SafetyInsights.org

Does the language we use in investigations shape allocations of blame? Quite possibly.

Today’s study is from Vesel, C. (2020). Agentive language in accident investigation: Why language matters in learning from events. ACS Chemical Health & Safety, 27(1), 34-39.

Spotify: https://open.spotify.com/episode/3PfEq7a3Bsq6zq48TDYMDe?si=AyiJO2O4Rbe-NmjbV5ot9w

Apple: https://spotifycreators-web.app.link/e/j6EE8d95bWb

Make sure to subscribe to Safe As on Spotify/Apple, and if you find it useful then please help share the news, and leave a rating and review on your podcast app.

I also have a Safe As LinkedIn group if you want to stay up to date on releases: https://www.linkedin.com/groups/14717868/?lipi=urn%3Ali%3Apage%3Ad_flagship3_detail_base%3Bhdg8uJYYT%2BmsMqZvpHBmdQ%3D%3D

Shout me a coffee (one-off or monthly recurring)

Transcript:

6.398562 11.678562 Does the language use in investigations? For instance, Bob spilled the chemical,
11.678562 16.638562 rather than the chemical was spilled. Lead to blame in short-circuit learning within
16.638562 22.238563 organisations. Let’s find out. G’day everyone, I’m Ben Hutchinson and this is Safe As,
22.238563 27.038563 a podcast dedicated to the thrifty analysis of safety, risk and performance research.
27.038563 30.238563 Visit safetyinsights.org for more research.
31.278562 37.518563 This paper from Krista Vessel explored how language can hinder or facilitate organisational learning
37.518563 43.678562 following events. It’s titled ‘Agentive Language in Accident Investigations – Why Language Matters
43.678562 49.438562 in Learning from Events’. Published 2020 in ACS, Chemical, Health and Safety.
49.438562 55.758562 This is a discussion paper, so there’s a lot of content, so the presentation today will be
55.758562 61.038562 a little bit different to usual. People naturally want to know who or what was responsible for an
61.038562 66.558562 action, and particularly something undesirable. The assignment of action is called agency,
66.558562 71.758562 where an agent is a personal thing that takes an active role or produces a specified effect.
71.758562 78.638562 So our example about Bob spilling the chemical. Bob’s the agent of the action. This simplistic
78.638562 84.238563 sentence, however, doesn’t tell us whether Bob spilled the chemical intentionally, by accident,
84.238563 89.598562 or simply near the chemical when it spilled. Despite this, we are likely assuming that the
89.598562 95.438563 agent of the action acted independently, and had some sort of free will, choice to act.
95.438563 101.678562 These assumptions influence learning from events. The assignment of agency is subjective,
101.678562 107.118562 and it’s influenced by a host of things like the cultures, experience and the language of the observer.
107.118562 113.118562 Notably, linguistic framing of events has been shown to directly affect the assignment of guilt,
113.118562 118.878563 blame and punishments of human actors. The language of written action reports play a
118.878563 124.798563 role in identifying these agents and linking them to events. Linguistic shortcuts and biases are
124.798563 129.918563 said to reduce the understanding of context around an accident. They can artificially stop
129.918563 135.358563 the learning process and lead to a false sense of improved safety. Language is one of several
135.358563 141.518563 influences on beliefs of causality, but it does have a strong effect. As the paper unpacks,
141.518563 146.958562 prior work has suggested that behaviour can be attributed to a person’s internal characteristics
146.958562 153.518563 or their disposition. Yeah, like personality, abilities, attitude, or to outside influences,
153.518563 160.158562 you know, the culture, society, norms, peer pressure. When observing others, there’s a tendency
160.158562 165.678562 to attribute that person’s behaviour to internal causes. For example, a person’s disposition or
165.678562 170.958562 mental state, they’re a lazy person. This is called the fundamental attribution error,
170.958562 176.478563 where observers underestimate the external or situational influences on another person’s behaviour.
177.118562 182.638563 And overestimate the individual’s traits or attitudes. They’re a lazy person rather than
182.638563 188.478563 looking at why they might be looking for shortcuts. This effect makes it easier to judge a person’s
188.478563 193.278562 negative actions as coming from their own volition and disposition. Interestingly,
193.278562 199.838562 different cultures may assign different agency. Western societies are more likely to identify
199.838562 205.038562 people other than ourselves as the person with the agency, particularly for negative outcomes.
205.038562 210.798563 When evaluating ourselves, though, we typically focus on success and we distance ourselves from
210.798563 216.478563 failure, which is referred to as self-serving attribution bias. So for instance, when something
216.478563 220.878562 bad happens in our life, we attribute other factors. It’s because of someone else or
220.878562 225.358563 bad luck. We don’t attribute it to ourselves. English speakers have also been shown to have
225.358563 231.358563 an agentive bias which could affect action analysis. So English speakers are more likely to use
231.358563 236.718562 agentive descriptions for events as compared to some other language speakers. Being exposed
