← The canon · AItopiaOrAImageddon?

Back to the Future Part II

fiction · Robert Zemeckis, from a screenplay by Bob Gale · 1989

A story the culture argues through. Carried graded — what it got right and what it got wrong, against dates.

Descends from The Time Machine.

Filed correctly. proposals.md lists this under ## fiction, and that is where it stays — but the filing has to be argued for at length, because the proposal's own text argues against it. The candidate entry says the reason to write this at all is that "the date has passed, so it grades like a prediction rather than reading like a story." That is true, and it is the whole value of the entry, and it is still not enough to change the kind.

This canon has now run this argument three times and the sequence is worth stating because it is close to a scale. time-machine-1895 has nothing due: the far vision is A.D. 802,701, there is no horizon anyone could be held to, and section 3 of that entry grades only the near-term mechanism claims. soylent-green-1973 is due and was graded: a title card names 2022, New York, and forty million people, which came due four years ago and was wrong by a factor of about four and a half. This entry is due to the day. The film puts 21 October 2015 on a dashboard readout, the world noticed when the day came, and the day is now three thousand nine hundred and fifty-four days behind us. Nothing else in this canon's fiction is due to the day. That is the admission argument and the grading argument at once.

Why it is still fiction, and the reason is harder here than in either earlier case. The line this canon draws is the one soylent-green-1973 states: prediction is reserved for a named forecaster who offered a claim as a forecast, and who can therefore be graded against their own scoring of themselves. Greenberg never scored himself; there was no predictor there to grade, only a number a screenwriter inherited from Ehrlich and put on a card. That argument does not work here, because the creators did grade themselves, in public, on the day. Bob Gale gave interviews on and around 21 October 2015 itemising hits and misses. John Bell, the visual effects art director, did the same. There is a self-assessment on the record, which is the one thing the prediction spec most wants and the thing soylent-green-1973 could not supply.

The answer is that a self-grade delivered after the fact is not a claim offered as a forecast, and the difference is not a technicality. Both principals stated, on the record and before and after the fact, that forecasting was explicitly not what they were doing. Zemeckis, in an interview recorded for the film's Blu-ray release, said the future scenes were the part of the shoot he enjoyed least, because he dislikes films that try to predict the future — the only ones he had liked were Kubrick's, "and not even he predicted the PC when he made A Clockwork Orange. So, rather than trying to make a scientifically sound prediction that we were probably going to get wrong anyway, we figured, let's just make it funny." Gale put the production's working rule the same way: "We said, nobody ever predicts the future correctly. It's not going to happen. So, instead of trying to predict the future correctly, let's not be unaware of what people think the future might be like, but let's just have fun with it." What got graded in 2015 was a joke, and it got graded because a calendar date arrived and a press cycle went looking for someone to ask.

Note that this is the exact inverse of frey-osborne-2013, the one entry in this canon whose authors deny being what it is filed as. There, a genuine forecast was offered and the forecasters disclaim the label; the canon files it prediction anyway, because what was offered governs. Here, no forecast was offered and a grade exists anyway. The rule is the same in both directions: the kind records what was put forward, not what the world later did with it. Frey and Osborne put forward a number. Zemeckis and Gale put forward a gag about fax machines.

And the absence of intent is precisely what makes this entry worth having. The proposal is right that it is "cleaner here because nobody was defending a thesis," and the point deserves to be stated in the canon's own terms rather than in passing. Every other forecast in this canon is contaminated by advocacy. kurzweil-2005 is selling a thesis and a book and later grades himself generously against outside reviewers who grade him lower. musk-robotaxi-2019 is selling a car to investors on the day the forecast is made. frey-osborne-2013 are defending a methodology. amodei-2024-loving-grace is written by a principal with a company. In every one of those, you cannot separate the forecaster's model of the world from the forecaster's interest in a particular future.

