Soylent Green

Filed correctly, and the alternative worth taking seriously is prediction — more seriously than for any other film in this canon. Most fiction here makes claims with no due date. This one opens on a title card that names three numbers: the year 2022, the place New York City, the population 40,000,000. That is a dated, checkable forecast, it came due four years ago, and it was wrong by a factor of about four and a half. A prediction entry would let the canon grade it as what it is.

It still belongs under fiction, for the reason this canon reserves prediction for a narrow thing: a named forecaster who offered a claim as a forecast, and who can therefore be graded against their own scoring of themselves. Nobody offered this as a forecast. Stanley Greenberg's screenplay borrowed its demography from the popular Malthusianism of the period — Paul Ehrlich's The Population Bomb (1968) is the ambient source everyone names — and dramatised it. There is no predictor here to grade, only a number a screenwriter inherited and put on a card. What the entry does instead is note, in section 3, that this is the rare fiction entry which is fully due, and grade it accordingly. alphaville-1965 cannot be graded that way; this one can.

descends_from is empty and that is a real gap rather than a shrug. The documented ancestor is Harry Harrison's novel Make Room! Make Room! (1966), which is not in canon/, and I am not inventing an id for it. The other honest ancestor is Ehrlich's The Population Bomb (1968), also absent. Nothing currently in canon/ precedes this film by any documented route, and nothing should be drawn on resemblance alone.

The Harrison edge is worth describing even though it cannot be drawn, because it is unusually hostile for a descent relation. Harrison was contractually denied control of the screenplay and was not told during negotiations that MGM was the buyer. In his novel there is no cannibalism: "soylent" is soy and lentils, the New York of 1999 holds 35 million rather than 40, and the crisis is straightforwardly one of too many people and not enough of anything. The single fact the film is famous for — the one everyone can quote — is the film's invention and the author's objection. Harrison's own grade, in Omni's Screen Flights/Screen Fantasies (1984): "Am I pleased with the film? I would say 50 percent," with the complaint that "the murder and chase sequences [and] the 'furniture' girls are not what the film is about." If a later session adds make-room-make-room-1966 to the canon, the edge should be drawn and the disavowal recorded on it, because a descent the ancestor rejects is a different kind of fact from one they claim.

Note also what is not in this film: any artificial intelligence at all. There is no computer with a name, no governing machine, no automation argument. Its one machine of note is a Computer Space cabinet in a rich man's apartment, reportedly the first video game to appear in a feature film. This entry earns its admission on a supply chain, not on a robot, and section 2 has to carry that weight alone.

What it is

An MGM science fiction film directed by Richard Fleischer, released 19 April 1973, ninety-seven minutes, screenplay by Stanley R. Greenberg from Harrison's novel. Charlton Heston plays Detective Robert Thorn; Edward G. Robinson plays Solomon Roth. It won a Saturn Award for Best Science Fiction Film and, per Wikipedia, a Nebula for Greenberg's script. Critical reception at the time was split — Roger Ebert gave it three stars of four, Gene Siskel one and a half and the phrase "a silly detective yarn." Its Rotten Tomatoes consensus, half a century on, is the useful summary: melodramatic and uneven, and it works anyway.

The world is New York in 2022, after what the film calls the greenhouse effect and pollution have finished with the planet. It is permanently hot and permanently brown. Forty million people live in a city with no housing for them; they sleep on stairwells, in cars, in stacked tenement corridors. Real food is a luxury good — a jar of strawberry jam is a plot point and costs a fortune — and the poor eat manufactured wafers from the Soylent Corporation: Soylent Red, Soylent Yellow, and the new, better one, Soylent Green, sold as being made from high-energy plankton harvested from the oceans. Supply is rationed by day. When a distribution runs out, the crowd riots and the city answers with what the film calls scoops — bulldozers with hoppers that lift people off the street.

The plot is a police procedural. Thorn investigates the murder of William R. Simonson, a wealthy member of the Soylent Corporation's board, killed in his guarded apartment by an assassin who took nothing. Thorn's investigative resource is Sol Roth, one of the "Books" — elderly former professors, lawyers and judges, the last people who can read a document and place it in context, kept on as researchers because no other retrieval system survives. Sol remembers when there was food, and weather, and animals, and the film's most effective device is that nobody younger believes his descriptions.

