We Have It on Camera
Ben Chanan's BBC thriller bet that the next crisis of evidence would be the footage itself, and three series later the courts are still drafting the rule.
Nearfi scores The Capture (2019) 7.6/10 as art. It built its whole plot on one question, whether anyone can trust footage, and asked it three times in three different ways. As prophecy, its P(n) is 96%: Real-time face-swaps have already moved millions on a video call, and police cameras now name people in the street, but nobody has shown a state quietly rewriting live CCTV.
- It built its whole plot on one question, whether anyone can trust footage, and asked it three times in three different ways.
- Each series is a chase thriller first, so the idea is dramatised more often than it is argued.
- Its villain is a government department that fixes evidence rather than a lone hacker, which is the less comfortable version.
Real-time face-swaps have already moved millions on a video call, and police cameras now name people in the street, but nobody has shown a state quietly rewriting live CCTV.
- ARRIVING Live video can be altered in real time, convincingly enough that investigators, courts and the public accept it as genuine footage.
- ARRIVING Police in a capital city can identify and follow people through a dense network of public cameras using live facial recognition.
- ARRIVING Courts operate formal rules for deciding whether contested video or audio evidence is authentic or a deepfake.
Real-time face-swaps have already moved millions on a video call, and police cameras now name people in the street, but nobody has shown a state quietly rewriting live CCTV.
The most frightening sentence in The Capture is a polite one. We have it on camera.
Ben Chanan's thriller for BBC One has run to three series of six episodes (2019, 2022 and 2026), made by Heyday Television, and it has never really been about a crime. It is about what happens to the word "evidence" once a picture can be edited faster than a jury can blink. It holds its one idea in both hands for eighteen episodes and never puts it down.
The art: a chase with a philosophy in its pocket
Series one starts well. Shaun Emery (Callum Turner), a former soldier, has had a conviction over video from Afghanistan overturned on appeal because the footage was potentially flawed. He walks out. Then London's CCTV appears to show him assaulting and kidnapping his barrister, Hannah Roberts (Laura Haddock). Detective Inspector Rachel Carey (Holliday Grainger) takes the case, and the show spends six episodes making you look at the same grey street camera from different angles until you stop being sure what a street camera is.
Grainger gives the show its spine: Carey is a detective who keeps her head while the evidence around her stops behaving. The design helps. For series three the production built a command centre in which a wall of more than 70 screens plays pre-recorded CCTV-style footage rather than relying on green screen, so the actors are looking at something.
Series two moves from the street to the studio. Isaac Turner (Paapa Essiedu), a politician, appears live on Newsnight backing a Chinese AI firm's bid to supply facial recognition for UK borders. He did not say it. The face and the voice were a live fake. Series three puts Carey in counter-terrorism, around a system called Operation Veritas that is meant to catch manipulated footage in real time.
The flaws are the flaws of the genre. It is a conspiracy thriller, so plot comes first, and the idea is dramatised more often than it is argued. That is the gap between a good thriller and a great argument. The show would rather get you to the next chase than let you sit with the question. The audience also shrank: average UK viewing went from 5.66 million for series one to 3.81 million and then 3.11 million, and no fourth series has been announced.
The prophecy: Correction
The show's central invention is a programme called Correction. Intelligence agencies alter live video feeds so that police and courts receive "evidence" that has been quietly rewritten. In series one a video engineer proposes that footage could be delayed slightly and then altered, with a passing bus giving cover for the join. The point is that the edit does not need to be perfect. It needs to arrive before anyone looks.
We should be careful about what has and has not arrived.
The real-time fake is here. In early 2024 a finance worker at the engineering firm Arup, in Hong Kong, joined a video call with what looked and sounded like the chief financial officer and colleagues. All of them were deepfakes. He made 15 transfers worth HK$200 million, about $25.6 million. Arup confirmed that "fake voices and images were used." That is series two's Newsnight trick, aimed at a bank account instead of a minister.
The all-seeing camera is also here. In London the Metropolitan Police ran fixed live facial recognition cameras in Croydon for six months from October 2025: about 470,000 faces scanned, 173 arrests and one false alert, according to the force's own figures. Civil liberties groups argue about the bias and the legal basis. The Home Office published its response to its own consultation today, and says it wants a statutory framework with an independent oversight body. Carey's London, in which every street is a feed, is closer than it looked in 2019.
