In the past two weeks, I saw sixteen paid ads promoting kitchen knifemakers on Instagram. Before 2024, I never saw any, despite following almost exclusively kitchen knifemakers for 4+ years. Where I’m not seeing paid advertisements, dozens of others makers are asking for help in sharing their posts — their accounts restricted or otherwise impacted by Meta’s policy crackdown on content. Some have begun abandoning the platform altogether, retreating to customer mailing lists or (admittedly, somewhat confusingly) fleeing toward X/Twitter and other platforms in the hopes that one burning ship somehow manages to right itself.

Yet, the really odd thing is that Meta never changed its content policy. Indeed, for those appealing, the most frustrating aspect has been the total lack of any human to speak to. For as bustling a market as Instagram may be, its police force appears to be silent.

How did this happen? Why is Instagram cracking down now, how can makers successfully appeal the spurious restrictions, and why is it so difficult to find a human to clarify all of this? Let’s dive in.

Stress risers and uncertainty

In preparing for this piece, I spoke to dozens of blacksmiths, kitchen stores, professional sharpeners, and more. Each has reported more or less the same experience:

• First, a wave of arbitrary content restrictions with seemingly no rhyme nor reason — a kitchen knife held in one hand to catch the morning light is flagged as “dangerous weapons” while that same knife laid flat on the table is 100% fine. Or, actually, two weeks later, a different knife laid flat on the kitchen table is flagged. Meanwhile, the CEO of Meta is posting videos of himself swinging an actual weapon. Alright. Cool.
• Second, each subsequent appeal to Instagram/Meta seems to undo restrictions on some posts, but not others. Again, there is no clear logic as to why that yellow-handled knife was fine upon review, but the blue-handled one wasn’t. In frustration, several photos are deleted just to get the gears unstuck and back into the good graces of The God at Menlo Park.
• Finally, Meta happily accepts your money to promote restricted content, even while refusing to speak directly on the matter. All those vaguely threatening messages about you disobeying the content policy? Hey, no worries! We’re happy to show your reel to thousands of users so long as you keep the money coming. But try to get a customer service representative to speak to beyond billing matters or literally reading you the terms of service verbatim from the website? Sorry, can’t help with that.

If that final point seems unbelievable, the better term may be dystopian. According to one North American maker, who requested to be left anonymous:

I’ve probably paid $200 total. First, I paid out for [Meta] verification, then more for my video to be [promoted]. They even have customer support for verified Meta accounts, but they’re mostly useless. I’ve shot them emails and called ten plus times, but it’s like talking to a brick wall…The last time I got someone on the phone, [all they did was] read the terms of use about dangerous weapons. When I asked if they now considered kitchen knives dangerous, they just repeated the same paragraph verbatim. I hung up on them.

Why is this all happening?

It’s at this point that we should address the elephant in the room. It’s no secret that Artificial Intelligence (AI) and Machine Learning (ML) made rapid progress in the last year or so. It seems as though every publicly traded tech giant is positively falling over themselves to integrate ChatGPT or other AI/ML technologies into their core products. They promise to make our lives easier and experiences more integrated, or whatever else will make their stock price go up.

But the promised experience of AI has a darker side. Rather than improving experiences or extending our enjoyment, these tools are leading down a gloomy path. Before going there, we need to take a brief diversion into the technical details of AI models and how they are used in real life.

Precision and recall: a detour into statistics

Hello, and welcome to my brief class on predictive models. These models are tools made with math that help us to predict certain things and I work with such models professionally (though not for Meta). So, while we’ll try to avoid over-complicating the subject, I’m about to make a controversial statement about Instagram’s behavior and want to make sure you are informed about the context before doing so.

For our purposes, we get to ignore how models are created (or “trained”/“tuned” in the technical jargon), and skip straight to evaluating their results. Instead of kitchen knives, let’s instead use models that predict pictures of cats. You can think of a model as essentially a black box, a mysterious machine that accepts input and gives an output. For example, this is how the model should behave when predicting whether several items are a cat:

Much as I’m confident you admire my artistic skills, you’ve likely guessed that professionals do not deal in neat icons of cats or airplanes, nor do predictive models actually give you a thumbs up or down. Instead, most models accept inputs that look like long lists of numbers — we call these “vectors” — and they output decimal numbers. So, let’s update our earlier picture.

