What emotions would AIs need to feel?

[this post is com­pletely fluffy and su­perfi­cial]

When talk­ing about AIs, the is­sue of emo­tions of­ten comes up. I nor­mally fi­nesse the ques­tion by say­ing that we don’t re­ally know if AIs would have things we would recog­nise and emo­tions, and would you stop pes­ter­ing me, I have some se­ri­ous and scary points to make.

But… If we take the ques­tion se­ri­ously for a sec­ond, and al­low our­selves to spec­u­late… The main an­gle of at­tack is to won­der: what are emo­tions for, what func­tions do they serve? And would they serve the same func­tions in an AI?

So, here’s a spec­u­la­tive cross-sec­tion of emo­tions, and whether AIs are likely to have ver­sions of them:

  • Very likely: Joy, Sad­ness, Bore­dom, Em­pa­thy, Interest

At their sim­plest, joy and sad­ness are re­sponses to the situ­a­tion turn­ing out bet­ter or worse than ex­pected. Th­ese are es­sen­tial to learn­ing, and the AI would cer­tainly have some ver­sion of them. Similarly, bore­dom is a sign that the AI should do more ex­plo­ra­tion of its en­vi­ron­ment, em­pa­thy is use­ful for un­der­stand­ing other agents, and in­ter­est is a sign that a cer­tain area should be in­ves­ti­gated in pri­or­ity.

  • Pos­si­ble, if the AI is limited or non-trans­par­ent: Anger, Kind­ness, Cru­elty, Trust, Distrust, Surprise

If the AI is able to cred­ibly pre-com­mit to cer­tain ac­tions (eg if at­tacked, it will re­tal­i­ate), then some of these emo­tions won’t be needed. But it might not be able to cred­ibly sig­nal that, in which case anger would be use­ful: other agents will be less in­clined to ex­ploit it, lest it re­act ag­gres­sively (note that giv­ing in to anger is gen­er­ally a dis­ad­van­tage, but be­ing known to give in to anger is of­ten an ad­van­tage). Kind­ness and cru­elty are similarly tools for man­ag­ing so­cial in­ter­ac­tions; if the AI is trans­par­ent and can pre-com­mit, they won’t be needed.

Em­pa­thy is the key AI emo­tion, but it’s pos­si­ble that it may award spe­cial cat­e­gories for other agents with es­pe­cially pos­i­tive or nega­tive traits. Sur­prise, fi­nally, is a sig­nal that un­likely events are hap­pen­ing, and a re-as­sess­ment is needed; if the AI can­not im­ple­ment its as­sess­ments and re­assess­ments in a smooth way, it may need some thresh­old of sur­prise to trig­ger a re­ac­tion.

  • Pos­si­ble, if the AI has a phys­i­cal body: Fear, Dis­gust, Anx­iety, Feel­ings of Security

Th­ese emo­tions are all con­nected with pro­tect­ing our own phys­i­cal body, so an AI would only feel them if it had a phys­i­cal body it­self, and couldn’t just think through the is­sues dis­pas­sion­ately.

  • Un­likely for an AI to feel: Love, Shame, Pride, Envy, Indig­na­tion, Ex­cite­ment, Pity

Th­ese are very spe­cific to the hu­man con­di­tion, and the hu­man so­cial cir­cum­stances, so they are un­likely to be felt by generic AIs. Now, it’s always pos­si­ble for AIs to feel these, but there’s no nec­es­sary rea­son for them to feel them, as far as we known now.

Fi­nally, we should con­sider that AI’s could have emo­tions that we never had, be­cause hu­mans were never in cir­cum­stances where we would use them. What do you think these might be like? I’ll pre­sent two spec­u­la­tive ideas: first of all, AIs should have an in­tu­itive grasp of statis­tics and prob­a­bil­ity that is much bet­ter than our own, so they may have statis­ti­cal in­stincts. Se­condly, if the AI gets copied a lot, self-loy­alty – loy­alty to copies and slightly-di­ver­gent-copies – may be­come use­ful, and in­stinc­tive ways of re­solv­ing de­bates with it­self may be­come com­mon.