the outcome of the action is known. However, the agentive language that we use for machines is often used to describe human action.

This point about outcome (success, failure) only being available after the events (in hindsight) is critical: at the time the acts took place, when the decisions were made, the outcome were not necessarily predictable. The mechanisms behind successes and failures tend to be the same, with only the result (which was unknown at the time) changing the classification. Forcing this classification forces the investigator into taking a judgmental position.

Many guides for incident investigators (the author cites the older version of USDA's guide) take a normative stance that presuppose humans failing as a cause for incidents:

The first paragraph in the Serious Accident Investigation Guide states, “The causes of most accidents or incidents are a result of failures to observe established policies, procedures, and controls.” This language presupposes that the cause of accidents is human failure. Though the guide goes on to develop three categories of “significant findings”, human, environmental, and material, only humans can meet the condition of failures to observe. The SAIG repeats the word “failure” 91 times in its guidance for investigators, particularly in regard to humans. It is not surprising that the word “failure” also appears multiple times in resulting accident reports to blame human action.

She mentions that such guidelines are priming investigators to look for and find failure. If investigators are trained to find agency, they're likely to overlook situational factors, and in turn lead to suggestions that are less effective since they work from a less complete picture.

A common way to work around this is to try and make the investigation report as factual as possible. The reality is that the choice of words in a factual re-telling still has an impact. She gives the example of a study done around the wardrobe malfunction in Justin Timberlake and Janet Jackson's half-time show in 2004. People were shown the video and then read a description of the event with distinct degrees of agentiveness:

Depending on which one was read, "an agentive report led people to think that Justin Timberlake owed more than $30,000 more (an extra 53%) in fines compared with a nonagentive report"—despite having access to the same facts and seeing it on video.

The author adds:

Small differences in language can have a large impact on causal attribution, with the sentence structure influencing how a reader perceives causality of the event. The active verb voice is one impactful linguistic technique that can lead to the assignment of agency. When a verb is presented in the active voice, the subject is seen to be doing the action, as in “Sara hit the ball.” Here, Sara is the subject of the sentence in relation to the ball. The passive verb voice would structure the sentence more like this, “The ball was hit.” Research has shown that attributions of control, causation, and dominance are all affected by the verb voice, even if an agent’s actions are presented as nonintentional. “Active voice apparently conveys a sense of control and causation that is lacking in the passive voice.”

When writers advice is then added to guides, such as the SAIG stating “Write causal factors in the active voice, clearly identifying the actor(s) and causal action, along with any necessary explanation,” they are inadvertently recommending that investigators write reports that prime the readers towards blame.

She points out that in fact, there were many reports where people were found as a causal source of incidents even when "humans were not acting on the environment, such as when wildland firefighters were simply walking through the forest and were struck by falling tree branches." The guide also recommended that reports be economical with their words, often removing nuance and directing people towards binary choices (safe, unsafe, and all unsafe acts in turn being either errors or violations).

By laying most of the blame to the proximate actors, those at the sharp end, it becomes easy to ignore and paper over the influence that can be had at the blunt end, far from the field of action. These reports can also end up used in criminal and civil prosecution, and the language used in turn can impact the type of penalties that result.

In fact, the directives about which language is acceptable to use can shift how the investigation unfolds:

Searching for causes restricted our teams from exploring some very critical aspects of our organizational culture and prevented us from asking hard questions regarding the perverse nature of some of the influences we discovered. For example, we had trouble making the case for the influence of overtime pay on the behavior of our crews. We had recorded admissions of workers indicating that overtime played a role in decision-making and risk acceptance, but we could not prove a causal link. Simply shifting the conversation to “influence” was enough of a softening of language to allow a dialogue to begin that could explore the possible ways that overtime nudged decisions.

Changing the language in the organization opened doors and oriented investigators towards asking better questions, and supported valuing learning from events above attribution of blame.

The author concludes:

Though most linguistic cultures use agentive language to describe intentional events, English speakers have been shown to exhibit a preference of assigning human agency to nonintentional events (accidents). English speakers are more affected by agentive priming language, like that found in certain investigation guides, which may lead them to assign more blame and punishment to people involved in events. All of these factors may lead to an agentive bias, where details of a situation are ignored or not recognized, and a simplistic causal attribution is applied to the human agent closest to the event.

[...]

Use of the active verb voice, where the subject of the sentence is “doing” the action, implies a sense of agent control and causation, even if the accident was truly a nonintentional event or the implied agent had no direct role in the action.

[...]

Agentive language can reduce inquiry into the conditions surrounding an event and lead to biased conclusions of causality. Changing agentive language to a language of inquiry can result in organizational changes that positively impact safety culture.