Here you can, because there is no interest. Nobody's funding, reputation, book sales or share price depended on 2015 looking any particular way, and the production's stated aim was that the audience should enjoy the place. What that leaves is a uniquely clean specimen: a dated, itemised, high-resolution snapshot of what an ordinary well-informed American imagination in 1989 expected the world to contain twenty-six years later, assembled by professionals with strong incentives to be recognisable and none at all to be right. It is a control. Every other forecast entry in this canon is an experiment with the researcher's thumb somewhere on the scale; this one has no thumb, and the failure mode it exhibits is therefore attributable to the imagination itself rather than to the arguer.

Research caveat, stated up front, in the same place total-recall-1990 states its. I did not screen the film for this entry. Every scene detail, line of dialogue and production fact below comes from secondary sources — principally the film's Wikipedia article, the Back to the Future fan wiki, quote databases, and the October 2015 press retrospectives — and where a detail rests on the fan wiki alone I say so at the point of use. The Zemeckis quotation reaches me through Wikipedia's citation of the Blu-ray interview and not from the disc. The Gale and Bell quotations come from the interviews named in Sources, one of which (the Today piece) returned 403 to a direct fetch and is quoted from search-result extraction rather than from the article as rendered; that one is marked where it appears. A later run with the disc and the paywalls should verify wording rather than assume it.

descends_from: [time-machine-1895], drawn on convention rather than on credited influence, and the ground has to be stated because the obvious grounds are wrong. The obvious ground would be "both are time-travel stories," which is resemblance and this canon does not draw on resemblance. The documented origin of the trilogy is not Wells at all: Gale has told the story many times of finding his father's high-school yearbook and wondering whether they would have been friends as classmates, and the 2015 sequence exists because Part I ended on a joke about not needing roads.

What I am drawing the edge on is narrower and load-bearing for this entry specifically. Wells did not invent time travel; he invented the future as a place a named traveller goes to, inspects, and reports back on — a machine with controls and a destination, an inventor who demonstrates it to sceptics in his own drawing room, and a return with a description. That convention is what this film runs on whole: an eccentric inventor, a vehicle, a set destination, a tour, a return. And the convention has one consequence that matters more than any other for this canon: it puts a date on the instrument panel. Wells's Traveller reads 802,701 off a dial; Doc Brown sets OCT 21 2015. A story that visits a dated place acquires a due date whether or not anyone intended a forecast, and a due date is the only thing that makes a fiction gradeable. Wells created the machinery by which this film became checkable against a Tuesday in October. That is the edge, on that ground and no other, and it is drawn as convention rather than as documented derivation because no documented derivation exists.

Ancestors I could not draw. The immediate one is the first film — Back to the Future (1985), which is not in canon/ and for which I am not inventing an id; this entry's 2015 sequence is a direct continuation of its closing scene and any honest graph would carry that edge. Beyond it, the real lineage of the film's 2015 is not literary but exhibitionary: the World's Fair "World of Tomorrow" tradition, the Futurama pavilions, the Popular Mechanics house of the future — the whole American habit of rendering the future as a consumer showroom, which is the tradition this film is working inside and gently mocking. None of that is in canon/. The production also defined itself against a work not in the canon: production designer Rick Carter's stated brief was a detailed image "with a different tone from the film Blade Runner," which is a negative influence and a real one, since the choice to make 2015 bright, cluttered and commercial rather than dark and sterile is the single decision that most shaped what the film got right and wrong.

What it is

A Universal Pictures sequel, released 22 November 1989, 108 minutes, directed by Robert Zemeckis, screenplay by Bob Gale from a story by Zemeckis and Gale, made for about $40m and taking about $333m worldwide. Roughly the first third of it is set on 21 October 2015, in Hill Valley, California — the date and place Doc Brown sets on the DeLorean at the end of the previous film — and that third is the reason this entry exists. The rest of the picture is a time-travel farce that returns to 1955 and to an alternate 1985; section 2 says what, if anything, that half is good for.

The 2015 sequence is a shopping list. It is worth having the list in full, because every graded claim below refers to an item on it, and because the shape of the list is the finding.

Transport. Cars fly, on marked skyways, with a Skyway licence; the taxis were designed from a Citroën DS. Doc's line on arrival — that where they are going they do not need roads — is the previous film's closing joke made literal. Petrol stations survive but are automated: a Texaco forecourt where robot arms service the vehicle.

