uspol, machine learning 

lawsuit.org/robotrump-an-ai-tr it's been possible to lie about things in writing, at scale, for a very long time. this blog post (like many others on the same topic) doesn't bother to imagine what *actual application* of these language models that will actually make the situation worse. consequently the post (like many others) comes off as aimless fearmongering

Follow

uspol, machine learning 

the actual potentially damaging effect of blog posts like this is that they leave unquestioned (and therefore reinforce) the "common sense" assumption that the trustworthiness of text somehow inheres in the text itself—i.e., that you can determine the authorship and intention of a text by looking at the text. this is very wrong, and reinforcing this belief makes it easier to "trick" people using spam, astroturfing and yeah, computer-generated text

uspol, machine learning 

using Trump's speeches is an especially ironic choice to make their point, since his entire rhetorical style is about saying outlandish, improbable, self-contradictory things just to get a rise out of people—his speech is already almost entirely phatic & meaning-free (you could replace any of his speeches with "I'm here, I'm loud, I'm cruel!" and little would change). like computer-generated spam, it's a good example of language that exists for its effect, not its meaning

Show thread

uspol, machine learning 

(Sarah Palin's speech style was very similar—outlandish, cruel, self-consciously "folksy," oftentimes incoherent—and it was often mocked at the time with language models, though it was markov chains in 2008 and not whatever the belching transformer model du jour is in 2019)

Show thread

uspol, machine learning 

@aparrish

(you could replace any of his speeches with "I'm here, I'm loud, I'm cruel!" and little would change)

thank you
that is the most concise and accurate summary of those rants i have yet encountered

🤬

uspol, machine learning 

@aparrish I also find it puzzling that these pieces never address the fact that it is quite cheap to pay humans to write those texts, regardless of the availability of sophisticated and "dangerous" AI.

Sign in to participate in the conversation
Friend Camp

Hometown is adapted from Mastodon, a decentralized social network with no ads, no corporate surveillance, and ethical design.