Turing's Razor

Robot ethics should be an extension of human ethics; and why that means we should urge programmers to simply neglect implementing trade-offs inspired by the trolley dilemma thought experiments.

Some months ago I wrote about robot vs. human ethics:

• Turing’s Mirror: How robots make us look into the mirror. To realize that ethical robot questions are really centered around our own morals. Reflecting our own continued failure to lead by example in a landscape shaped by power imbalance.
• Turing’s Duck: How the essence of the imitation game equally applies to ourselves. How we too just learn to pretend and to conform to expectations. Leaving ample ground for our infamous hubris.

Since then I've been struggling with this third one.  Pondering around what emerges like some guiding principle in these topics. Such "decision guides" deserve being formulated as ‘razors’, like the infamous ones from Occam or Hanlon.  And being the origin of my inspiration, the obvious choice is to name this one after hero Alan. All while happily giving in to the stylish fit of making a sequence of three 'Turing' blog titles :)

Image credits : 

Jeff Osier-Mixon. CC BY-SA 4.0

Here is what I’m calling “Turing’s Razor” :

Do not ask robots to achieve ethical norms no human needs to adhere to.

Some derived versions spontaneously sound like:

Do not blame on technology what humanity has successfully mismanaged for ages.

Don't attribute to technology those moral failures we humans couldn't resolve (without it).

And it leads to this obligatory corollary:

Stop it already with endlessly letting those meat-brains get off the hook when artificial brains show a path to doing it better.

With one eye on this shiny new compass, we should be able to navigate this modern age landscape, avoiding the steep cliffs of false ethical questions.

-oOo-

No way back

Before getting deeper though, it should be clear that there is simply no way to avoid thinking about these issues.

Take for instance the discussion on DNA Databases.  Obviously this gets condemned as too strong a violation on our public rights, freedom and privacy. Still, there should be room for the careful progressive approach: if current state of legislation, practice and lack of transparency urges us to strongly oppose, than do so, but also start thinking about the right way to do it. You want to be ready for that time when this does become allowed practice: did you spend all your energy strictly at opposing, or did you, in the meantime, also craft up the process to correctly supervise the usage? Principles are useful, less so: being blindly principled about them. Finding a good set of regulations to control possible but avoidable mistakes really is due.

The cry to return to happier times when these questions played no role is a false one: There is simply no ‘undo’ to progress. Once escaped from pandora’s box we just have to reckon with every possible fringe abuse of even the most benign invention.
The other extreme is equally untenable: “To abandon all regulation so not to inhibit 'Future' itself along its disruptive path.” No, we can do better than just hoping for the best. Guard the balanced outcome for all involved.

Both extreme visions will accuse the other side of being overly naive, while neglecting that both are entangled in a race in naivety. The useful cases are born in the same nest as their evil twins, the fruitful trees seeded amid the tares: careful weeding is a balancing act requiring attention. Controlling the usage must be separated from prohibiting the technology. The boundary between the beneficial and the dangerous is too thin to use anything else than the sharpest lancet. So, lay aside your sledgehammers.

We have been here before, at least if you can believe Ash Carter in "What I Learned from the People who Built the Atom Bomb"

I'll give you a topic...

Checking how well this new razor actually shaves requires searching for a suitable beard: What about lethal accidents of self driving cars?

This topic has been covered more than just a bit. Quite a challenge to find anything to add to what is already in the mix. Some references to get up to speed:

• Philosophy Tube | The Ethics of Self-Driving Cars - Olly Lennard (video)  
• TED ed | The ethical dilemma of self-driving cars - Patrick Lin (video)  
• WiReD | To save the most lives, deploy (imperfect) self-driving cars asap - Aarian Marshall (article) 
• TEDxCambridge | The Social Dilemma Of Driverless Cars - Iyad Rahwan (video) 
• Radio1 | Interne Keuken van 17 februari met Maarten Boudry (podcast / dutch)  
• RadioLab | Driverless dilemma (podcast) 

Some of those reference the trolley dilemmas; some take that down to the grim boundaries of "counting lives" (which made me think of this old Chernobyl documentary); some succeed in considering this as non-harmful practice, even beneficial and necessary, as it resembles (almost) the practice of medical triage; some strongly aim for the pragmatic "don't let the perfect be the enemy of the good"; some call upon society to "at least ask the right questions?'; some just have the Frankenstein Reflex written all over (term coined by Isaac Asimov); some arrive at the conclusion that "robot ethics should be a natural extension of human ethics". But none of them listened to grandma.

Best effort

Grandma assured us over and over: "It never hurts to do your very best." Applying Turing's razor we should investigate first what our reaction is to failing human drivers. Particularly those who pass grandma's test with flying colors (i.e. didn't do drinks nor drugs, slept enough, kept themselves and their car in shape, adapted speed to circumstances, kept attention on the road, not on media device, etc... ) If we can allow some of those to actually get off the hook, and be pardoned by admitting there was no possible better outcome; that whatever horrific turnout can be mildly covered under the cloak of "just bad luck", … if so, then why would we not be able to come to the same conclusion for robots? Or their programmers?

