In partnership with

Up until a few years ago, you could recognize a semi-personalized, yet mass-mailed cold email because it included details about you that could, in theory, apply to hundreds or thousands of other people: "like other graduates from X..." "Since you're one of the top Y..." "As a Z..." etc. Get enough of these emails and you can spot the mad-lib portion and learn to ignore them. But in the last year or two, this has flipped around, and now the details are more specific, but almost any email you get from a stranger that mentions a few facts about you in the first line is probably LLM-produced spam.

There's a general pattern in media where lowering the barriers to entry (lowering cost of production or distribution) means lowering the average quality, but potentially raising the peak. Pre-printing press, the high cost of the written word meant that the works most likely to be published were things like the government records, official pronouncements, devotional texts, and the bible. A lower cost for printing meant that there was room to publish many more bibles in many more languages, persuasive essays on why those monks busily illuminating manuscripts were living pretty well off of tithes, and popular entertainment. But it was still popular entertainment for a narrow slice of the populace; literacy rates in Northern Europe were 5-10% before the Gutenberg press, and were around 15-25% after a century, and while books got cheaper in this period, they were still something most people had to save up for: an English worker in 1600 was paying about a day's income for a Shakespeare quarto. But that example illustrates that there was a wider variety of written material available. Plays predated the printing press, but they were experienced live; that was the most cost-effective way to transmit the text.

As some of the examples above indicate, cheaper media led to some problems: anyone could say whatever they wanted, and if someone claimed the narrative high ground, their story might spread fast before someone else figured out that it was false. There was an actual literary genre called libels, which consisted of every defamatory thing the author had heard or chose to make up about some public figure. These benefited from a media dynamic that's familiar today: if something takes the form of a kind of content that used to be expensive to produce, it feels credible, even if it's fictional; a well-produced AI video's credibility is somewhat anchored to the historical cost of producing such a video, and an AI-generated nonsensical research paper is still a LaTeX-formatted PDF that implies some credibility: the medium is the message. The Alchian-Allen effect tells us that if there's a high fixed cost associated with something (such as producing a video), the marginal cost of making it better (through fact-checking, for example) is relatively low. But that means that if the cost of media production declines over time, we routinely overestimate how credible a given media source is. That applies even to the written word; if you were to get a column like this delivered to you a generation ago, it would likely have shown up in print, and overseen by an editor; the hiring process for someone to write it would have stricter filters, and the content would have to attract sufficiently valuable eyeballs that even 90s-era print ad targeting could make it profitable. Whereas today, in the time that it took you to read this paragraph you could have started a free account on a newsletter service and started typing (or copy and pasting from Claude) whatever it is you wanted.

In one sense, media economics is a big deal because media consumption takes up a substantial fraction of people's free time, and first- and secondhand media consumption determines what they think is going on in the world—unless you happen to be personally acquainted with Donald Trump, and the two of you are in the habit of talking about work regularly, your view of Trump's policies and their impacts either comes from news you've read, or from talking to people about news they've read. But media economics is actually becoming a bigger deal over time, because the information economy is in some sense a media-mediated economy; email is just a very niche digital magazine with a wildly fluctuating subscriber list and an unpredictable publication schedule.

We're seeing this now, as one of the fundamental tradeoffs in media—lower costs of production make verification more important, but the demand for verification shows up on a lag—is now the tradeoff that defines returns for AI, which itself is the majority of US GDP growth right now.¹ The verification problem takes a while because at first, demand is invisible: it's flattering to get cold emails, it's delightful when an LLM confirms that your writing is absolutely wonderful and the only edits are minor, and it's comforting that there's photographic or video evidence for whatever you happen to think is going on in the wider world. But in a way, all of these AI outputs should be interpreted, not as the media artifacts that they copy, but as prompts, which are shared with an artificial or natural intelligence that tries to figure out if they're real (and valuable), or if they're just a way to exploit certain unpatched bugs in our media consumption.

These bugs compound, and they create a very interesting dynamic in areas where there's lots of inefficiency but scarce context: a company that's working with messy data that isn't well-represented in the training set can treat that messiness as a liability, and focus their AI efforts on very nice-looking powerpoint presentations that smooth over whatever inevitable issues they have² , or they can look at that messiness as a valuable proprietary asset—as a captive source of unique data that's valuable to either AI labs or to anyone who wants to have a competitive moat while using fairly commoditized models. As with other cases where AI can be a substitute or complement to human effort, it's really a test of willpower. An undergraduate today is one of the luckiest people in history, in that they have four years with—outside of a handful of majors—minimal work requirements, plus access to artificial intelligence and living domain experts, and protection against the autodidact's curse of having grad student-level knowledge in some parts of their field and a high school understanding of others. Some students use this well, many more convert AI and the opportunity for learning into leisure time. Corporate America has plenty of organizations that will follow the path of the CheatGPTer, and a few that will eat them alive.

We've written a lot in The Diff about media economics, spam, and of course AI content generation:

• We've covered the media cycle by way of Buzzfeed, most recently here ($).

• Back in 2022, we looked at the generative AI landscape and asked what happens when the median audience size for a given piece of content is one.

• AI moves fast enough that this piece about AI spam is already obsolete in its particulars but still generally true ($).

• You can use AI to help with writing, but you need to be very judicious about using AI to actually do your writing.

• We'll spend a lot of tokens on responding to the token use of adversaries ($).

4x your communication output. Same quality. No burnout.

The bottleneck isn't what you want to say — it's how long it takes to type it. Wispr Flow removes the bottleneck.

Speak naturally and get polished, send-ready text for executive summaries, client updates, board recaps, investor notes, or just the 30 Slack messages you're behind on. Flow strips filler, formats numbers and lists, and preserves your tone.

Used by teams at OpenAI, Vercel, and Clay. 89% of messages sent with zero edits. Works in every app on Mac, Windows, and iPhone.

Start flowing free

Join the discussion!

1  That's true in a literal accounting sense, where you add up all of the incremental AI capex happening and compare it to economic growth. But it's also true if you look at productivity growth, which has been running at around the midcentury and late 90s/early 2000s level since roughly the launch of ChatGPT.

2  A common problem in big organizations is a subtype of lying through omission: lying through compression, by squashing growing problems into better-looking averages.