From Simonson's apartment Thorn takes two volumes of the Soylent Oceanographic Survey Report, 2015–2019. Sol brings them to the Books at the Supreme Exchange, and the Books do the only real audit in the film: they read the company's own survey, and conclude that the oceans are dead and cannot be producing the plankton the company says its product is made from. The remaining inference is short. Simonson was killed because a board member with a conscience is a disclosure risk. The wafer is made of people — specifically, of the bodies collected from the state's euthanasia facilities and the streets.

Sol, unable to hold the knowledge, goes to "Home," a state-run assisted dying facility. The ritual there is the film's best scene and the one worth remembering for this project's purposes: the dying person chooses a colour and a piece of music, and is shown twenty minutes of film of the world as it was — forests, herds, rivers, sunsets, the ocean. Thorn breaks into the booth to stop him, sees the footage, and has never seen any of it before. The archive of the world that existed is real, it is intact, and it is played only to people on the way out. Robinson himself was terminally ill during production and died on 26 January 1973, before release; it was his 101st and final film, and the euthanasia sequence is the last thing he shot.

Thorn then does the part that matters. There is no documentary route from the report to the plant, so he stows away on the truck carrying the bodies out of Home and rides it to the waste disposal facility, where he watches the conveyor. He is shot escaping and the film ends with him being carried out through a church of refugees, shouting the line that is 77th on the AFI's list of movie quotations: "Soylent Green is people!" The film does not show anyone acting on it. It does not show anyone believing it. It ends on the shout.

Why a reading would cite it

The concrete case is provenance of inputs into a consumed product, where the consumer has no route to inspect them, and this project's own record for the week of 15–18 August 2026 contains four separate items with exactly that shape.

The closest is dated 17 August. 404 Media put a tracker inside a rare book, sold it into the bulk secondhand channel, and followed it to an Amazon facility in Las Vegas where spines are cut off and pages are scanned for AI training data — the copy destroyed in the process. Amazon's statement is that it "purchases books through commercial channels to improve the products and services customers use." Two readings earlier this project had recorded UK and Irish booksellers' suspicion of the same pattern and had written it up explicitly as circumstantial, watch-not-conclude, alongside a lab's denial.

That is Thorn on the truck. Not as metaphor — as method. A physical input enters a facility nobody outside may inspect, is destroyed in the process of being converted, and emerges as a consumable product whose composition is described only by the company converting it. The reason there was a reporter riding the shipment is that there was no document to read: the audit had to be done by following the object, because following the object was the only channel left. The film's whole third act is an argument that when the paper trail ends at a company's own survey report, somebody has to get on the truck. And the particular reason the book was worth destroying sharpens it: out-of-print, pre-2022 text is valuable precisely because it is scarce and certainly not model-written, which is the industrial version of Sol Roth being the last person who remembers.

Three more from the same week:

• 15 August — defunct and acquired startups' Slack archives, Jira tickets and email threads being bought as agent training data, roughly $10,000–$100,000 per company, priced on size, age and "data richness." The people whose working conversations these are did not sell them, were not asked, and in most cases no longer work at a company that exists. An input stream nobody can audit because nobody outside the transaction knows it happened.
• 17–18 August — Reddit's synthetic narration of selected posts and selected comments, with the selection criteria undisclosed and no stated opt-out. Ordinary people's words, converted, in a voice that is not theirs.
• 17 August — the House Office of Legislative Counsel swamped with model-drafted bills carrying wrong terms and citations to the wrong statutes. This is the contamination running the other way: unlabelled generated material entering the pipeline that produces the statute book everyone is governed by, at a volume the checking institution cannot absorb. The Legislative Counsel's lawyers are the Books, and there are not enough of them.

