Shoulder devil, Shoulder angel and Gut Instincts.

What would you do 

if you walked into a murder in progress?
  

The 'Elevator Murder Experiment' simulated the situation. 

  

Results

The 29 individuals (and 2 dogs) shown in this clip are faced with a tough and fast choice to make. 11 of them instinctively run away as soon as they witness the dangerous situation. 12 of them (and the 2 pets) take action: by shouting at the murderer, calling for help or attacking the murderer. As for the last 6 individuals, they more or less hesitate or interact with the victim before deciding to run/walk away.

Note: this video currently being used for the 'Dead Man Down' movie's promotional activities, its results can powerfully represent the diversity of reactions - but not their probability.

Analysis
Instincts are inherited rather than learned behaviours, non-rational and innate ways to respond to a simuli often associated with animal behaviour. If this was the main drive for the experiment subjects' behaviour, how come not all of them - including the dogs themselves - run away when witnessing the dangerous situation?

Individuals who stopped and analysed the situation did not necessarily engage into a profound and rational processing of information. It is most likely that in this "uncomfortable" psychological confrontation of "inconsistent" ideas, they relied on time efficient mental shortcuts. 
Subjects are pressured to "reduce or eliminate cognitive dissonance". Three strategies arise: (1) lower the importance of one discordant factor (2) add consistent elements, and/or (3) change one of the dissonant factors. (Festinger, 1957) 

In this case, the dissonant factors are - risking your life and helping a stranger (ALTRUIST SCENARIO) or saving your life and running away (SELFISH SCENARIO). Examples of how the experiment's subjects could have attempted to reduce cognitive dissonance and choose a path of action are listed below:

ALTRUIST SCENARIO

• Selfish scenario seems less tempting thanks to the ability to call for emergency and fullfil the altruistic scenario (1:05 - calling 911, 0:39 - shouting for help)
• Altruist scenario is supported by strategic advantage over the murderer (e.g. 0:55 - using a fire extinguisher, 0:40 - hitting with a newspaper)
• Altruist scenario is easier to perform because the subjects are more numerous than the murderer alone: the situation seems no longer life-threatening (e.g. 0:33 - group of two).

SELFISH SCENARIO

• The altruist scenario is less important because the subject doesn't personally know the victim (all cases)
• The selfish scenario seems better because the subject feels in no position to help. As such, the altruistic ambition to save one's life is not valid anymore. This is particularly true when subjects talk with the victim before walking away (e.g. 1:20 and 1:26)
• Sense of duty excuses the subject from performing the altruist scenario (e.g. 0:30 - men delivering a couch).

Those potential explanations do not necessarily imply one course of action or the other. For example, the superiority in number did not consistently lead to an altruist behaviour (e.g. 0:55 - three subjects walking away).

Helping the victim depends on how many excuses you can think of when confronted with the dangerous situation. 

You cannot predict this in advance.

References
Festinger, L. (1957). A theory of cognitive dissonance. Stanford, CA: Stanford University Press. pp.3,  pp.18.