Five weeks ago, I published a post about an artificial intelligence backlash, or reactions against it from various directions. Since then, when we factor out the news stories speculating about AI’s effects on employment, telling us about new or not-so-new robotics advances, and subjective investment advice, there hasn’t been much good, and it adds up to something worse.
First, as shown in “CheatGPT,” by Aki Ito in Business
Insider on August 1st, we have workers using AI against company
rules, taking advantage of lagging or overly cautious policies as well as thoughtfully
reasoned ones. That is really an issue
between employer and employee but could turn out to cause the former real
problems when accuracy is not what it seems.
Such isn’t rare, as “When Hackers Descended to Test A.I., They Found
Flaws Aplenty” (Sarah Kessler and Tiffany Hsu, The New York Times,
August 16th). At the Defcon
hacker’s conference this month, participants “tried to break through the
safeguards of various A.I. programs in an effort to identify their
vulnerabilities.” They were able to get
the software to make racist and discriminatory statements, select sample job
applicants by Indian caste, and write factually about nonexistent people,
places, and constitutional amendments.
Efforts like these, in this case strictly benevolent, are invaluable, with
experts revealing previously unknown flaws – the bad news is that those
shortcomings are real.
Is, or will it, be true that “AI is ruining the internet”
(Arantza Pena Popo, Business Insider, August 8th)? The programs are now able to defeat many of
the human-verifying “captchas,” causing sites to make theirs so difficult many
actual people cannot get through. A
Europol study estimated that within “a few years,” ninety percent “of internet
content” will be produced by AI systems and will contain significant
errors. That could result in online
resources seeming “engineered for the machines and by the machines,” its value could
greatly decrease, and it may in turn poison future AI improvements.
The most telling of these articles, though, is “Digging for
digits” (The Economist, August 19th). It told us about how the world is running out
of accessible, non-copyrighted data, for which companies are competing and
paying increasing amounts to use. Some
of that will be inputs into ChatGPT 5, which may be released before the end of
the year.
But how about ChatGPT 6?
If we see that one someday, it may not be fundamentally better. Why?
Although conceptually it could be boundless, artificial
intelligence is being stopped on several different fronts. First is copyright problems, with
unauthorized incorporation of legally protected work well integrated into
ChatGPT 4, and certain to result in further lawsuits and massive settlements,
with a chance that its entire knowledge set could need to be erased and
rebuilt. Second is chip shortages –
developing large language models at contemporary strengths consumes more
silicon than it can get. Third is calls
for more regulation, which, with governmental authorities perennially behind
the curve with technical issues, could impede or even halt development,
production, or both. Fourth is an
upcoming data shortage as above.
The fifth limit, and perhaps the ultimate, is money. ChatGPT 5 will cost in the billions of
dollars. Since growth is exponential,
there is a real chance that ChatGPT 6 or its equivalent would run in the
hundreds of billions. Even in as
future-oriented a field as information technology, companies, venture
capitalists, and other lenders and resource providers want results quickly and
will see the other four problems.
Six years ago, driverless cars seemed poised to take over
world roads within a generation. Now,
although they have done well in useful if controversial niches such as taxicab
travel, and their technology is turning up more and more in human-driven autos,
their overall development has apparently stalled, with little press about them
becoming the norm. A serious around-1970
prediction said that within two decades artificial hearts would be, after cars,
the second largest American industry – it is now possible to get one, but its
implementations are measured in dozens.
The same may happen with artificial intelligence. It may find its role in business writing, and
as a lower-end substitute for human effort.
By 2040, people may laugh about the idea that this business tool could
take over the world, and compare its dangers to those of copiers. Or not.
Will artificial intelligence stay way short of its potential? I don’t know, but we need to seriously
consider that possibility.