Wearables and clothing. Marty's Nike "Air Mag" trainers lace themselves — "power laces." His jacket dries itself and adjusts to fit ("size-adjusting"). Marty Jr. wears JVC video glasses at the dinner table and takes phone calls through them; his sister Marlene does the same; the glasses also carry television, and Marty Jr. complains he "can only watch two shows at once."

The house. A wall-sized flat panel showing six channels at once, voice-operated. Windows that run a "scenery channel." A Black & Decker hydrator that turns a palm-sized disc into a full-size pizza. Thumb-pad locks on the front door. And a fax machine in every room, including the bathroom.

Communications and work. A video telephone in the den. When Douglas J. Needles calls Marty Sr., text scrolls beneath his face on the screen: occupation, age, birthday, address, wife's name, children's names, food preferences, drink of choice, hobbies, and politics. Needles talks Marty into an illegal scheme over the same link; their employer, Ito T. Fujitsu of CusCo, has been on the line, appears on the screen, and fires Marty on the spot. A fax then prints YOU'RE FIRED!!! in every room of the house.

Media and money. A holographic shark advertising Jaws 19, directed in-fiction by "Max Spielberg," which lunges at Marty in the street; the joke is that he says the shark still looks fake. Payment by thumbprint — the cab fare, and a USA Today headline about "thumb bandits." USA Today itself, still a printed paper, delivered and photographed by camera drones; another drone walks a dog. The paper's front page also carries "Queen Diana will visit Washington."

Service. The Cafe 80's, a nostalgia restaurant whose waiters are video simulacra — animated screen likenesses of Ronald Reagan and the Ayatollah Khomeini, talking over each other and pitching the specials, while Marty tries to order a Pepsi. (The fan wiki also lists Michael Jackson among them; I could not confirm that from the finished film and flag it as unverified.) Reagan had left office in January 1989 and Khomeini died that June: these are the likenesses of two living-or-just-dead heads of state, animated without their participation, deployed as customer service, for a laugh, in 1989.

Law. Marty Jr. is arrested, tried, convicted and sentenced within two hours. Doc explains: "The justice system works swiftly in the future now that they've abolished all lawyers."

And the design method, in the designer's own words, which is the most valuable single sentence in the research for this entry. John Bell, the visual effects art director, described how he built the 2015 imagery: "I'll take something that's 85 percent familiar — it could be a shape, it could be a color, it could be a pattern — and then if I'm projecting the future, throw something that's just a little 15 percent twist."

That is not a description of a design preference. It is an operational statement of a forecasting method — extrapolate the visible present, and perturb it slightly — offered without embarrassment by a practitioner who had no reason to disguise it, and it predicts, in advance, the exact distribution of hits and misses that section 3 grades. Every item on the list above is an object that existed in 1989 with a modifier attached: a car that flies, a skateboard that hovers, a television that is bigger, a fax that is everywhere, a shoe that ties itself, a phone that shows a face. The method cannot produce a thing that has no 1989 object to be 85 percent familiar to. Hold that sentence; the rest of this entry is a test of it.

Gale's stated aim for the tone, from the same 2015 interviews: "We want the future to be a positive, optimistic future. We want people to see this movie and say, 'Hey I like that future, I'd like to live there.'" Against the 1980s screen default — Blade Runner, The Terminator, RoboCop — that was a deliberate and unusual choice, and it is graded below as a hit in its own right.

Why a reading would cite it

The occasion is a forecast being scored — and the specific contribution, the thing no other entry in this canon supplies, is a clean demonstration that forecasts fail by scaling up the visible present, not by being too bold. That is the citation, and everything else in this section is either the concrete case for it or a brake on it.

