By Charles C. W. Cooke
Tuesday, December 07, 2021
The Washington Post’s Dana Milbank believes
that he has hit upon a journalistic scandal. “After a honeymoon of slightly
positive coverage in the first three months of the year,” Milbank wrote last week, President “Biden’s press for the past four
months has been as bad as — and for a time worse than — the coverage Trump
received for the same four months of 2020.” This is a problem, he added,
because it has led to the “news media” giving “equal, if not slightly more
favorable, treatment to the authoritarians,” and thereby “serving as
accessories to the murder of democracy.” On CNN yesterday morning,
Milbank expanded his argument. “We see it as our job to be
negative, to be adversarial,” he said. “But there’s a real problem when we are
being just as adversarial ‘cuz a guy didn’t pass a bill as we are when a guy is
trying to overthrow democracy.”
If, upon reading this, you thought to yourself, “there is
no human being alive who could possibly have concluded that this is happening”
. . . well, then you were correct, because it turns out that no human being
alive did conclude that this is happening. Instead, Milbank’s evidence — which
he describes hilariously as “painstakingly assembled” “proof” — came from a
bunch of servers. “At my request,” he explained, “Forge.ai, a data analytics
unit of the information company FiscalNote
combed through more than 200,000
articles — tens of millions of words — from 65 news websites (newspapers,
network and cable news, political publications, news wires and more) to do a
“sentiment analysis” of coverage. Using algorithms that give weight to certain
adjectives based on their placement in the story, it rated the coverage Biden
received in the first 11 months of 2021 and the coverage President Donald Trump
got in the first 11 months of 2020.
Responding to Milbank’s conclusions, the statistician
Nate Silver noted drily that “the degree to which the extremely
nontransparent ‘AI’ analysis cited by Milbank should shift our priors” on this
question “is somewhere between zero and less than zero.” Silver is correct.
Milbank’s credulous talk may impress the partisan laymen, but the harsh truth
is that what he is selling here is closer to snake oil than to “artificial
intelligence” (itself a marketing term). In his piece, Milbank claims that
“artificial intelligence can now measure the negativity with precision.” But
this isn’t true — it can’t be. Human communication is extraordinarily complex,
and it remains the case the most sophisticated algorithms struggle to parse it
usefully. Absent heavy-handed intervention, “AI” is unable to comprehend
commonplace linguistic tools such as irony, sarcasm, cynicism, in-jokes,
callback-humor, and self-deprecation, and because they are so heavily
contextual, it is of extremely limited use when attempting to judge the tone or
scope of quotidian human sentiments. In a categorical sense, “she’s not a good
violinist” and “she’s the worst f***ing violin player in the world” are both
“negative.” But they do not represent the same critique. You know that. I do,
too. But does an undisclosed algorithm, utilizing a set of unrevealed input
variables, deployed by a researcher of unknown quality, at the behest of a
partisan journalist? Let’s say that the odds aren’t great.
Nothing about political language excludes it from these
structural deficiencies. “President Biden failed to implement his vaccine
mandate,” “President Trump is a failure and has left the country in an
unhealthy state,” and “Senator Dole’s health eventually failed him” are superficially
similar statements. But they do not amount to the same sentiments when
processed by nuanced human ears. Likewise, straight pieces documenting adverse
things that have happened to a political figure, recording
their poor polling numbers, or chronicling disagreements within their party,
are not intrinsically or deliberately “negative” in nature, but are all liable
to be interpreted as such by drones.
There is a reason that Milbank’s piece provoked such
hysterical laughter upon publication, and that reason was that its readers
were people, who were in possession of sophisticated brains, of
ears that can process distinctions, and of political memories that go back
earlier than January of this year. I am, as my readers know, both an enormous
devotee of technology and a self-confessed tinkerer (you can’t tell me to
“learn to code,” because I already can), but even I have not been so blinded by
the lights of science that I consider a glorified Ctrl + F program that has
been told to “give weight to certain adjectives based on their placement in the
story” to be more capable of grasping attitude, tone, scope, disposition,
proclivity, and sensibility than real-time, real-life human observers.
Does Forge.ai have a service that detects obvious
gaslighting?
No comments:
Post a Comment