Images won the feed.
Then the feed grew subtitles.

One: Egypt solved this and ran the solution for three thousand years.

Egyptian scribes wrote unpronounced signs at the ends of words to mark what class of thing the word named. Orly Goldwasser calls them semantic classifiers and maps each one as the head of a conceptual category. They were silent, visible, and combined with phonetic spelling rather than competing with it. Anyone who appends a picture to a sentence today is doing that job with different pictures.

Two: a picture has already signed a contract.

In March 2021 an employee of South West Terminal texted a photo of a flax contract to Chris Achter with the words “Please confirm flax contract.” Achter answered with a thumbs-up emoji and then delivered no flax. The Court of King’s Bench treated the emoji as acceptance and awarded about $82,200 in damages. Saskatchewan’s Court of Appeal upheld the decision two to one, finding the emoji a valid electronic signature on the facts.

Three: arrangement does the grammatical work that sequence cannot.

Neil Cohn and colleagues tested whether emoji strings behave like sentences. No word order beat chance, verbs came in below chance, and participants produced no prepositions or determiners at all. Their explanation is more useful than their result. Linear sequencing suits speech, and graphic communication was never linear. Templates encode their operations spatially, and once you look at layout the operations are all present.

Four: published linguistics treats memes as constructions.

Barbara Dancygier and Lieven Vandelanotte analysed image macros as multimodal constructions in Cognitive Linguistics in 2017. Their account gives the image a constructional slot, filling a role the text leaves open. One of their examples reads “Uses shopping cart / Leaves it in empty parking space.” Nothing in that text supplies a subject. The picture supplies it.

The argument is sixty-four years old.

Marshall McLuhan published it in 1962. The Gutenberg Galaxy predicted that electronic media would dissolve print culture and restore something closer to oral tribal life. Walter Ong’s secondary orality descends from the same reasoning. What arrived after McLuhan was email in 1971, Usenet in 1980 and the web in 1991, three text protocols in a row.

The format with the most grammar lost.

The construction-grammar evidence comes from image macros, and image macros are gone. Advice animals are catalogued as a fad of the late 2000s, and the format is described as bled dry of its novelty. Platforms now favour original short video and longer watch time, which throttles static jokes. Compositional structure did not turn out to be the adaptive trait.

Video did not shed its language. It grew a text layer that most young viewers now depend on.

Subtitles turned video into a reading medium.

Preply found 70% of Gen Z watching with subtitles most of the time, against 53% of millennials, and the BBC put viewers aged 18 to 35 near 80%. Hearing loss explains little of it. Respondents cited muddled audio at 72% and hard accents at 61%, and about 74% watch in public where the sound is off. Viewers also report using captions to move through content faster than it plays. Short-form video now ships with burned-in type as standard.

Memes are parasitic on syntax rather than a substitute for it.

Construction grammar is a theory of language, so filing memes as constructions places them inside language. The shopping-cart example works because English has a subject position available to leave empty. A picture can fill a gap that grammar opens. Filling that gap is a different achievement from replacing the grammar that opened it.

Memes and writing solve opposite problems.

Research on meme competence describes it as the boundary between members in the know and outsiders, where a failed reading can exclude. Density comes from knowledge both parties already share. Writing was invented to reach strangers across distances nobody could verify. A system engineered to be opaque to outsiders cannot take over from a system engineered to reach them.

Speed disqualifies the orality frame.

Ong lists oral cultures as conservative, and they must be, because memory is their only storage. Homeric formulae are stable for that reason. Meme popularity windows have run the other way, from roughly two years in 2008 to about four months in 2023 on indicative figures. Volatility on that scale describes fashion rather than tradition. Writing exists to defeat time, and cuneiform stays legible after four thousand years.

The frame itself is contested.

Secondary orality inherits the great-divide problem, the sharp split between oral and literate cultures that anthropologists have attacked as ethnocentric and unsupported. Ong draws standing charges of technological determinism. Most of what gets called digital orality is written, in chat windows and comment threads. Calling text oral because it feels informal names a mood rather than a mechanism.

No population declined.

A slope needs a t above 2.31 across thirteen fitted years to count as a trend. None of the four approaches it. Reddit fits at −0.17% a year with a t of −0.34, Hacker News at 0.40% and 0.71, Stack Exchange at 1.75% and 2.00. The imageboard fits at −0.48% a year with a t of −0.54.

The directional prediction fails harder than the flat one.

4chan cannot open a thread without a picture. Hacker News cannot host one. Their slopes are −0.48% and +0.40% a year, each indistinguishable from zero and from the other. The platform where images dominate by construction behaves like the platform where images are impossible.

4chan carries a second series nobody else can supply.

Its posts arrived with an image 24.5% of the time in 2012 and 26.9% in 2024, fitting at 1.41% a year with a t of 1.48. On an actual imageboard, across thirteen years, posts are not increasingly bringing a picture with them.

Three things the measurement cannot carry

The Stack Exchange verdict moves with the window. Starting in 2011 rather than 2012 returns a t of 2.61, which clears the threshold and reads as rising. The conservative reading is used and the other one is recorded rather than dropped.

On 4chan the endpoints alone suggest a 16% decline, from 16.7 words to 14.1. The fit says flat, because 2012 is a high outlier and every year since sits between 10.5 and 14.1 against 1.6 words of scatter. Subtracting endpoints would have produced a finding on the one source where a finding would have been most welcome.

All four populations measure people who are still writing somewhere. Anyone who abandoned writing altogether for video appears in none of them. That remains the strongest surviving form of the theory, and no public archive can reach it.

A register, working inside literate culture.

Memes compress at extraordinary rates over shared context, encode their operations in layout, and coin new units without asking permission from a standards body. Emoji cannot say the same. Unicode added eight base characters in 2025, against thirty to sixty a release before 2020, and its own membership has spent a decade arguing over whether emoji distract from the work of encoding historical scripts. What memes cannot do is inform somebody who does not already share the reference. That limit is structural, and it keeps them alongside writing rather than ahead of it.

The economics moved while the theory was being written.

The attention economy paid for capture, and capture favoured pictures. That payment rail is thinning. Small publishers lost about 60% of search referral traffic over two years, and chatbot referrals still account for under 1% of publisher page views. Work on generative engine optimisation finds citation rates rising by up to 40% when content adds statistics, quotations and authoritative language. Whatever replaces the feed rewards structure and propositional text.

The pattern worth keeping

Sixty-four years of this argument have produced one reliable regularity. Each time somebody announces that pictures are taking over from writing, the reply arrives in writing.