236.718562 241.998562 to undesirable events can trigger a defensive attribution bias where people defend themselves
241.998562 248.318562 from the concern of being seen as the cause or victim of a mishap. Hence, according to paper,
248.318562 254.158562 if we can categorise a serious action in some way as someone else’s fault, it’s reassuring
254.158562 258.878562 because we simply need to assure ourselves that we are a different kind of person from the victim,
258.878562 264.078563 or we would have behaved differently in that same situation. Researchers highlighted that
264.078563 268.718563 simple linguistic changes to an event’s description can affect the assignment of agency.
268.718563 274.718563 But, according to paper, most of our own agentive language variations are invisible to us.
274.718563 280.798563 Language priming, like agentive languages, for instance, “he crashed the car” or “none
280.798563 285.678562 agentive, the car crashed” have been studied in different ways. I can’t go through all of the
285.678562 292.158563 findings, but one example is our English speakers can better remember situations involving agentive
292.158563 297.758563 language compared to non-agentive. According to paper, placing attention on individuals involved
297.758563 302.718563 in accidents can actually improve memory for people when they recall in the accident. So people
302.718563 307.998563 are more likely to remember, “Bob crashed the car rather than the car crashed.” Another really
307.998563 314.558563 interesting study that was covered in here was examination of over 197,000 criminal trials in
314.558563 323.038563 London from between 1684 to 1913, found that the use of agentive language priming resulted in more
323.038563 329.198563 guilty verdicts. So, irrespective of the presented particulars and the facts of the case, the way
329.198563 334.958563 that the information was framed around agentive language influenced the guilty verdicts. So,
334.958563 341.038563 the language acts investigations. There is a need for linguistic shortcuts. We make things easier
341.038563 346.238563 for ourselves, but this can lead investigations towards binary oppositions where your pairs of
346.238563 351.758563 terms are seen as polar opposites like right and wrong, success and failure, good and bad. We
351.758563 357.918563 create these either or situations and they can limit learning. So, the success, failure, separation
357.918563 363.118563 is another binary commonality in safety, where the language is said to be rooted in theories
363.118563 368.878563 surrounding simple or complicated machine systems. But these machinery systems have a limited number
368.878563 373.678562 of parts that can break and function in my modal manners. You know, they can be either on or off,
373.678562 379.598563 working or not working, but humans can’t. Humans act in different and inherently complex dynamic
379.598563 385.278563 ways. So, we can really only ever say that a human failed or succeeded in hindsight once the outcome
385.278563 390.478563 of the action is known. This machine agentive language is often used to describe human action
390.478563 396.638563 when it really isn’t suitable for it. But from the research, it’s unlikely that a single objective
396.638563 402.798563 story exists around an event. Memories, experiences and language will differ when an event is retold
402.798563 408.878563 by different stakeholders. And written descriptions will be subjected to different sorts of linguistic
408.878563 413.918563 biases and shortcuts compared to people’s own recollections. Another element is active
413.918563 419.038563 versus passive verbs. So, small differences in language can have a large impact on causal
419.038563 425.038563 attributions like active verb voices. We covered some before, but Sarah hit the ball. That directs
425.038563 430.238563 agency to Sarah being the one that hit the ball. And this heightens attributions of control
430.238563 435.998563 compared to passive voice, like ball was hit. This attribution effect might still even remain
435.998563 440.878563 if the agent’s actions were presented as non-intentional. So, Sarah didn’t mean to hit the
440.878563 446.318563 ball. It can still be seen as more in control of the action. So, discussing the findings,
446.318563 451.998563 again, I’ve skipped a lot. Investigations may search for human failure and place significant
451.998563 456.478563 weight on human agency. Blame can be further influenced by the fundamental attribution
456.478563 461.758563 error where dispositional or personality-based qualities of people are prioritised over other
461.758563 467.438563 factors in the environment, the work, pressures, the leadership, all that stuff. Lots of other
467.438563 471.278563 studies have looked at blame. Here’s some other findings that weren’t directly covered here,
471.278563 475.918563 but I’ve actually covered on my site. So, one study, these are mostly from healthcare,
475.918563 480.558563 as is concept of blame conformity. This is where blame for an accident can be shifted
480.558563 485.678562 between individuals as a result of a leading eyewitness statement. So, they found that over
485.678562 489.998563 a third of participants who read an eyewitness statement that blamed the individual for the
489.998563 495.838563 accident went on to then blame the same person for its occurrence. So, it’s been well established that
495.838563 500.878563 when eyewitnesses encounter post-event information that contradicts or supplements
