It’s easy to see both the up- and
downsides of artificial intelligence.
Just a few upsides: more accurate
medical diagnosis, safe, fast automated vehicles, AI-driven instruction that
ever adapts its style and pace based on the student’s ongoing performance.
On the downside, luminaries such as Bill
Gates and Elon Musk worry that self-teaching AI computers could get smart
enough that humans won’t be able to stop them from nefarious ends.
It would seem that, per Stuart
Russell, author of Human Compatible, that we optimize risk/reward if we take
two steps: 1. Don’t let the computer “know” the goal of the software. 2. Block
the computer from making decisions beyond a certain magnitude—that when
implications of a decision go beyond a certain point, human override is
required. It’s kind of like the car salesperson who has discretion to give a 10
percent discount, but if more seems required, the boss must approve.
Another fear about AI is that an evil
individual or entity could use it to nefarious ends. A few examples: release a
murderous virus into the water supply, threaten to close down the electric grid
unless paid a zillion dollars, or develop an algorithm for manipulating people
into voting for Candidate X. (Whoops, that already pretty-much exists.) Of
course, most powerful things, notably nuclear energy, could be used
cataclysmically, yet most experts conclude that, rather than prohibit it, it’s
wiser to install in-computer and human oversight.
A similarly moderate stance could
apply to genomic research. On the upside it could better address such diseases
as cancer, diabetes, and cardiovascular disease and create drought-resistant-,
high-protein, and insect-repellant crops. Gene editing might eventually be used to create a super-intelligent human. Yes,
that person’s brainpower could be used for social good but what if the
gene-editing also caused him or her to have to live in physical pain? Or the
person could use the hyper-intelligence for personal gain even if it causes
great pain to the world. Again, it would seem that regulation, both built-in and legal might yield the risk/reward sweet spot.
What worries me is that such
restrictions may move such research to jurisdictions that have looser restrictions. For
example, the worldwide consensus has been that for now at least, gene editing be conducted only
for research, not clinically. He Jiankui defied that by using CRISPR to edit
the embryos of two recently born twin girls in what he said was an effort
to prevent them from contracting HIV. Russian scientist Denis Rebrikov
is planning to insert a gene into an embryo that would enable deaf couples to
produce hearing babies. Geneticist Bing Su has inserted a gene for brain
size, which is correlated moderately with intelligence. So, in this big world of ours, amid ever
advancing gene editing tools, it seems quite likely that, given the huge
stakes—even a country or non-state-actor creating an army of supergeniuses—that
restrictions would move some non-complying research underground.
En toto, it’s probably wise to
establish restrictions on AI: laws, professional standards, and, more
difficult, building-in limitations to AI software: forcing them to switch off
when the stakes are great or the implications unclear. Social norms and fear of
punishment will facilitate research that has a positive risk-reward ratio while
restraining less advisable research. Outright bans would likely yield worse
net results, as occurred when religion restricted scientific
research in the Dark Ages. It seems we must accept that the perfect is the
enemy of the good. Despite the likely excesses, it seems wise to bet on
humankind, that, net, we’ll probably derive positive effects from AI. It
certainly will be interesting to watch.
No comments:
Post a Comment