The courtroom rule is half written. This is where fiction runs ahead. In the United States the federal evidence committee has spent years on a draft Rule 901(c) that would make a party who says a video is a deepfake produce enough evidence to support the claim, after which the other side must show the item is more likely than not authentic. It has been prepared but not approved. A separate proposed Rule 707 on machine-generated evidence was withdrawn for further study after more than 70 written comments. The Justice Department has said 901(c) is not ready for publication. Chanan imagined a state that forged evidence. The real difficulty is quieter: real footage can now be denied, and nobody has settled who must prove what.
The fix is half built, and it is not a detector. Operation Veritas catches fakes as they happen. The real answer is the reverse: sign the real thing at the moment it is made. Google says every JPEG from its Pixel 10 camera now carries C2PA Content Credentials, signed with keys held in the phone's security chip, and that the camera app reached the highest assurance level the standard defines. Google also calls it "not the sole solution." Credentials say where a file came from. They cannot say whether the scene was staged, and they are lost when a platform re-encodes the file. A provenance label tells you which pictures to trust. It cannot tell you what to make of the rest.
What the show missed is the boring route. Correction needs a conspiracy, a bus and a technician. The cheaper threat is a world where anyone can say "that could be fake" and be half right.
The reading
The Capture is a very good thriller with one idea, and the idea has aged better than the genre around it. It got the fake right, the cameras right and the courts nearly right, and it was too early to see that the dull defence, a cryptographic signature in a phone, is the one that might actually hold. No fourth series has been announced, and perhaps none is needed. The sentence it began with will outlive the show.
We have it on camera used to close an argument. In the mid-2020s it only opened one.
What it predicted, and what happened
The Capture is strongest on the threat of the live fake and on the all-seeing London camera, both of which now exist in early forms. It is ahead of the law, which has not settled how courts should treat contested video, and it misses how much of the damage comes from plausible denial rather than undetectable forgery.
How this P(n) was calculated
P(n) is the chance that this work's future arrives within 20 years, by 2046. For each prediction we turn our estimate into a probability: arrived is 100%, diverged is 0%, and otherwise the low end of our 80% range counts as a 10% chance, the median year 50% and the high end 90%. Each prediction is weighted 1–5 by how pivotal it is to the story, and P(n) is the weighted average. P50 does the same by 2076. P(n), explained.
| Claim | Prediction · why this weight | Weight | Status · estimate | By 2046 | By 2076 |
|---|---|---|---|---|---|
| CAPTURE-1 | Live video can be altered in real time, convincingly enough that investigators, courts and the public accept it as genuine footage. Core premise: the Correction programme, and every twist in all three series, depends on footage that cannot be trusted. |
5 | ARRIVING 2031 (2028–2042) | 94% | 100% |
| CAPTURE-2 | Police in a capital city can identify and follow people through a dense network of public cameras using live facial recognition. Major: the camera network is the setting and the detective's main tool, and it carries the whole plot. |
4 | ARRIVING 2030 (2027–2037) | 100% | 100% |
| CAPTURE-3 | Courts operate formal rules for deciding whether contested video or audio evidence is authentic or a deepfake. Major: the trials and the question of what a court will believe drive each series' stakes. |
4 | ARRIVING 2029 (2027–2034) | 100% | 100% |
| CAPTURE-4 | Police have a system that checks footage for manipulation in real time and whose verdict investigators and courts rely on. Supporting: Operation Veritas drives series three, but the story is about its failure. |
3 | ARRIVING 2032 (2028–2045) | 91% | 100% |
| P(n) = P20 = Σ(weight × chance by 2046) ÷ Σ weight = 15.4 ÷ 16 = 96% | |||||
| P50 = 16.0 ÷ 16 = 100% | |||||
Already here today (arrived 1, arriving ½, weighted): 50%.
-
CAPTURE-1
Live video can be altered in real time, convincingly enough that investigators, courts and the public accept it as genuine footage.
Weight 5/5 · Core premise: the Correction programme, and every twist in all three series, depends on footage that cannot be trusted.
Real-time face and voice fakes have already fooled people on live video calls: in 2024 an Arup finance worker in Hong Kong paid out about $25.6 million after a call in which colleagues were deepfake re-creations. That is a small, controlled audience, not a public broadcast or a government CCTV feed, and no public evidence shows any state altering live camera systems. The date reflects how fast the fakes improve against how hard it is to compromise a chain of custody at scale. 123
-
CAPTURE-2
Police in a capital city can identify and follow people through a dense network of public cameras using live facial recognition.
Weight 4/5 · Major: the camera network is the setting and the detective's main tool, and it carries the whole plot.