500.878563 505.518563 information contained within a witnessed event, they often incorporate this new information into
505.518563 510.798563 their own testimony. So, let’s say somebody observed an incident in the street and then if
510.798563 514.878563 they read someone else’s testimony, a written testimony, they can then tend to incorporate
514.878563 520.878563 that into their own visual eyewitness event. So, this priming is a relevant issue. In fact,
520.878563 526.558562 some investigation techniques actively control the flow of information to the investigators
526.558562 531.758563 to reduce that priming effect. They do their best to ensure that the investigators aren’t biased by
531.758563 535.838563 other information until they’ve had their own chance to sort of reflect on the information.
535.838563 540.878563 Another study from Healthcare found that when the investigation reports described
540.878563 547.198562 one or more contributing factors, the investigators were more likely to direct blame to the individuals.
547.198562 551.678562 And interestingly, this is something I wouldn’t have expected, that reports listing
551.678562 557.918563 organizational factors had higher odds of apportioning blame towards individuals. And the authors
557.918563 562.878563 really sort of suggest that perhaps those reporting incidents may still make fundamental
562.878563 568.718563 attributionaries, blaming individuals, despite recognizing there were systems factors at play.
568.718563 574.078562 So, it is a very human thing to blame. And this study also found that related to the
574.078562 578.878563 fundamental attribution error, it’s possible that the voices of those most directly connected the
578.878563 583.838563 events, the frontline operators who were involved in the event, aren’t having an active enough presence
583.838563 589.118563 in the post event learning process. And then another healthcare study that I covered on my site
589.118563 593.998563 also found that investigations that had higher word counts, you know, our larger reports,
593.998563 599.358563 or listed a communication related issue, were more likely to attribute blame,
599.358563 602.158563 which sort of seemed contradictory because you would think if they had more space,
602.158563 606.638563 they would have more space to talk about the context. But this study actually found the opposite.
606.638563 612.638563 The authors conject that perhaps having more space, you know, a larger word count allows more
612.638563 617.198562 opportunity to incorporate assumptions, hypotheses and value judgments, even if the
617.198562 622.318562 investigator didn’t intend it. So what can we make of the findings? Blame attributions
622.318562 627.438562 are a very human thing. So this won’t easily be changed, at least systematically or permanently.
627.438562 633.038563 Let’s first dispel the myth that investigations are an objective, fact-finding exercise.
633.038563 639.118563 They’re not. They involve humans. They can’t be. This isn’t a bug though. It’s just a feature.
639.118563 644.798563 So investigations are shaped by social factors as they are technical, but not always equally.
644.798563 649.118563 Now, first, the author argues that changing the assumptions of accident
649.118563 653.838563 causality and the analysis can have a profound impact on practices and
653.838563 659.038563 logics and belief systems. So the author says that you have to change the assumptions that drive
659.038563 664.318562 our belief systems. And language can be a very powerful vehicle to do this.
664.318562 669.678562 My own example is, take an investigation that says something like the operator failed to follow
669.678562 675.438562 the procedure. It leaves me thinking, and that’s not a finding. That’s the start of the sentence.
675.438562 681.518563 Why didn’t they follow the procedure? Was it known? Was it tested? Was it actually useful to them?
681.518563 686.718563 How’s the task normally done by that person or by their peers? Is the procedure normally used?
686.718563 692.318562 Why wasn’t it used this time? What was different? What makes following the procedure difficult?
692.958562 697.598563 Using the word “failed” over just reporting that they did not use the procedure
697.598563 701.358563 could be seen as injecting an unnecessary value judgment.
701.358563 707.998563 Indeed, saying the operator failed to follow the procedure is itself an example of agentive language
707.998563 713.998563 over something like the procedure was not followed. This isn’t a justify work around some procedural
713.998563 719.438562 departures. It’s about using those as critical insights about the mismatches between prescribed
719.438562 726.158563 systems and actual practice. Also, blame and attributions of error mask creators’ substantive
726.158563 731.678562 findings when they rarely are. Even things like situational awareness or complacency, according
731.678562 737.598563 to Holnagel, can take on the appearance of folk models. These are ideas which purport to explain
737.598563 742.718563 a real world phenomenon, but don’t really when you scratch below the surface. The author also
742.718563 747.118563 discusses one approach that’s been used successfully by them, is the learning team,
747.118563 750.878563 and they say that this can help map the interrelations between the systems factors.
750.878563 756.798563 But, of course, learning teams are still susceptible to the same very human heuristics and language.