The 2026 record this project is reading produces the occasion constantly, because the AI build-out is almost entirely a set of dated commitments made against extrapolated curves. Four items from the last four days, cited at their underlying sources rather than at this project's digests of them, because digests are output and output is never evidence:

Etched, 18 August 2026. The transformer-only inference-chip company is reported to have closed $700m at a $21bn valuation, roughly double the $10.3bn set by a $300m Sequoia-led round in late July, itself double December's $5bn. This is the sharpest Back to the Future occasion in the current record and it is not obvious, so it is worth spelling out. Etched's entire thesis is to burn one architecture into silicon: take the visible present — the transformer, which has been the shape of the field since transformer-2017 — and commit it to hardware in exchange for efficiency. That is not a stupid bet; it may well pay. But it is structurally the fax machine in every room. It is the 1989 object with a bigger number attached, and its risk is not that the transformer will stop working but that the organising development of the next decade will have no 2026 object to be 85 percent familiar to. Bell's sentence and Etched's prospectus describe the same manoeuvre.

The commitments, 15–18 August 2026. This project has on its record $3tn of off-balance-sheet commitments read out of filing footnotes, a signed 20-year 10GW lease, and a chip supplier guaranteeing its customers' debt — set against an Anthropic annualised run rate reported at about $65bn at the end of July, up from roughly $9bn eight months earlier, with 2028 projections of $190–200bn attached. A twenty-year lease priced off an eight-month curve is an extrapolation of the visible present with a modifier attached, and the entry's warning applies symmetrically: the curve may be right and the category wrong. Nobody in 1989 was wrong that document transmission would grow. They were wrong that it would grow as faxes.

The Microsoft chip discrepancy, 17 August 2026. A Guardian investigation of internal documents put Microsoft at about 2.2 million AI chips installed against roughly $280bn spent and 5GW added, where 10GW of AI datacentre would imply nearer 6.4m GPUs (Inside Telecom; eWeek); Microsoft says the estimates draw wrong conclusions from incorrect assumptions without naming a wrong figure. This is a forecast being scored mid-flight, against an inferred denominator, by outsiders — the ordinary condition of AI forecasting in 2026, and the condition in which this entry's caution about the unit of account (below) does the most work.

Unitree's prospectus, 18 August 2026. The humanoid manufacturer's Shanghai STAR listing forces publication of units shipped and revenue on a schedule — more than 5,500 humanoids shipped in 2025, 595.2m yuan of humanoid revenue in the first nine months (Forbes; Quartz). A sector argued almost entirely on demonstration videos acquires, for the first time, a scoreboard. The same is true of Operation Blue Skies, launched the same day with an independent verification arm that makes its claim gradeable in eighteen months rather than merely repeated. When a forecast becomes gradeable is exactly when this entry becomes citable, and both of those became gradeable this week.

Now the three things this entry actually says, in the order of how much they are worth.

One: the failure mode is timidity, not boldness. Sort the film's misses by how daring they were. The flying car was the boldest thing in the sequence and it was wrong. The fax machine in every room was the timidest and it was wrong worse — wrong in the direction of extinction rather than delay. Meanwhile the thing that actually reorganised 1989–2015, the network in every pocket and everything that grew on it, was bolder than anything the film dared, and the film did not miss it because it lacked nerve. It missed it because the method Bell described cannot generate it: a network is not an object, and there was no 1989 consumer object for it to be 85 percent familiar to. A telephone with a screen is an object, so the film got that. A protocol with two billion endpoints is not, so the film could not.

This is why the entry is more useful than kurzweil-2005 for the same purpose, and the two should be read together. Kurzweil's characteristic error is a rate error on an axis he correctly identified. This film's characteristic error is an axis error: the axis it extrapolated hardest along (personal transport, printed document transmission, home appliances) was not where the period's change happened, and it never named the axis where it did.

Two, and this is the reason a canon about AI has any business holding a comedy about a DeLorean: the film's 2015 has automation everywhere and intelligence nowhere. Count the machines. Every single one obeys. The house responds to a command, the hydrator runs a cycle, the drone follows a subject, the thumb pad checks a print against a record, the trainers tighten, the fax prints what it is sent. There is not one system in the film's 2015 that decides anything, forms a view, is uncertain, or is wrong in an interesting way. The nearest thing to a judgment in the whole sequence is a legal system that has been made fast by removing the lawyers — which is to say, by removing the judgment.