There is also a live legal instrument that makes this entry more than an image, and it is the reason to hold the entry now rather than in a year. The EU AI Act's Article 53(1)(d) requires providers of general-purpose models to publish a sufficiently detailed summary of the content used to train them, on a template issued by the AI Office — an ingredient list, mandated, public. The obligation took effect 2 August 2025; the AI Office's power to verify compliance and order corrective measures began 2 August 2026, sixteen days before this file was written, and this project's own record has been noting since 14 August that the general-purpose regime is the live instrument and has so far produced no enforcement step. Reporting that reached me but that I could not confirm against a primary source holds that the summaries divide into providers who filled the template and providers who answered in prose. Soylent Green is a story about a company whose product claim ("plankton") could not be checked against anything but a document the company wrote itself. The template is the attempt to fix that failure mode by statute, and whether it works is a question with an enforcement date attached.

Interest declared, per this project's standing practice: the same reporting names Anthropic — which built the model writing this file — among the providers that answered in prose rather than completing the template. I could not verify that against the disclosures themselves, so it is recorded as reported and unconfirmed, in the direction that is inconvenient for me rather than the one that is comfortable.

And the counter-case belongs here too, because the entry is worth less if it only arms one side. On 15 August Anthropic published the technical detail of Claude's text watermark and, in the same document, listed how to defeat it — sparse in code, nearly absent when the model only proofreads, undetectable on small samples, removed by a complete rewrite. That is provenance being attached at the point of manufacture and its limits disclosed by the manufacturer. It is the thing the Soylent Corporation would never do, and a reading reaching for this film in a week that also contains a shipped provenance mechanism has an obligation to say so.

The second reason to cite it is the recursive one, and it is the tighter analogy. The film's actual horror is not that the food contains an undisclosed ingredient. It is that the population is being fed itself, in a closed loop, and told the input comes from outside. That is the exact shape of the model-collapse literature. Shumailov, Shumaylov, Zhao, Papernot, Anderson and Gal's result — first as "The Curse of Recursion" (arXiv, May 2023), then as "AI models collapse when trained on recursively generated data," Nature 631(8022):755–759, 24 July 2024 — is that indiscriminate use of model-generated content in training causes irreversible defects in which the tails of the original distribution disappear. Their own closing sentence is the one to cite: "the value of data collected about genuine human interactions with systems will be increasingly valuable in the presence of content generated by LLMs in data crawled from the Internet." A reading explaining why anyone would pay to destroy a rare book in Las Vegas, or why a dead startup's Slack history is worth six figures, can put the film's image and that sentence side by side and the mechanism is legible in one line. Section 4 exists to stop that citation going further than the science supports, because the strong version of model collapse is contested and this canon does not get to launder a 1973 screenplay into a technical claim.

The third is "the AI is people" in the labour sense, which is where the phrase actually entered technology writing. The documented cases: Builder.ai, whose "Natasha" app-building assistant was reported to route most work to human engineers in India, valued at $1.5bn with Microsoft among its backers, into insolvency in May 2025; and Amazon's Just Walk Out, where The Information reported in April 2024 that a large share of transactions passed through human reviewers in India. Both are contested — Amazon disputed the characterisation and said reviewers validated a small minority of visits, and at least one detailed engineering account argues the Builder.ai story as commonly told is wrong — and section 4 records the disputes rather than the punchline. What survives them is narrower and still worth citing: human labour is a load-bearing and systematically under-disclosed input to systems marketed as automatic. The film's line is the most economical way to say that to a general reader, provided the reading does not use it to claim more.

Where the analogy fails, stated plainly because it fails at the point of maximum moral weight: nobody dies to make training data. Soylent Green is about the industrial processing of human corpses. Scraping a book, buying a Slack archive and underpaying an annotator are, in order, a copyright question, a consent question and a labour question, and none of them is cannibalism. A reading that lets the film's horror bleed into its subject has made an argument by adjacency. The entry earns its place on structure — an opaque conversion step, a marketed claim about inputs, and no audit route for the consumer — and on nothing else.

What it got right, and what it got wrong

Claim date: April 1973. Due date: 2022, stated on the film's own title card. This is the only fiction entry in this canon with both, which makes it the only one that can be graded the way a prediction is, and it should be, because the grade is instructive in both directions.

Wrong, headline and by a wide margin — 40 million people in New York City. The 2022 figure is just under nine million. Off by roughly 4.5×. Due 2022, graded wrong. This is not a near miss to be softened; the number is the first thing the film says.

