It’s technically impressive, yet functionally specious.

The algorithm is vain - in all three definitions of the word. Here’s some data from the past week to back this up.

Post 1 describing how the LinkedIn feed is on the path to becoming as ignorable as our email spam folder: 7,000 impressions.

Post 2 offering commentary on Canada’s AI Strategy survey and important questions that remain unasked: 480 impressions.

Post 3 calling LinkedIn users who post AI slop the Zombies of the Dead Internet. 17,000 impressions.

Post 4 announcing TYPECHA, an AI-slop-fighting text editor that provides proof of work for your writing and editing process: 170 impressions.

Post 5 synthesizing a decade of experience in medical imaging and how solutions fell short of accounting for the myriad of cognitive and workflow factors that would affect the adoption of AI by radiologists. Most importantly, it offered the insight that every domain is full of such AI blind spots. 160 impressions.

Vain as in ingeniously contrived - yes.
Vain as in marked by ineffectiveness - yes.
Vain as lacking in real value - yes.

I think the lack of attention to human factor engineering principles is real issue in the current design and implementation of AI systems - both at the consumer and enterprise levels. I would love it if a post on the topic got tens of thousands of view. It could save someone millions of dollars.

And if those who generate their LinkedIn content with AI are doing so to reach those remaining human accounts…

And if those remaining human users, seeing LinkedIn’s algorithm reward the AI-generated content begin seeking solutions to do the same…

And if, after adopting those solutions the pressure grows on the remaining human users…

Does that mean that those who post with AI are zombies?

This has rendered email inboxes largely ignorable. the bulk of it goes to spam. The rest is filtered. What remains is quickly skimmed over. Most emails that don’t compel an immediate response will likely be immediately forgotten - or deferred until it is. Unless a sender has earned a place on the recipient’s white list, whether explicitly or implicitly, most emails will never be seen.

Information feeds taken over by AI-generated content will soon meet the same fate. Short and long form content, with and without images, can be generated at very little cost and in very little time. it can be topically relevant, context-relevant while fooling most of us as to its provenance. There is so much of it in our online information feeds, including here on LinkedIn, that it is mostly ignored.

Email is still used everyday. It’s critical both at the individual and enterprise levels. But to the recipient, the first question is always “who is the sender?”

As we browse our LinkedIn feeds, we all ask ourselves a version of that question. Who is behind this content. but not just who posted it - who created it. Did they create it themselves, use a ghost-writing service, chatGPT or some other LLM? Unconsciously (or consciously for some!), we look for any reason to ignore it.

But if the content is good and the creator is genuine, then maybe it’s worth our time. Unfortunately, the cues we seek are hard to find.

This is part 2 of “101 Ways AI Can Go Wrong” - a series exploring the interaction of AI and human endeavor through the lens of the Crossfactors framework.

View all posts in this series or explore the Crossfactors Framework

If you know, you know. And until then, you don’t.

Algospeak is coded language or modified terms, often with intentional misspellings, meant to discuss a sensitive topic online while evading detection by content moderation algorithms.

Why It Matters

Coded language is nothing new, but its need and use is being shaped by technology. Content is ranked and many serious topics are tightly controlled due to an impressionable audience and the inability of moderation algorithms to distinguish between a serious but constructive discussion and nefarious content that may glorify or otherwise inappropriately discuss the topic.

Real-World Example

In January 2025, “cute winter boots” was a term trending on TikTok and other platforms. One may initially think this refers to a seasonal item many of us need. In actuality, it references the wave of raids carried out by the U.S. Immigration and Customs Enforcement on sanctuary cities. The acronym for this agency (ICE) is the apparent connection to the term, perhaps aided by the coincident seasonality.

Key Dimensions

Speed of trends: new terms that trend quickly can be favoured by ranking algorithms, which goes beyond the initial purpose of evading moderation. The existence of influencers with large followings helps speed up the trend.