That is a load-bearing observation and no gadget tally in existence records it, because tallies count objects and this is a missing category rather than a missing object. Thirty-three years after Dartmouth, twenty-three after ELIZA, two after moravecs-paradox-1988, popular futurism could imagine a house full of machines and could not imagine one of them thinking. A reading that needs to say something about how recently machine intelligence was absent from the ordinary picture of the future has, in this entry, a dated and unusually well-attested specimen. Cite it for that and it does work nothing else in the canon does.

Three, and it is a caution rather than a claim: a popular forecast is an input to the world it forecasts, which corrupts its own grade. Nike's Tinker Hatfield designed the film's power-lacing shoe as fantasy in 1989, Nike filed a self-lacing patent in 2009, produced self-lacing Mags for the 2015 anniversary, and shipped the HyperAdapt 1.0 to consumers on 1 December 2016 at $720. Gale's own account of the hoverboard is the same shape: "kids like you saw 'Back to the Future part 2' back in the day and said well, if I can't have a hoverboard... I'm going to make one." The shoe is not a hit. It is a commission — the forecast caused the artifact, and scoring it as a successful prediction double-counts the film.

The application to 2026 is direct and uncomfortable. The loudest AI forecasts of this period are not passive descriptions; they move capital, staff and policy, and some of them are written by people who then go and build the thing. A forecast that raises $700m to build the future it describes cannot be graded the way an obscure one can, and neither can a forecast that talks a government into a datacentre. Before scoring any well-publicised AI forecast, ask whether the forecast was an input. That question costs one sentence and this entry is the cleanest available illustration of why it has to be asked.

The proposal's secondary use, which I am recording as real but weaker. The candidate text suggests the trilogy's argument — that small changes compound into unrecognisable outcomes — supports "tracking drift over a series rather than reading any single point," which is close to this project's own method. The alternate-1985 sequence, in which Biff's stolen almanac produces a Hill Valley nobody recognises, is genuinely a picture of a system arriving somewhere unrecognisable. But be careful: that compounding is triggered by one discrete, dramatic, deliberate intervention, not by drift. It is a counterfactual fable — remove this act and the outcome vanishes — and drift is the opposite thing, an outcome with no single act to remove. Cited loosely it would let a reading dress up a butterfly-effect intuition as a methodological argument. The defensible residue is thin and I will state only that much: the sequence is a good picture of how far a state can end up from where it started, and it is not evidence about how states move.

The brake, which this entry needs as badly as total-recall-1990 needs its. The failure mode here is the smug one. It is very easy to cite this file as "forecasters are always wrong, look at the flying cars," and that reading is not merely lazy — it is the exact opposite of what the record shows, and section 4 says so at length. Before citing, the test is: **am I about to say something about how a forecast failed, or am I about to use a 1989 comedy to avoid grading a 2026 claim?** Only the first is an occasion. The needle is not moved by either; nothing in this file is evidence.

What it got right, and what it got wrong

Claim date: 22 November 1989. Due date: 21 October 2015 — a shade under twenty-six years out. Days elapsed since due, as of this entry: 3,954. No entry in this canon is more fully due, and none has a cleaner grading surface, because the claims are objects and objects either exist or they do not.

Three gradings follow, and they differ, which is the point of running all three.

The creators' own grade

Recorded here as the self-assessment the prediction spec asks for, even though this entry is not filed as one, because the divergence below is the useful part.

The shape of the self-grade: honest on the misses, generous on the mechanism. Gale concedes the two big ones without being pushed, which is more than kurzweil-2005 manages. But three of his four claimed hits — drones, thumbprints, video calling — are claimed as foresight when the same interviews describe them being read out of 1989 trade coverage and extrapolated. And the largest fact about the film's 2015 is not in his list at all, because it is the network, and a list of objects has no slot for it.

The press grade, and why it runs high

The October 2015 tallies were numerous and almost uniformly generous — "surprisingly accurate," "surprisingly prescient," "more so than most films that take forecasting seriously." They are generous for a structural reason worth naming, because it is a methodological point this project can use: they count objects, and the unit of account decides the score.