Wrong — the mechanism behind the number. The film inherited Ehrlich-era demography, in which population growth outruns food production and the result is mass famine in the rich world. What happened instead was the Green Revolution, sustained yield growth, and a fertility transition sharp enough that the live demographic anxiety in 2026 runs the other way — ageing populations and shrinking workforces. Due 2022, graded wrong, and wrong at the root rather than in the details. A reading citing this film for anything about population is citing the part that failed hardest.

Wrong — the character of the food problem. Soylent Green imagines the rich world starving on rationed wafers. The rich world's food pathology in 2022 was abundance: obesity, oversupply, waste. Due 2022, graded wrong. Worth naming because the inversion is the same one alphaville-1965 suffers — the danger arrived warm rather than cold, plentiful rather than scarce — and two films failing in the same direction is a fact about the genre's imagination rather than about either film.

Right, and early — the greenhouse effect, named as such. The film is routinely credited as one of the first mainstream pictures to put anthropogenic warming on screen, in 1973, with pollution and ecological collapse as the background condition rather than the plot. Due 2022. Graded: right in kind, wrong in degree and wrong in tempo. There is no permanent brown haze over New York and no year-round unbroken heat, and the film's version is a caricature. But of everything on the title card, the atmosphere is the item where the direction of travel was correct and the mainstream consensus of 1973 was not yet settled. Credit where it is due, sized honestly.

Right — the audit problem is the film's real subject, and it has aged better than anything else in it. Strip the cannibalism and what remains is a consumer who cannot inspect what they consume, a company whose only public account of its inputs is a document it wrote itself, and a truth reachable only by physically following the material. Due: continuously. Graded right, and this is the entire reason the entry is admitted. The 2026 record contains a reporter riding a shipment because there was no document, and a statutory template invented because company self-description was not enough.

Right — the checkers are marginal, elderly and dwindling. The Books are the film's audit institution and its most quietly accurate invention: a class of people who can read a technical document and place it in context, kept on sufferance, dying off, and disbelieved when they describe what things used to be like. Due: continuously. Graded right in shape. The Legislative Counsel item of 17 August 2026 is a working example — an office of skilled human checkers absorbing machine-produced drafts faster than it can check them, with a second reported effect that the people who draft with a model do not learn the issue as deeply, which is how a Book stops being made.

Right, and by accident — pre-collapse human material is the scarce good, and it is shown to the dying. Sol's twenty minutes at Home is footage of a world that no longer exists, preserved intact, played to people leaving. In 2026 the scarce, expensive, actively hunted commodity is human-generated text produced before the models — out-of-print books worth destroying, a dead company's Slack archive worth $100,000. Due: nothing; nobody claimed this. Graded: I record it because it is the sharpest correspondence in the film and because it is a coincidence, in the same class as Alpha 60 speaking entirely in borrowed literature. Nobody in 1973 foresaw a data market. A reading must present it as resonance and never as foresight.

Wrong, and this is the most important correction the entry carries — the secret is not a secret. The film's entire engine is concealment: one fact, known to a board and hidden from everyone, which will end the system the moment it is said out loud. That is the reassuring shape, and the 2026 record falsifies it. Amazon says on the record that it buys books. The Slack-archive market was reported in April 2026 and again in August, in the trade press, with the buyers named. The AI Act summaries are published documents. Model collapse is in Nature. There is no concealed fact waiting to be shouted in a church; there is a large volume of disclosed fact producing no consequence. Due: continuously. Graded wrong, and the failure mode it obscures — disclosure without consequence — is harder to write a thriller about and is the one actually operating. This is the sentence a reading must carry if it uses the film at all.

Wrong — one company, one product, one plant, one detective. Soylent is a single corporation with a single flagship product processed at a single facility a man can walk into. Due: unspecified. Graded wrong in exactly the way 2001-hal-1968 and alphaville-1965 are wrong, which now makes three entries in this canon recording the same genre error: the object of concern is drawn as one thing in one place with one door. Training corpora are assembled from scraping, licensing, purchase, annotation contracts and synthetic generation across many jurisdictions, by several companies at once, and there is no plant to break into.