That probably looks much less pleasant, but it’s a closer representation of what computer scientists use. The nice thing now is that instead of just a “yes” or “no” value, we now have numbers that can tell us a story about how confident the model is for each result. You could say that the model is about 98% sure it saw a cat on the first image, but the airplane is doubtful, maybe 7% confidence. This is helpful if, for example:

• What should we do with a fox which has pointy ears and whiskers like a cat, but also has a very unfeline nose? Is it “almost” a cat?
• Or what if I passed in a tiger? How should the model behave?
• How about a cat I photoshopped to give it human teeth? Is that still a cat?

The model still needs to output something, so it’s nice if we have a sense of how certain the model is. A value between 0.00–1.00 helps communicate this relative doubt.

What is accuracy?

Wait, so how accurate is our model above? To decide upon this, we need to introduce a new term: threshold. If we choose, for example, 0.95 as the yes vs. no line, then that’s our threshold and our three examples above have perfect accuracy.

Again, reality can be far more tedious and difficult to measure in terms of accuracy. It turns out that almost no model is perfectly accurate, so a single “accuracy” metric is rarely used because we care more about learning in what cases a model makes errors. To illustrate this, consider the following table of 200 imagined images we passed through our model.

While it would be accurate to say the model is “mostly right”, you should notice that it makes some mistakes more commonly than others. There were 7x more cat images misclassified as “not cat” than vice versa. It would be really useful if we could somehow represent this in our reported metrics, especially in cases where our numbers aren’t so even. Fortunately, there are such metrics!

• Precision is typically defined as how many relevant instances were retrieved out of the total retrieved instances. For example, “when our model said it was a cat, how often was it right?”
• Recall instead tracks how many relevant instances were retrieved compared to all relevant instances. In our case, “of all the cats we passed in, how many did you actually identify?”

Quick aside — why do we need precision and recall?

Precision is more important when we want to avoid false-positive cases. I’d get annoyed, after all, if I asked my phone to show me cat pictures and it showed me airplane pictures instead. So what if it leaves out a few “technically” cats like tigers? I don’t mind.

Recall is important when false-negative cases have serious consequences. Imagine, for example, an authoritarian dog regime who wants to flag all cats trying to enter its borders. A single cat slipping through would be disastrous, so who cares if you accidentally deport a few foxes along the way?

Equipped with our two new terms, we can happily report that our precision score is 93/94 (99%) and our recall is 93/100 (93%). That is, our model does pretty good at only outputting ≥ 0.95 if it’s really a cat…but we’re sometimes letting real cats fall under our threshold and ignoring them. Hmm. Wonder if we should do something about that.

Thresholds and curves

If you’re observant, you may have wondered about that 0.95 number I arbitrarily chose as an example earlier. It turns out that you can pick just about anything you want for this number and get different precision/recall numbers as a result. What if those 7 false-negative cat images were mostly quite close to the 0.95 threshold? Maybe, if we dropped our threshold down to 0.90, we could get a higher recall score…

It turns out that as you move your threshold number around, you can optimize more toward precision or recall — but never both. We call this the P/R (or precision/recall) curve and it looks something like the following.

P/R curves are a neat tool for the data scientist because they allow you to customize the types of predictions being made without changing your model. Need to just find cat photos and nothing but cat photos? Tune your precision up and recall down. Must find all cats and don’t care if your net is too wide? Ramp up on recall.

Now with our detour into statistics over, we can return to Instagram and knifemakers. Don’t forget thresholds and P/R curves, though; we’ll be back to that soon.

Costs and incentives

Back at Meta headquarters in 2023, there’s a crisis unfolding. Not for users — who cares? — but for stockholders. The social media giant, having seen a decade of unprecedented growth, was now buckling to the same pressures as the rest of the tech industry. Beyond the specter of AI and the increasing cost of capital investment, their CEO had just spent a year throwing money into the ill-advised pit of the Metaverse. Investors noticed, and turned up the pressure on Meta’s senior leadership.