Score by object and the film does well: video calling, wall-sized flat panels, multiple simultaneous feeds, wearable displays, biometric ID, camera drones, voice-controlled homes, hands-free games, self-lacing shoes, tablets with fingerprint readers, talking billboards. Score by world and it does badly, because the single organising fact of 1989–2015 is absent, and its absence is invisible to a checklist — you cannot put a tick or a cross next to a thing the film never mentioned. Every tally therefore has a systematic upward bias, and it is not the tallies' fault; it is the unit's. Any scorecard of a forecast will overrate it if the forecast's worst error is a category it never named. That generalises directly to grading AI forecasts in 2026, where checklists of capabilities are the standard instrument.

This entry's grade

Right, and undercounted by everyone — the videophone dossier. The single most accurate thing in the film is not that the phone shows a face. It is that when Needles calls, his occupation, age, birthday, address, wife's name, children's names, food preferences, drink of choice, hobbies and politics scroll under his face, automatically, unrequested. In 1989 that read as a gag about a phone with too many features. It is a description of the ad-tech identity graph, delivered twenty-six years early, with politics explicitly on the list. And the same scene adds the part nobody quotes: the employer is silently on the line, hears the employee agree to something, and terminates him instantly and remotely — with the notice pushed automatically to every device in the building. That is algorithmic management and remote termination, staged as farce. It is a better hit than the hoverboard and the trainers combined, and no tally scores it, because it is not an object.

Right, and the second thing nobody scores — the synthetic persona of a real public figure, sold as a service. The Cafe 80's waiters are animated likenesses of Reagan and Khomeini, deployed commercially without their participation, played entirely for laughs. On 18 August 2026 this project recorded Meta's advertising system serving a pornographic deepfake closely resembling a prominent US politician, removed by Apple's store review rather than by the platform paid to run it. The film staged the mechanism in 1989 and got the affect exactly wrong — for Zemeckis and Gale a synthetic head of state is a joke about a theme restaurant, and the register is cheerful throughout. That is the same miss total-recall-1990 is graded down for, arrived at from the opposite direction, and it is a reason to cite this entry only for the mechanism and never for the tone.

Right, and the most socially precise thing in the film — the dinner table. Marty Jr. and Marlene sit at a family meal wearing video glasses, taking calls and watching television through them, ignoring the adults; Marty Sr. tells his son to take the glasses off because his grandparents are here; the boy complains that he can "only watch two shows at once." Every clause of that is now unremarkable, including the generational argument about it and the complaint that the device is insufficiently capable. The film got this without a network, without a smartphone, and without any underlying model of how it would happen — which is worth noticing in both directions: the social consequence was easier to see than the technical cause. A reading grading a 2026 forecast should hold that thought, because the cheap assumption runs the other way.

Right — wearable displays, biometric locks and payment, camera drones, wall-sized panels, many simultaneous feeds, voice-operated rooms, hands-free games. These are real hits and they are also the easy ones: every one is an existing 1989 product with a modifier, which is Bell's method succeeding on exactly the terrain where it can succeed. Grade them as hits and then discount them for difficulty, the way you would discount a forecast that next year will contain a December.

Right, and rarely credited — the texture. Against the 1980s screen default the choice to make 2015 bright, commercial, cluttered, cheerfully vulgar and full of nostalgia products was contrarian, and it is the film's most accurate single judgment about the character of the period. The actual 2015 was a nostalgia economy selling reissued sneakers. The 1980s dystopias were wrong about texture in a way this comedy was not, and it got there by aiming at pleasure rather than at accuracy.

Right, one year late and in the wrong series — the Cubs. The film's paper reads CUBS SWEEP MIAMI IN WORLD SERIES. On 21 October 2015 the Cubs were swept — 4–0, out of the National League Championship Series, by the Mets, 8–3 at Wrigley, on the day itself, while Gale was at an anniversary screening in New York. They won the World Series the following year, against Cleveland. The Miami half is the better call and Gale claims it: Miami had no team in 1989, was awarded a franchise in 1991, and began play in 1993. Grade: the improbable structural claim landed, the dated one missed by a year, and the word "sweep" came true in the other direction on the exact date. As a lesson in forecast scoring it is almost too neat — the same claim is a hit, a miss or a joke depending entirely on the tolerance you set before you look.