I admit that pulling in the trolley dilemmas starts from a defensible position: after all: good prevention, like good robust programming is all about considering "what if" scenarios to a large extend.  Passed that initial outlook of logic it took me quite some time to see how this line of thought unwittingly opens the door for something else to sneak in. Two things in fact.

1. for the robot: We silently extend the check-sheet for evaluating his 'very best' driving skills.
2. for the coder: We ignore that doing his 'very best' plays in a different realm.

Doing your best driving around safely does not overlap at all with being the best moral judge in split-second valuation of fellow human beings.  If it would, getting a driver's license would include obligatory passing university classes in advanced ethics, moral judgement and human potential assessment. Why do we now add this in the mix? Is this just because we (think the robots) can? Are we so satisfied with predicted safety levels of self-driving cars that we think it is (ever) good enough? So good at least that doing your best at pure safety and avoidance will reach some limit, after which it gets redefined to finding the best trade-off? Or is that giving in to mediocrity? Let me tell you this: Judge "grandma" will be less than pleased with a robot explaining it dedicated part of its processing power to evaluating and counting lives rather than strictly avoiding all possible harm to the fullest of its ability.

But where to shift blame? Since we don't know yet how to prosecute and sentence robots themselves, the programmers suddenly show up in the crosshairs. They can make judgments, right? They can negotiate trade-offs. They can agree on a set of moral rules. They too have to do their best. I agree, but we should notice that they play on a different level.  They're not the ones driving, they are coding. Coding is policy making: defining the rules of engagement that will eventually be executed. We have a different set of standards for those.
The trolley-equivalent of their role is that of the railway-security-engineering team, and it is simply not present in the trolley dilemmas. The questions they are solving would be: How could any person ever end up on a track where trolleys pass? How do we avoid that unsupervised bystanders are operating our switches? How to design security fences on all bridges to keep (even fat) men to drop down from them? Admittedly, the execution of this job does involve trade-offs, balancing and what-if scenarios. But doing your 'very best' in this realm assumes neutrality, removing all possible personal bias, extracting generic strategies out of specific cases, aiming for statistical overall best performance. It too refrains from valuing individuals and counting lives. Acceptable policies are as blind as Lady Justice. Good performing policies are designed to do equally well regardless of the person involved.

Thus programmers are relieved from coding in trolley-like (case-based) ethical decisions, and the robots are relieved from negotiating which people get to live. Their common guiding principle becomes the one we've known for ages: "do your very best to do no harm". Zero tolerance on making any victims.  Such principle becomes a clear instruction towards programmers, one that should get them to feel proud about any invention they add towards that goal. More beer and pizza to them!

-oOo-

Bad luck

If we all keep doing our best the failure count will go down. Nevertheless, remaining cases will exist. And so returns the notion of "bad luck".

Most likely this will hit hard with the paradox of "totally unacceptable improbability". The lower the chance that something bad happens to me, the more unacceptable that becomes. Why me? This must be intentional! Who is behind this?  The better these robotic drivers become, the less prepared we will be to accept the remaining fatalities as ill-fated.
Consolation (probably less so for those directly involved) comes from knowing that these versions of "bad luck" no longer should be "in vain".  Each specific failure will end up being analysed and turned into fixing the policy - I meant the code -  when ready: instantly deployed, not to happen again.

We know how to do this, since we have such procedures in place for dealing with accidents in air traffic.  Whatever the reasoning behind it: We seem to apply these analyse-to-remediate procedures to larger extend as the transportation mode is more automated or mechanic. Considering Turing's razor we should question this strange, unmotivated privilege towards those ape descendants.

One more thought, in a desperate attempt to avoid receiving the label of devout dataïst or naive futuroptimist.

Maybe, just maybe, the only way for us humans to really trust these robots, is for them to be as failure-prone as we are?

-oOo-

This article is part of a series: Turing's mirror - Turing's Duck - Turing's Razor

update: I came across this quote attributed to Arthur C. Clarke. It says it all.

One cannot have superior science, and inferior morals.
The combination is unstable and self-destroying.

– Arthur C. Clarke

update: Found this project "The Moral Machine" at MIT for measuring what humans generally think should be done in certain traffic-dilemma cases.  While this is a great idea for measuring how humans think about these cases, it remains totally unclear why that corpus of facts would be a good/reliable source or guiding design principle for actually engineering "automated" safety.  We should aim higher then opt for that variant of failure that will be mostly acceptable by humans. We should stop thinking like lawyers about the blame shifting, and more (like engineers?) about doing things better. I still think Arthur's quote is right on, just make sure your morals team understands the science.