If you’re interested in machine learning, or, and especially if you have any involvement in big data, analytics, and related matters, before today is over you must read “Why scientific findings by AI can’t always be trusted,” (Maria Temming, Science News, Vol. 195, No. 7, 4/13/2019).
It describes research by Genevera Allen, a data scientist at Rice University, that attempts to answer a question asked in this space not long ago: With neural networks, which can’t explain their logic when presenting a conclusion, aren’t we just substituting trusting a machine’s guts for our own?
Allen’s conclusion: Yes, we are, and no, we shouldn’t.
Machine learning can, she says, be useful for providing preliminary results humans can later validate. “More exploratory algorithms that poke around datasets to find previously unknown patterns or relationships are very hard to verify,” she explains. “Deferring judgment to such autonomous systems may lead to faulty conclusions.”
Reinforcing the parallel with humans and their guts, Allen points out one of the more important limitations of machine learning: “… data-mining algorithms are designed to draw conclusions with no uncertainty.”
The people I know who trust their guts also seem to lack uncertainty.
Among those who should be less certain are those who figure the so-called “technological singularity” represents the biggest risk AI poses to humanity at large. The singularity — runaway AI where automated improvement cycles beget ever-more-advanced non-biological superintelligences — is the least of our concerns, for the simple reason that intelligence and motivation have little to do with each other.
To choose a banal example, Watson beat all human opponents at Jeopardy. We didn’t see a bunch of autonomous Watsons vying to become the next game-show contestants. Watson provided the ability; IBM’s researchers provided the motivation.
If we shouldn’t worry about the Singularity, what should concern us?
The answer: GPT-2 and, more broadly, the emerging technology of AI text generation.
And is so often the case, the danger doesn’t come from the technology itself. It comes from us pesky human beings who will, inevitably, use it for nefarious purposes.
This isn’t science fiction. The risk is now. Assuming you haven’t been living in a cave the past couple of years you know that Russian operatives deployed thousands of ‘bots across social media to influence the 2016 election by creating a Twitter echo chamber for opinions they wanted spread to audiences they considered vulnerable.
Now … add sophisticated text generation to these ‘bots capabilities.
You thought Photoshop was dangerous? Take it a step further: We already have the technology to convincingly CGI the faces of dead people onto living actors. What’s to stop a political campaign from using this technology to make it appear that their opponent gave a speech encouraging everyone to, say, embrace Satan as their lord and master?
Oh, and, by the way, as one of those who is or soon will be responsible for making your company more Digital,” it likely won’t be long before you find yourself figuring out whether, in this brave new world, it is more blessed to give than to receive. Because while less politically alarming, do you doubt your Marketing Department won’t want to be the last one on their block to have these new toys to play with?
The same technologies our geopolitical opponents have and will use to sell us their preferred candidates for office will undoubtedly help marketeers everywhere sell us their products and services.
How to solve this?
It’s quite certain prevention isn’t an option, although, as advocated in this space once or twice, we might hope for legislation restricting first amendment rights to actual human persons and not their technological agents, and, beyond that, explicitly limiting the subjects non-humans are allowed to speak about while also requiring all non-human messagers to clearly identify themselves as such.
We might also hope that, unlike the currently pitiful enforcement of the Do-Not-Call Implementation Act of 2003, enforcement of the Shut the ‘Bots Up Act of 2019 would be more vigorous.
Don’t hold your breath.
What might help at least a bit would be development of AI defenses for AI offenses.
Way back in 1997 I proposed that some independent authority should establish a Trusted Information Provider (TIP) certification that information consumers could use to decide which sources to rely on.
What we need now is like that, only using the same amplification techniques the bad guys are using. We need something a lot like spam filters and malware protection — products that use AI techniques to identify and warn users about ‘bot-authored content.
Of course, we’d then need some way to distinguish legitimate ‘bot-blocking software from phony alternatives.
Think of it as full employment for epistemologists.
A truly grim situation. The problem is as difficult as it is important.
I just don’t see an AI solution for two reasons:
1. Wittgenstein’s “the meaning is the use and the use is the meaning” says to me that there are no speech patterns that can’t be used to by intelligence bad actors to promote fascist and hate provocation text.
Not only words, but sometimes sentences have wholly different meanings between the same two people, as when to black people say, “He is a bad man”, which can be admiration for Malcolm X or disgust for Bill Cosby.
2. Godel’s Theorems on the incompleteness of language can easily guide bad actors to “logic holes” in any system of thought where the intended meaning gets twisted in the most toxic way at logically undecidable points.
AI systems are, after all, just algorithms, which necessarily have logic blind spots within them.
I think Facebook, et al, have to get out of the news business. Just stay with people you know, as one does with snail mail.
It may slow constructive revolutions, but unless and until it is shown possible for automated effective to controls to work, at least in theory, control must remain in our human hands only.
It seems our survival is at stake.
Agree, though it’s not just the scale and speed that’s most troubling. Consider the threshold at which the “A” in AI behind such systems transcends from being an “artificial intelligence” to an “autonomous” one. (I’ll accept groans and laughter for the pun on self-driving vehicles.)
Let the “Battle of the Bots” begin!
I like the concept of TIP, and it can work like the certificate authority (CA) used to support SSL and secured Internet communication these days. Perhaps, in addition, we can apply technologies like blockchain or immutable ledger to further strengthen the authenticity of the news/messages. With the help of CA and Blockchain, the machine learning algorithms just might be able to help us by classifying the news correctly and detecting fraud.
https://www.youtube.com/watch?v=DIZf7eRlD4w
Thanks for sharing the link. As I said, this isn’t science fiction. This is today’s news (and tomorrow’s even faker news).
I remember from the olden-times when IBM was still a major factor in the computer market. I am in the Process Control profession. In the 1960’s IBM wanted to enter this market using their “software expertise.” We had an IBM 1800 connected to a paper machine, but of course control of the paper-making process was a giant unknown. IBM’s crew of PhD’s told us That all we had to do was take lots of data sets over a long period of time, then use regression analysis to build a process model. An early approach to data mining, but taking lots of data will NOT “find” correlated data. You need to perturb the process manipulated variables in order to find the effect on the measured variables. This is why it is necessary to have placebos in double-blind clinical drug trials. Without a statistically designed experiment, you cannot determine causation or correlation. Another view – at lunchtime in most restaurants, notice that most of the overweight people are eating salads, while the skinny people eat burgers. Conclusion – salads make you fat. Correlation does not reveal causation.