The problem with algorithmic social-media feeds is that they presume to know better than the user what they will want to see. The problem with generative AI tools embedded in social-media is that they presume to know better than the user what others will want to see.

LinkedIn prompt: try writing with AI

I have been complaining about LinkedIn quite a bit, and I am about to do so again.

Lately, LinkedIn’s algorithm seems even worse than usual at surfacing anything interesting. It’s literally showing me weeks-old posts, and worse, ones that I already interacted with!

This is why people are sceptical of algorithmic feeds: if you’re making decisions on behalf of your users, the results had better be good.

(Me, on both Mastodon and Bluesky)

Today’s provocation is this piece in Wired, surely a publication that should know better.1 The problems start right from the sub-head:

A new analysis estimates that over half of longer English-language posts on LinkedIn are AI-generated, indicating the platform’s embrace of AI tools has been a success.

The site is over-run with AI-generated slop — [Borat voice] Great success!!!

I don’t want to over-index on something that may have been added by an over-enthusiastic editor, so let’s continue to the actual content of the article.

A woman talking on the phone; a giant head in the background

Adam Walkiewicz, LinkedIn’s head of “feed relevance”, sees AI as “a tool that can help with review of a draft or to beat the blank page problem”. I could spend the rest of this post just ripping this one statement to shreds. THERE IS NO BLANK PAGE PROBLEM! If you don’t have something to post — don’t post! Problem solved!

There is a vanishingly small group of people for whom this is an actual problem in their lives, and that is social media managers for brands. They are the only people who need to put something out regularly on LinkedIn, day in, day out. I can well imagine creativity flagging at some point — and in that situation, the sort of predictable output that LLMs are good at is exactly what is called for. After all, it’s not as if anyone is ever going to read all the posts from a brand in sequence. Instead, brand posts need to be created on a regular schedule (and liked or otherwise interacted with by employees and partners) in order for the brand to appear regularly in people’s feeds. Their goal is, to a first approximation, simply to exist.

This is a mode of content creation which is very much about regularity and consistency. And sure enough, the Wired piece quotes a woman who ”uses Anthropic’s Claude to spin up rough drafts of posts she creates on behalf of clients in the tech industry”. This makes perfect sense! The author of the posts is not herself a domain expert, and it is quite possible that the people she is working with are not experts either; they may work for the brands concerned, but probably in their social-media or marketing teams, not as technical experts themselves. LLMs are well suited to that situation: they will generate something obvious and anodyne, and that output can be validated based purely on a brand bible or similar document.

Individual humans don’t post like that. Aside from the hustle-porn LinkedIn influencers (Cthulhu help us all), real people tend to do well on LinkedIn by providing actual content: an opinionated take on an issue, based on long and proven expertise in a relevant domain. My colleague Brian Greene is a perfect example: he posts roughly once a day, and sometimes it’s just a gag, but sometimes it’s a deeper reflection. All of those posts come from him; ChatGPT wouldn’t come up with those takes, and it wouldn’t occur to him to ask.

This is why I find these justifications so disingenuous:

I find it fascinating how polarizing this technology can be, especially since tools like calculators or spellcheck, which are also forms of assistance, are widely accepted.

Personally, I think it all boils down to quality. If it’s a good read, I don’t think people are bothered much.

GenAI is not like a calculator or a spellchecker. Those only operate once the user has put something onto the blank page. LLMs are considered to be useful precisely because they will put something onto the blank page in the first place — although if there is any mathematics involved, you probably still want to keep a calculator handy.

And what about “quality”? A lot of LinkedIn was slop before GenAI; all the LLMs have done is to accelerate the production of slop. Instead of individual posts being slopped out by hand, GenAI promises to create slop automatically at scale. Not sure that’s a win, actually.

From one angle, LinkedIn may have inadvertently created the ideal laboratory for AI writing. Nobody’s logging on expecting profundity, hilarity, or sincerity. It’s the place where people strive to be the most anodyne versions of themselves, pleasant and inoffensive. Artificiality, in other words, is what everyone is expecting.

I hated those posts (and posters) before. Now, they have overrun my feed, and my hate is as the fire of ten thousand suns.

The LinkedIn app prompt is (still) “What do you want to talk about?” Of course you have to write with an audience in mind, but that audience wants to hear from you. If you sound exactly as boring and predictable as the rest of the feed, you won’t stand out in any way, and won’t get the engagement you were presumably after by posting in the first place.

If only a few people had access to GenAI, it would indeed be like a superpower for them. In our current social-media environment, where it takes effort to avoid the “helpful” GenAI features, all it does is to swamp the few human voices still trying to express an actual opinion.

I am not sure what the end-game is for LinkedIn. If the feed is overrun with AI slop that adds nothing to people’s understanding of a topic, we will all stop using the social feed, and LinkedIn will go back to being just a website for sharing our CVs. While it is darkly amusing to think of AI bots chit-chatting to each other in the feed with no human users to read their output, it would be a wasted opportunity for LinkedIn — and for us users, who kind of liked having a place to talk about work stuff.

🖼️  Photo by Priscilla Du Preez on Unsplash

  1. Yeah yeah, Wired today bears only a passing resemblance to the 90s Wired that was actually cool. I literally found a bunch of copies in a Swiss squat I was staying at while Interrailing, hoarded by a local nest of cipherpunks and phreakers; I particularly remember a feature on Atari Teenage Riot, which led to me resolving to seek them out at my earliest convenience — a decision I only partly regretted. ↩