Tag Archives: Kathy Pretz

Machine Learning

Sensei Jordan gets grumpy about the term “AI”

1 Comment

As I have remarked, the term “Artificial Intelligence” has long been loosely defined.  While originally coined to describe the idea of creating human like “intelligence” in computers, the term has been an umbrella for any technology that replicates (or exceeds)  human level performance in any fragment of human behavior:  speech and text recognition, for example, or visual analysis, or locomotion (i.e., robotics).  Anything that’s “magical” is usually called AI, at least until it becomes ubiquitous (and therefore less magical).

The thing is, these technologies, however magical and clever, aren’t particularly “intelligent” or human-like.  In fact, some are definitely not human-like, such as Markov based speech recognition, which works amazingly well, but has nothing to do with how humans do it—probably.

(And don’t get be started on what the term “intelligence” might mean—that’s another giant semantic mess.)

This spring Kathy Pretz discusses these issues with old gray-head Sensei, Michael I. Jordan [1]. He’s been doing this stuff for a long time, so he knows what he is talking about.  In particular, there is a lot of technology (which he helped invent) that is clever about pattern recognition, and, of course, a who galaxy of machine learning techniques.  These technologies aren’t especially “intelligent”, and aren’t really even supposed to compare to human behavior.

So “Stop Calling Everything AI” he says!

“People are getting confused about the meaning of AI in discussions of technology trends—that there is some kind of intelligent thought in computers that is responsible for the progress and which is competing with humans,” he says. “We don’t have that, but people are talking as if we do.”

“While the science-fiction discussions about AI and super intelligence are fun, they are a distraction,” he says. “There’s not been enough focus on the real problem, which is building planetary-scale machine learning–based systems that actually work, deliver value to humans, and do not amplify inequities.”

(Quoted in [1])

My own view is that the term AI has long ago escaped from its technical and academic origin, now running free in the wider culture where “AI” means “magic”.  As Humpty Dumpty said in Through the Looking Glass, “When I use a word, it means just what I choose it to mean.” 

It’s not really a mystery why AI has become a staple of advertising and popular fiction .  It is used to evoke—and embody–fantasies, both pleasant and unpleasant. Essentially technologically realized angels and demons.

Sigh.

So, Sensei Jordan is right, but there is nothing that can be done. I think the only way to cut down on the abuse of the term is to come up with a replacement term.  Which I don’t know how to do.

  1. Kathy Pretz, Stop Calling Everything AI, Machine-Learning Pioneer Says, in IEEE Spectrum – IEEE Member News, March 31, 2021. https://spectrum.ieee.org/the-institute/ieee-member-news/stop-calling-everything-ai-machinelearning-pioneer-says
Linked Data, Machine Learning, Sociotechnical

New IEEE Standard to Rate the Trustworthiness of News Sites(?)

Leave a comment

There has been a lot of chitter-chatter about fake news and the general problem of the untrustworthiness of digital information.  The global scale and relatively unmediated nature of digital communications has amplified perfectly normal human weaknesses (e.g., not listening carefully) into unpleasant and possibly dangerous problems at a vast scale.

As often happens with the Internet, some people are sure that with the right technical fix, we can take care of these deep sociotechnical challenges.  In the case of fake news, there have been a variety of proposals for algorithms that “detect” iffy content, and for rating systems to manually or automatically grade and filter information.  The idea seems to be for people to be able to set criteria for what they trust, and be able to easily ignore or block information that doesn’t meet these criteria.

As far as I know, none of these proposals has been shown to work.  Given the nature of human thought and communication, I’d be surprised if anything like this actually works.  (It’s an open question whether this would be a good idea.)

Imagine my surprise to see the announcement of a new working group, intending to create a technical standard for this technology! [2]  Call me old fashioned, but I generally don’t bother writing standards for stuff that doesn’t work and doesn’t even exist yet.

IEEE P7011: Standard for the Process of Identifying and Rating the Trustworthiness of News Sources

So what in the world could be in this “standard”?