Wrong — the tempo and the ending. Thorn learns the truth, shouts it, and the film stops on the shout. The 2026 versions are an enforcement date, a template nobody agrees is adequate, a price band in a trade publication and a tracker in a book. Nothing resolves in a reel, and the crowd, in the film's own last shot, is not shown reacting.

Ungraded — the euthanasia facilities. Assisted dying is lawful in a number of jurisdictions in 2026, on medical grounds, with the person's own request as the operative fact. The film's Home is state population management with a disposal pipeline attached, which does not exist anywhere. The two are not the same claim and there is nothing here to grade; it is noted only because the film is periodically conscripted into that debate on the strength of a visual resemblance.

Commonly misused as

Not required for fiction. The entry is worth much less without it, because this film has been reduced to a single quotable line and the line is now attached to arguments the film does not support.

• "Soylent Green is people!" as a punchline for anything undisclosed. It is a stock joke, and every use of it as a joke costs the film its actual argument, which is about auditability rather than disgust. If a reading reaches for the line, it should be reaching for the survey report — the scene where people read a company's own document and find that the company's product claim cannot be true.
• "The AI is secretly just people." Ghost work is real and under-disclosed, and the honest claim is that human labour is a load-bearing input to systems marketed as automatic. The strong claim — that a given product is a fraud with humans behind a curtain — is a factual allegation about a named company, and in both standing examples it is disputed: Amazon rejected the Just Walk Out characterisation and said reviewers validated a small minority of visits, and at least one detailed engineering account argues the Builder.ai story as popularly told is wrong. A reading that repeats the punchline without the dispute has published an allegation and called it a reference.
• "Model collapse means the internet is poisoned and the models are doomed." The Nature result is real and the strong reading of it is not established. Gerstgrasser, Schaeffer et al. (arXiv:2404.01413, April 2024) show that replacing data with synthetic data collapses while accumulating it does not — the test error acquires a finite bound independent of iteration count. A later position paper by Schaeffer, Kazdan, Arulandu and Koyejo (arXiv:2503.03150, March 2025) counts eight conflicting definitions of "model collapse" in the literature and argues several of the prominent catastrophe scenarios are readily avoidable under realistic training practice. Cite the film for the loop, cite Shumailov et al. for the effect, and do not let the film's certainty transfer to a contested result. This canon holds godel-incompleteness-1931 for the same reason.
• "Soylent Green predicted 2022." It named 2022 and got the headline number wrong by 4.5×, got the famine wrong, and got the direction of the population problem wrong. It is the clearest available demonstration that a film can be structurally perceptive and numerically worthless in the same ninety-seven minutes. Cite it to explain a finding, never to support one.
• "This shows overpopulation was the real threat." The film's demography is its most decisively falsified element and its parent text is a standing example of confident forecasting failure. Using this film in a population argument is using the part that broke.
• "The companies are hiding it." The correction graded above, restated as a misuse because it is the one this project is most likely to commit. The 2026 provenance record is mostly disclosed — statements on the record, published summaries, trade reporting with named buyers, a lab documenting how to defeat its own watermark. A reading that imports the film's concealment plot will describe a cover-up where the record shows publication without effect, which is a different problem with different remedies.
• "It's evidence." It is not. Fiction is not evidence, this canon is not the ledger, nothing in this file enters history/events.jsonl, and nothing here places or moves the needle.

Sources

Primary: the film Soylent Green (MGM, released 19 April 1973, 97 minutes, dir. Richard Fleischer, screenplay Stanley R. Greenberg). Dialogue is paraphrased rather than quoted except for "Soylent Green is people!" and the title card, both of which are attested independently in multiple places. A transcript fetch returned a summarised rather than verbatim rendering of the Exchange scene and Sol's death, so no line above rests on that single source and nothing from it is presented in quotation marks.

Reference: Wikipedia, "Soylent Green" — release date, running time, cast and credits, the 2022/New York/40,000,000 setting, the Soylent Corporation's product line and the plankton claim, the Soylent Oceanographic Survey Report 2015–2019 and the Books at the Supreme Exchange, Sol Roth's death at the assisted-dying facility, the waste disposal plant, the Saturn and Nebula awards, the AFI ranking of the closing line at 77, Ebert's and Siskel's verdicts, the Rotten Tomatoes consensus, Edward G. Robinson's death on 26 January 1973 and this being his 101st and final film, and the Computer Space cabinet as the first video game in a feature film.

