I suppose I shouldn't be surprised by bad-faith comms from Mozilla after the past twelve years. Even so, I'm surprised by how bad their response to the MDN AI debacle is. For more context on the debacle, see @lexyeevee's original post.
Speaking of the human element, LLMs also have the useful attributes of having zero ego and infinite patience. LLMs don’t mind answering the same question over and over again, and they feel no compulsion to gate-keep for online communities. It’s sadly not uncommon today for learners to face the discouraging experience of asking technical questions in online forums only to be dismissed or shamed for their “dumb question.” When done well, there is a clear opportunity to employ AI-based tools to improve the pace of learning as well as inclusivity for learners.
"LLMs have zero ego", wrote the people who built confidently wrong machines to amplify their own confident wrongness.
You know what LLMs also have zero of? Superego, which makes them very useful tools for executing unethical actions while disclaiming responsibility for those actions.
And finally, it's incredibly slimy to weaponize "inclusivity" (something Mozilla has rejected over and over again) to justify business decisions that are -- under the most charitable interpretation -- lazy and sloppy.
None of what I'm about to say is original, but maybe there's value in as many people as possible repeating it in as many ways as possible:
- LLMs generate text that looks as if a person wrote it. That's it. That means they might be useful tools for tasks like "here's a draft email I wrote, rewrite it to fix grammatical mistakes and increase formality a little" -- so long as the human reads the output to make sure *meaning is preserved*. I haven't tried it for that, but I hear it helps. That's fine; anything that eliminates inequality based on different levels of writing skill in a particular language is good, to me. *As long as it's used with human oversight.*
Because LLMs manipulate language without modeling its meaning, they are terrible for generating language that faithfully reflects an external reality. How could it do that when it has no sensory organs that create an independent model of reality outside of language? If you've created -- for example -- what's supposed to be a trustworthy, human-edited reference manual for a programming language, something that looks like an authoritative reference and so far has been one, and then you replace that with nonsense text generated by a statistical model, people are reasonably going to treat it like it's a trustworthy, human-edited reference, because the nonsense text superficially looks like a person wrote it. (See the first sentence.) For programming documentation, this can cause at worst frustration (something I care about because I've spent my career, such as it is, trying to make programmers' lives less frustrating). You can imagine contexts where the implications are much worse.
- As a corollary, we're headed down the road to an increasingly shitty world where all text is machine-generated because that's cheaper than paying human writers and editors, where the absence of human oversight means there is no truth anymore except that which an individual can verify with their own senses, and where the people lucky enough to have learned critical thinking skills know that reliable sources no longer exist while the unlucky majority loses themselves in a sea of lies. Navigating the modern world requires trust in secondary sources. For example, I know that vaccinations are safe and effective (and don't cause autism) not because I've personally done clinical trials, but because I trust the people who did. Even before the rise of LLMs, disinformation created by humans acting in their own economic interest (and sometimes non-economic interest) was a threat. Injecting machine-generated disinformation into the information economy accelerates the threat. Machine-generated disinformation is motivated not by human self-preservation or human pleasure -- it's not motivated at all (LLMs don't have an id, either). It just runs on its own until someone pulls the plug. Or, until all media that have even slightly responsible editorial oversight are dismantled because it's no longer profitable to pay editors and journalists to have a tiny bit of a moral compass, when cloud computing time is cheaper and can generate more content faster. Who wants to pay extra for truth?
LLMs are dismediation at scale:
Dismediation isn’t discourse. It doesn’t disinform, and it’s not quite propaganda, as that term has long been understood. Instead, dismediation seeks to break the systems of trust without which civilized society hasn’t got a chance. Disinformation, once it’s done telling its lie, is finished with you. Dismediation is looking to make sure you never really trust or believe a news story, ever again. Not on Fox, and not on NPR. It’s not that we can’t agree on what the facts are. It’s that we cannot agree on what counts as fact. The machinery of discourse is bricked.
-- Maria Bustillos, "Dismediation, revisited"
Bustillos' article, written in 2016 and republished in 2018, pre-dates the rise of LLMs. But there's only one detail that doesn't apply to the present moment: AI-powered dismediation isn't looking to make sure of anything; it doesn't have intentions. The people who created it have intentions, but I don't think most of them intend to make anyone "never really trust or believe a news story, ever again". They are neutral, disinterested. They have no investment (financial or psychological) in making people distrust news stories, but they also lack personal investment in not making people distrust news stories. The machines they build are neutral as well, which is why they tend towards untruth. Being honest requires constant, continuous effort. Lying requires none.