Ethical dilemma: Platforms are to balance free speech with a need to curb harmful content.

Context collapse: Users unaware of the algospeak terms may misunderstand the intended meaning. Large language models need to be retrained or fine-tuned to include the new context that’s been created and may be short-lived.

Hijacking: Algospeak terms that are trending may hijack or overshadow the conventional meaning of the term.

Is algospeak only a concern for large social media platforms or users of these platforms? I don’t think so. Large companies have grown the use of their chat-based tools (Teams, Slack, etc) and while employees may seek to cloak topics for different reasons than TikTok users, these companies may see their monitoring tools react to coded language in unexpected ways.

Technology has changed the way we listen to music.  In moments, we can listen to an album on Spotify that we had forgotten about for 15 years.  Or be notified that our favourite band has released their first album in 7 years. Or find a long lost one-hit wonder that you once couldn’t get out of your head.  All from our phones.

40% of streams on Spotify come from their own playlists and 55% of users report that they discover new music through Spotify’s recommendations.  So what’s behind these recommendations?  Or more precisely, what’s behind these playlists?

Algorithmic recommendations are ubiquitous and can sometimes feel like magic. We can be forgiven for not caring what drives them, or for thinking too simply about the elements that factor in.  In the case of Spotify playlists, they remove friction, especially if you just want to put something on while doing some other task.  This isn’t new.  Over-the-air radio still serves this purpose for many.  And their advertising-based revenue model seems simple.  If they keep you tuned in, you’ll hear their ads.  Stations optimize the tempo of music to commercials, the song selection, relevant information, and promises of an interesting story or hit song to achieve this end.  Or is it that simple?

The listening and the radio stations are not the only parties involved.  It is artists and record labels that provide the music and they have their own incentives too, which is the hope that listeners will want to buy an artist’s album.  Payola is the practice of a station receiving payment in exchange for airplay without disclosure to the listener.  It is illegal, has been for a long time, but still considered to be occurring by industry insiders despite the radio landscape being heavily regulated.

Music streaming, however, is not regulated - at least not to radio’s extent.

Spotify sees 88% of its revenue coming from premium users and they maintain a very low (<4%) churn rate. They are incentivized to keep their premium users happy and have them bring their streaming into many parts of their daily routines.  In practice, this means removing friction from their lives by providing music with little to no input from them - and to keep the stream playing in a way that becomes essentially unnoticeable. This is witnessed by the popularity of “mood” playlists, where the music is there to support a mood or activity as background, as opposed to listening to music in an intentful way.

This creates an interesting problem.  The recommendations do not want to risk upsetting your mood.  If you skip songs on a playlist too many times, you’ll abandon this playlist.  If it happens across playlists, you’ll abandon the playlist entirely. Or you’ll make your own - without Spotify’s input.  Therefore, recommendations won’t take any risks in helping you find your new favourite song.  In actuality, they’ll favour songs that will go mostly unnoticed or at least have a familiar sound. So while more than half of Spotify listeners report discovering music on the platform, is it really discovery?

Just like radio, there are other stakeholders - those who provide the music.  And just like radio, there are incentives for artists and labels to pay for exposure (which they do).  Spotify also exerts massive leverage and resources.  They can choose which artists to include on their playlists, and much like song selection, artist selection will be influenced by an incentive to keep a premium subscriber’s stream uninterrupted.

By now, I hope you’ll agree that those recommendations are not generated in a straightforward way - and not entirely for your benefit.

For a deeper dive, I recommend Liz Pelly’s book - Mood Machine: The Rise of Spotify and the Costs of the Perfect Playlist.

Cold outreach is nothing new. It predates LinkedIn, email and the internet.

If you’re like me, you’re on LinkedIn to expand your knowledge and network and perhaps even be entertained a bit. Most of us willfully accept a volume of cold-outreach in service of this goal. We easily ignore cold, unpersonalised connection requests and messages while giving more consideration to those showing more effort and relevance.

