Robots Should Not be Equipped with Human-like Emotions
Next: Remarks
Up: Formalized Self-Knowledge
Previous: Observing its Motivations
Some authors, e.g. Sloman and Croucher [Sloman and Croucher, 1981], have
argued that sufficiently intelligent robots would automatically have
emotions somewhat like those of humans. We argue that it is possible
to give robots human-like emotions, but it would require a special
effort. Moreover, it would be a bad idea if we want to use them as
servants. In order to make this argument, it is necessary to assume
something, as little as possible, about human emotions. Here are some
points.
- Human reasoning operates primarily on the collection of ideas of
which the person is immediately conscious.
- Other ideas are in the
background and come into consciousness by various processes.
- Because reasoning is so often nonmonotonic, conclusions can be
reached on the basis of the ideas in consciousness that would not be
reached if certain additional ideas were also in consciousness.
- Human emotions influence human thought by influencing what
ideas come into consciousness. For example, anger brings into
consciousness ideas about the target of anger and also about ways
of attacking this target.
- Human emotions are strongly related to blood chemistry.
Hormones and neurotransmitters belong to the same family of
substances. The sight of something frightening puts certain
substances in our blood streams, and these substances may reduce the
thresholds of synapses where the dendrites have receptors for these
substances.
- A design that uses environmental or internal stimuli to bring
whole classes of ideas into consciousness is entirely appropriate
for a lower animals. We inherit this mechanism from our animal
ancestors.
- According to these notions, paranoia, schizophrenia, depression
and other mental illnesses would involve malfunctions of the
chemical mechanisms that bring ideas into consciousness. A
paranoid who believes the Mafia or the CIA is after him and
acts accordingly can lose these ideas when he takes his medicine and
regain them when he stops. Certainly his blood chemistry cannot
encode complicated paranoid theories, but they can bring ideas about
threats from wherever or however they are stored.
These facts suggest the following design considerations.
- We don't want robots to bring ideas into consciousness in an
uncontrolled way. Robots that are to react against people (say)
considered harmful, should include such reactions in their goal
structures and prioritize them together with other goals. Indeed we
humans advise ourselves to react rationally to danger, insult and
injury. ``Panic'' is our name for reacting directly to perceptions
of danger rather than rationally.
- Putting such a mechanism in a robot is certainly feasible. It
could be done by maintaining some numerical variables, e.g. level of
fear, in the system and making the mechanism that brings sentences
into consciousness (short term memory) depend on these variables.
However, human-like emotional structures are not an automatic
byproduct of human-level intelligence.
- It is also practically important to avoid making robots that are
reasonable targets for either human sympathy or dislike. If robots
are visibly sad, bored or angry, humans, starting with children,
will react to them as persons. Then they would very likely come to
occupy some status in human society. Human society is complicated
enough already.
Next: Remarks
Up: Formalized Self-Knowledge
Previous: Observing its Motivations
John McCarthy
Thu May 25 00:33:25 PDT 1995