With no option to claw back the costs already sunk into its speculative projects, Meta looked toward its maturing cash cow of Instagram and decided it was time to start milking for all it was worth. The rest is history, with Zuckerberg slashing 13% of employees in one go in late 2022 and rounding off 2023 with over 20,000 employees gone. Suddenly, the social media giant was insisting its remaining staff rein in more money with fewer employees.

I hinted earlier that I was going to make a controversial statement about Instagram’s behavior. Here it is: In 2023, Instagram cut its human moderation staff to near-zero and has instead relied on faulty AI. In fact, I suspect that any seeming human oversight, such as on appeals, is mostly feigned.

To understand how, we need one final piece — advertisers. Recall that most Instagram users are not paying for services and so the primary customer of Meta is not the person uploading photos, but rather the company paying to for the ad interstitial between stuff you’d rather see instead. Advertisers are a notoriously flighty bunch, you see, and they tend to gripe and moan when their advertisements don’t get prime real estate. Volkswagen, for example, doesn’t like it when their car ad is shown right after a Reel of a gruesome car crash. Neither does HBSC appreciate it when their personal loan advertisement is stitched between illicit sales of firearms in somebody’s Stories. Want your advertiser revenue to go up? Then you need to flag and remove undesired content as quickly as possible. And what if you had just the model to do that?

Back to P/R curves

It’s time to recall (heh) our discussion of P/R curves and thresholds from earlier. If you suddenly had (1) a lack of human moderators to make manual judgment calls, (2) hundreds of millions of non-paying users, and (3) a model with the above P/R curve for recognizing undesired content, what would you do? The decision is not a difficult one to make. If your pre-2022 threshold was the first point on the graph — nearly 100% of all flagged content really was truly bad, but you only caught 75% of it — then why not move to a point further right? Sure, you’re suddenly flagging a significant percent of legitimate content as restricted, but who cares? Those users can appeal the decision if they want, and now our advertisers are happier.

Then somebody comes up with another clever idea. What if you moved your threshold even further to the right?! You’re mistakenly flagging way more content, but your recall score has never been so high! Now, instead of actually having humans review the appeals process, just let a machine run it instead by briefly lowering the threshold score by a few points and checking again. It’s amazing how many salaries you can avoid paying through this trick. Heck, if you make the users go through this enough, they’ll even stop asking you altogether. The best part of all? You have no legal duty to answer to why their content was flagged. No transparency required! Users will guess en masse trying to figure out what about their video or photo was in violation and will self-censor in hopes of keeping their access unfettered.

In short, no human plans to review your appeals. That was the old way of doing things.

I will raise up a profit from among them

We find ourselves at the end. The end of this three part series about Instagram and knifemakers, but also what I predict is the end of Instagram as a platform for knifemakers and other craftspeople. I started this article with an observation about the amount of paid promotions I’ve seen in the past two weeks for knifemakers, and there seems like no more fitting harbinger of the coming apocalypse.

How strange it all feels. Many knifemakers and craftsmen first joined Instagram because of what it promised: a no-frills approach to seeing pictures of interesting stuff. As features for exploration like Reels arrived, it seemed possible to capture outside attention in innovative, scalable ways so long as you put in the effort. Now, here in the end, we’re back to the pay-to-play game that most of us first fled to Instagram to escape. It’s pure marketing again.

What does this all mean for the kitchen knife community? Nothing good, I’m afraid.

1. At a minimum, it means a lost avenue for newcomers to find their way to small-time makers.
2. These avenues will likely shift over to stores with marketing budgets who will become the new landing strip for the curious. It wouldn’t surprise me at all to see knife shops take over more of the Western-made high-end market while independent craftsmen work twice as hard for each new sale.
3. Forums are similarly consolidating, though there is some hope for emerging craftsmen breaking out among enthusiasts there.
4. Or perhaps it will become the golden age of organic networks again, where line cooks swap tips about the best possible sources of high-performance knives.

Wherever it is, I’ll see you on the other side.