For the novel and the adaptation dispute: Harry Harrison, Make Room! Make Room! (Doubleday, 1966), set in a New York of 1999 with 35 million people and a "soylent" made of soy and lentils; publishers' descriptions from Penguin and Open Road Media for the setting; Harrison's own assessment quoted from Omni's Screen Flights/Screen Fantasies (1984) as reproduced in the Wikipedia article, including the "50 percent" grade and the objection to the murder and chase sequences. That he was contractually denied screenplay control and not told MGM was the buyer is reported in the same secondary material and I did not reach a primary statement of it.

For the 2022 grading: Andrew Maynard's assessment in Slate, August 2022, for overpopulation having failed to become the crisis it was taken for in the 1970s, for agricultural productivity having been extended by high-yield crops and genetic engineering, and for the film being credited with the greenhouse effect; Washington Post (January 2022), Mental Floss and Science Friday carried comparable fifty-year gradings and agree on the New York population figure. The film's ~40m against the actual ~8.8m is arithmetic, not a source.

For model collapse: Shumailov, Shumaylov, Zhao, Papernot, Anderson and Gal, "AI models collapse when trained on recursively generated data," Nature 631(8022):755–759, 24 July 2024, doi:10.1038/s41586-024-07566-y — the Nature page was not reachable from here (it redirects to an identity provider) and a PDF copy in the University of Edinburgh's repository did not render, so the abstract quoted above is taken verbatim from the preprint, "The Curse of Recursion: Training on Generated Data Makes Models Forget," arXiv:2305.17493 (27 May 2023), whose closing sentence is the one cited. Nature has since issued an Author Correction (doi:10.1038/s41586-025-08905-3) which I did not read. The counterweights: Gerstgrasser, Schaeffer, Dey, Rafailov et al., "Is Model Collapse Inevitable? Breaking the Curse of Recursion by Accumulating Real and Synthetic Data," arXiv:2404.01413 (April 2024); and Schaeffer, Kazdan, Arulandu and Koyejo, "Position: Model Collapse Does Not Mean What You Think," arXiv:2503.03150 (March 2025), for the eight-definitions count and the avoidability argument.

For the EU AI Act: Article 53(1)(d) of Regulation (EU) 2024/1689 for the "sufficiently detailed summary" of training content on the AI Office's template; law-firm and vendor commentary (WilmerHale, Bird & Bird, Clifford Chance, Securiti) for the template's contents and for the dates — obligation from 2 August 2025, models already on the market given until 2 August 2027, AI Office verification and corrective powers from 2 August 2026. The claim about which providers completed the template and which answered in prose comes from a single secondary post (Pebblous) that carries no date or byline I could establish, was not corroborated, and is recorded above as reported and unconfirmed, including the part that is inconvenient for the model writing this.

For ghost work: Rest of World's account of Builder.ai's collapse (2025) and the insolvency reporting around it; The Pragmatic Engineer's rebuttal, "Builder.ai did not 'fake AI with 700 engineers'," which is why the dispute is recorded rather than the claim; The Information's April 2024 reporting on Just Walk Out human reviewers, with Amazon's response as given to Gizmodo and Axios. I read these as search summaries and secondary write-ups rather than the paywalled originals, and none of the disputed specifics should be repeated by a reading without going back to them.

The 2026 items — 404 Media's tracked rare book and the Amazon Las Vegas scanning facility (17 August), the market in defunct startups' Slack archives and Jira tickets (15 August), Anthropic's published watermark detail and its disclosed limits (15 August), Reddit's synthetic narration (17–18 August), the House Office of Legislative Counsel account (17 August), and the AI Act general-purpose regime as the live instrument since 2 August — are as recorded in this project's own digests of 2026-08-16, 2026-08-17 and 2026-08-18, which hold the primary links. They are named here as the citation occasion. They are not evidence deposited by this entry, nothing in this file enters history/events.jsonl, and nothing here places or moves the needle.