I believe this balance is delicate and LLMs could upset it.

There had been an inverse relationship between the scalability of the outreach and the personalization it could offer.

On LinkedIn, those who use LLM tools hide the provenance of their outreach - it shows up as coming from their account and signed with their name. The message text has the appearance of being highly relevant to your profile details or a recent post or comment. It often misses the mark - especially if you’re already writing nuanced takes on the use of AI, like me.

I’d rather get a cold outreach that’s truly personalised by a human, or not personalised at all. The middleground is dishonest at best. I don’t want to spend time figuring out if something which has the right keywords at first skim actually makes sense and if it was written by the person who is presenting it as theirs.

Cold outreach is the first step to engage in a business relationship. If you’re using LLMs to do this on your behalf, you are starting on the wrong foot.

I brought up how YouTube provides results, unprompted, for prior and highly specific search queries on the homepage - days after the initial query itself did not satisfy your search.

I did a bit of digging, but unsurprisingly, there’s no information about the inner workings of YouTube’s search algorithm. Also, the bulk of the information I found focused on optimising content for the algorithm.

However, Google has stated that they’ve transitioned from optimising for watch time to optimising for user satisfaction. Watch time is easy to measure. How does one measure user satisfaction without using watch time as a surrogate measure!

But I’d like to infer a possibility from this.

All of YouTube’s recommendations, including search results, are based on the profile they have of the user. The update frequency of these profiles is also not disclosed. However, it is reasonable to assume that it does not happen in real time. Even if it were, the data collected from an abandoned search would not exist to impact the algorithm until the search was, in fact, abandoned.

If an abandoned search is a sign of an unsatisfied user and the data relating to an abandoned search isn’t available until the search is abandoned, this explains the delayed impact on the algorithm and the search results it drives. If we add to that the fact that the user recommendation profiles are likely not updated in real time, this all begins to make sense.

At the risk of stating the obvious, there’s a fundamental problem with this approach. It guarantees that a well-profiled user will leave empty-handed when seeking something new. This is a big issue for someone like me. I want to fix the hinge on this vintage cooler before the weekend. I don’t need 10 cooler hacks that will change my life.

This is something that happens when searching for something specific. For example, a search relating to the repair or maintenance of a specific item around the house.

I observe this pattern as follows. The first 3 or 4 results are OK. They show some relevance but are not very promising. Following these are a few results of “People also watched” which are hardly related at all. Perhaps the algorithm wishes to redirect my attention towards a path it can better monetize. Then a few more results that continue to degrade in quality and relevance. Then a series of “For You” results, which is the content I’d see on the homepage. I’ll scroll past these, see a few more mediocre hits and often give up.

Now here comes the weird part. Two or three days later, I’ll return to the YouTube homepage and a selection of good results relating to that previous query are now on my homepage. This has occurred many times and continues to surprise me. These results are the ones I wished to find, but were not available initially. They are often older videos, with very low view counts and lower production quality.

It is as if I had been looking for a needle in a haystack, gave up, and someone had found it for me while I was sleeping and placed it where I’d find it when I woke up.

Since Google initially bought YouTube and they now share a parent company, we should ask: is this a feature or a bug? What limitations or incentives are at play here?

That’s what content recommendations algorithms have been doing for the past week.

Did you post something silly on April 1st? Know that it’s been reposted many times beyond its ‘best before’ date. As the context of the first of April quickly fades, how many readers are left confused or even perhaps questioning your cogency.

Automated recommendation systems are not context-aware. That is, unless they’ve been explicitly programmed to be so, by a human. It took years for Facebook to stop reminding its users of painful memories on their anniversary dates.

April Fools and the anniversary of the passing of loved ones are the easy, clear cut cases. However, these systems are connected to a veritable firehouse of new content, often relating to developing or nuanced events. The lack of context-awareness will continue to be a notable shortcoming of recommendation algorithms and guarding against all edge cases will be impossible.