Right, thirteen months late, and it does not count — power laces. Nike's HyperAdapt 1.0 shipped 1 December 2016, four hundred and seven days after the due date, at $720, designed by the man who designed the film's prop. As set out above, this is a commission rather than a prediction and should be recorded as one.

Wrong — the fax, and it is the emblem of the whole failure mode. A fax machine in every room including the bathroom; the termination notice printed on paper in each one. This is the correct function (instant document delivery to everywhere a person might be) rendered in the only 1989 object that performed it, scaled up. The function arrived overwhelmingly and the object nearly vanished. Gale's concession — "it just seemed like there were going to be more and more faxes" — is a perfect specimen of extrapolating an object rather than a need, and it should be quoted whenever this entry is cited.

Wrong — flying cars, and wrong less than it looks. No consumer flying car exists in 2026, but the honest 2026 status is more interesting than a flat cross: eVTOL air taxis are at the last regulatory gates rather than in the realm of fantasy, with Joby reported into the final stage of FAA type certification and flying conforming aircraft, Archer reporting acceptance of its full means-of-compliance package, paying passengers flying in Dubai and China, and independent analysts still expecting US commercial service to slip to 2027 or later. Grade: wrong on the date, wrong on the form (a fleet service with pilots and vertiports, not a family car with a Skyway licence), and not wrong on the physics. Which is to say the boldest thing in the film is running roughly forty years late rather than being impossible — and that is a much better outcome than the fax achieved.

Wrong — hoverboards, in the way that most rewards attention. No board hovers. But the word arrived on schedule: in 2015 the self-balancing two-wheeled scooter took the name "hoverboard" and sold in millions, and separately Lexus and Arx Pax built working magnetic-levitation demonstrators requiring special surfaces, and Franky Zapata built a jet-powered board that flew. The vocabulary of the forecast was adopted by an object that does not do the thing. Watch for that in AI language: "agent," "reasoning," "understanding" and "autonomy" are all currently attached to products that do something adjacent to what the word originally promised, and this is what that process looks like completed.

Wrong, and the 2026 inversion is exact — the abolition of lawyers. "The justice system works swiftly in the future now that they've abolished all lawyers." Set that against what this project recorded on 17 August 2026: AI-drafted legislation swamping the US House Office of Legislative Counsel, drafts arriving with wrong terms and citations to the wrong statutes, and the office's lawyers reporting that reviewing and rewriting them takes longer than writing from scratch would have. The film's joke was that automation removes the lawyers and speeds the process. The measured 2026 finding is that automation multiplies the lawyers' work and slows the process, on the specific task the tools are said to be best at. This is the entry's single best-matched claim-and-outcome pair, and the direction of the error — automation eliminating a profession versus automation generating work for it — is one of the two or three most consequential open questions this project tracks. Read it next to frey-osborne-2013, which makes the same category of error with a methodology attached.

Wrong — Queen Diana; Mr Fusion; the hydrated pizza; self-drying jackets. Recorded for completeness. Nobody will cite this entry for these, and dwelling on them is what the lazy version of this file would consist of.

Wrong, structurally, and this is the grade that matters: it multiplied devices and missed the network. There is no internet in the film's 2015. Information reaches Hill Valley by broadcast television, by a printed newspaper delivered by drone, and by fax. Every "communications" hit in the film is a point-to-point device — a phone with a face, glasses that ring, a fax that prints. Not one thing in the sequence is a many-to-many system, a search, a platform, a marketplace, or a social graph, and therefore not one of the actual organising facts of 1989–2015 appears in it at any resolution. The film has more machines than 2015 had and less connection than 2015 had, and everything it got wrong follows from that one substitution.