My first thought is that one feature to trust is Provenance.  There has been considerable work on this (which I contributed to, some twenty years ago), and much of it could apply to news sources.  Have a reliable and standard way to identify the actual source of information would be useful.

Of course, there already are standards for Provenance, as well as deep research on the automation of this reasoning about sources.

But it turns out that the P7011 group is much, much more ambitious than just provenance.  They want to standardize methods for detecting/flagging.  These include:

  • Factual accuracy of spot-checked articles 

  • Degree and consistency of bias 

  • Usage of misleading headlines 

  • Existence and utilization of effective retraction policies and procedures 

  • Clear distinctions between advertisements and content
    (from [1], p.64)

Wow!

These ideas are explained in a paper, “Addressing fake news: Open standards & easy identification” [1].

Boldly, the proposal hopes to identify algorithms for analyzing text to identify “moral, bias, and functional risks”.  The idea is to analyze (a sample) of “news stories”, and provide statistical tests of the language which are, they hope, indicative of conflict of interest, misrepresentation, or other deviations.  The proposal also includes human cross checking to maintain external validity of these tests.

What could possibly go wrong?

First of all, I have to question the fundamental goal, “This solution should build an easy way to help the public begin to differentiate between real news and fake news purveyors.” ([1], p.67)   Does “the public” (whoever that is) want to “differentiate”?  Or do they just want to listen to amiable sources that they agree with?  Even if all the goals of this standard were met, I have no reason to think that people would use it.

Second, they project rests on a semantic quicksand of undefined and probably undefinable terms.  “Bias”. “Misleading”. “Factual accuracy”.  What is the definition of “fake news”, and how is it different from “real news”?

These concepts are difficult to define, and certainly not universally agreed upon.  How can you make a tool to measure something that we can’t define?  Worse, a tool that uses one definition of is worse than useless if I disagree with that definition .  A standard I disagree with then contributing to “bias” by labeling it in a “standard” way that I disagree with.

Third, it seems to me that the idea of an industry supported standard will mostly give cover to the large platforms who are making money off of this mess.  Facebook can point to IEEE P7011 and say, “look how virtuous we are, we have fully implemented the international standard that fixes this problem.”  Whether it works or not (and it surely is unlikely to do anything useful), this helps the big companies deflect responsibility.

(And wait until governments start mandating the use of this standard–and other governments ban its use.)

Is there any possibility that this will work?  I’d rate the chances as slim at best.  One problem with an open standard is that it will be, well, open.  Trolls and advertisers will game the system.

Let me make some suggestions.

First of all, P7011 should have a very careful look at the work on Provenance and Provenance standards and the Semantic Web in general.  These are hard problems, so pay attention to what has already been learned.

Second, the P7011 should have a sub group for Metrics.  How can you tell if this system is actually working correctly?  Is the goal to have happy users who “like” your app?  Or is the goal to change what information people use?  Or maybe the goal is to change the effects of getting information from the Internet.  Whatever the goal, how can we measure the results?

This process will be difficult because much of the imagined technology is circular.  “Bias” is defined by someone, an algorithm is taught to flag that kind of “bias”, and the problem is solved.  But is this really doing the right thing?  And according to who?  How can we show that it is working?

For that matter, the working group needs to define what success would be for the WG.  Just what is the goal of this standard creation activity?  Write some specifications? Demonstrate some technology? Deploy a prototype?  Create a sustainable software suite?

There is nothing wrong with trying and failing, but no one should spend time and effort on something that can never succeed or fail because it has no clear goal.

Personally, I think P7011 is going to go nowhere.

  1. J. Hyman. Addressing fake news: Open standards & easy identification. In 2017 IEEE 8th Annual Ubiquitous Computing, Electronics and Mobile Communication Conference (UEMCON), 2017, 63-69. http://ieeexplore.ieee.org/document/8248986/
  2. Kathy Pretz, IEEE Standard to Rate the Trustworthiness of News Sites: Automated rating system to analyze factual accuracy, headlines, and more, in The Institute. 2018. http://theinstitute.ieee.org/resources/standards/ieee-standard-to-rate-the-trustworthiness-of-news-sites