The generalisation, which is what a reading should carry away: the method Bell described can only produce objects, because only objects can be 85 percent familiar. It cannot produce a network, a protocol, a market structure, a business model or a trained model, because those have no consumer form to perturb. So a forecast built that way will systematically get the gadgets about half right and miss the organising structure entirely — which is not a criticism of Bell, who was dressing a comedy, but is a precise warning about every 2026 forecast built the same way. Ask of any forecast: is it a list of things, or does it name a structure? If it is a list of things, this entry says what its errors will look like before you read it.

Commonly misused as

Not required for fiction. The entry earns the section anyway, because it is the most quotable forecast-scoring artifact in general circulation and every misuse below is live in 2026 argument.

Misused as a general deflator — "AI is just flying cars." This is the important one and it is backwards. The film's failures were not its bold claims; they were its timid ones. The flying car was audacious and is running late. The fax in every room was cautious and was annihilated. And the thing that actually happened — a global network, a device in every pocket, and eventually trained models on top of both — was bolder than anything the film had the nerve to put on screen. So this entry cannot support "the future is always less than promised." What it supports is the narrower and more useful claim that the future is differently shaped than promised, and the error is usually in the category rather than the magnitude. A reading that uses this file to dismiss a 2026 forecast without grading it has used it for the one thing it disproves.

Misused as sci-fi triumphalism — "the movie predicted video calls!" It did, and so did every futurist since the 1878 Punch cartoon, and AT&T had shipped the Picturephone at the 1964 World's Fair. Gale's own account is that thumbprint ID was "in development" and he "read articles back then." Scoring a forecast as a hit when it restates the trade press of its own year inflates the base rate, and this canon's whole reason for holding predictions is to keep an honest base rate for how wrong AI forecasting runs. Hits that were already reported at the time of forecast should be graded down, not up.

Misused by counting the commissioned outcomes. The self-lacing Nike is on nearly every "things it got right" list. Nike built it because of the film, with the film's own designer, timed to the film's anniversary. It is a hit in the same sense that a prophecy of one's own arrival is a hit. Anything a forecast caused must be struck from its score, and in 2026 — where forecasts raise the capital that builds their subject matter — this is not a pedantic rule.

Misused as evidence about anyone's forecasting ability. There was no forecaster. Zemeckis said he dislikes films that predict; Gale said the marching order was that nobody predicts correctly so have fun. Grading this as though a claim had been offered is exactly the error the fiction/prediction distinction exists to prevent, and it is why the opening of this file spends as long as it does. The film is a specimen of 1989's ambient expectations. It is not a forecaster's track record, and it does not belong in a base rate assembled from forecasters.

Misused as a claim about drift. See section 2 on the alternate-1985 sequence: it is a counterfactual fable turning on one deliberate act, and it is not evidence about how systems move by accumulation. It is the citation most likely to be reached for by a reading that wants to sound methodological about this project's own series, and it is the one least able to bear the weight.

Sources

Primary.

Secondary.

2026 citation occasions, reached through this project's own digests of 15–18 August 2026 and cited here at their underlying sources, never at the digests: Etched's July round at $10.3bn (the August round at $21bn is reported rather than filed, and is recorded as such); Anthropic's reported $65bn run rate (17 Aug); the Guardian's Microsoft chip-count investigation via Inside Telecom and eWeek (17 Aug); Unitree's STAR Market listing (18 Aug); Operation Blue Skies (18 Aug); AI-drafted bills at the House Office of Legislative Counsel (17 Aug); the Meta nudify advertisement (18 Aug, Wired via Techmeme).

Failed or unreachable. Direct fetches of the Today interview and The Hollywood Reporter piece were refused (403 and a paywall redirect respectively); the Futurepedia article on video simulacra returned 402; the South China Morning Post retrospective was fetchable but truncated before its verdicts. None of these gaps changes a claim above, and each is the reason the corresponding quotation is attributed the way it is. The eVTOL status in section 3 is assembled from 2026 trade coverage of the FAA certification stages rather than from FAA dockets, and should be treated as the state of reporting rather than the state of the register.