The Metropolitan Police uses live facial recognition on vans and, in a six-month Croydon pilot, on fixed cameras: about 470,000 faces scanned, 173 arrests and one false alert, on the force's own figures. It matches faces against watchlists and deletes the rest, which is not the same as following anyone. The government has just published its response to a consultation promising a statutory framework and an independent oversight body, and the High Court has ruled the Met's use lawful. Citywide tracking is the part that has not arrived. 456
-
CAPTURE-3
Courts operate formal rules for deciding whether contested video or audio evidence is authentic or a deepfake.
Weight 4/5 · Major: the trials and the question of what a court will believe drive each series' stakes.
The US federal evidence committee has drafted a Rule 901(c) that would put the first burden on whoever says an item is a deepfake, then require the other side to show it is more likely than not authentic. It has been prepared but not finally approved. A separate proposed Rule 707 on machine-generated evidence was withdrawn for more study in 2026 after more than 70 written comments. Federal rules take years, which is why the range starts in 2027. This ledger covers the US federal system only; England and Wales has no equivalent rule drafted that we could verify. 789
-
CAPTURE-4
Police have a system that checks footage for manipulation in real time and whose verdict investigators and courts rely on.
Weight 3/5 · Supporting: Operation Veritas drives series three, but the story is about its failure.
The nearest real thing is provenance, not detection: signing footage at capture. Google's Pixel 10 attaches C2PA Content Credentials to every JPEG it takes, and the camera app reached the highest assurance level the standard's conformance programme currently defines. The standard covers files from cooperating tools only, and it is a label of origin rather than a deepfake detector. Nothing public shows a police force relying on a real-time manipulation verdict. 101112
- FOR British engineering giant Arup revealed as $25 million deepfake scam victim. CNN Business, 2024-05-16. An employee was duped on a video call by deepfake re-creations of the CFO and colleagues, and sent 200 million Hong Kong dollars; Arup confirmed fake voices and images were used.
- CONTEXT Advisory Committee on Evidence Rules agenda book, October 15, 2026. Administrative Office of the U.S. Courts, 2026-09-22. The committee has spent seven meetings on deepfakes as an evidence problem, and calls audio and video authenticity an open issue.
- AGAINST How Pixel and Android are bringing a new level of trust to your images with C2PA Content Credentials. Google, 2025-09-10. Hardware-signed capture makes tampering with new photos detectable, which narrows the room for undetectable edits to files that carry credentials.
- FOR Met Police tout arrests, crime drop from permanent LFR camera pilot. Biometric Update, 2026-05-13. Fixed lamp-post cameras in Croydon scanned about 470,000 faces over six months and led to 173 arrests; the High Court ruled in April that the Met had not breached human rights or privacy law.
- CONTEXT Consultation on a new legal framework for law enforcement use of biometrics, facial recognition and similar technologies: government response. Home Office, GOV.UK, 2026-11-04. The government intends to legislate a framework with a new independent oversight body, after a consultation that ran from December 2025 to February 2026.
- AGAINST Metropolitan Police to expand live facial recognition use even amid legal challenge. Biometric Update, 2025-11-03. Ten people were falsely flagged in the Met's annual report period, eight of them Black, and Big Brother Watch and a wrongly identified man brought a legal challenge.
- FOR Advisory Committee on Evidence Rules agenda book, October 15, 2026. Administrative Office of the U.S. Courts, 2026-09-22. A draft Rule 901(c) on deepfakes is prepared but not finally approved; the committee withdrew proposed Rule 707 for further study.
- AGAINST Machine-Generated Evidence Challenges the Federal Rules: Inside the Contested Proposed Federal Rule of Evidence 707. Proskauer Rose LLP, 2026-07-29. In June 2026 the Standing Committee declined to act and sent Rule 707 back for revision, alongside the separate deepfakes issue.
- CONTEXT New Evidence Rule 707 Would Set Standards for AI-Generated Courtroom Evidence. Barnes & Thornburg LLP, 2025-08-21. Explains the original Rule 707 draft, which would apply expert-reliability standards to machine output offered without an expert.
- FOR How Pixel and Android are bringing a new level of trust to your images with C2PA Content Credentials. Google, 2025-09-10. Pixel Camera signs every JPEG, with keys held in the phone's security chip; Google calls it a first step and not the sole solution.
- AGAINST Big Tech's standard for fighting AI fakes puts privacy on the line. Fortune, 2025-09-18. Reports a World Privacy Forum finding that C2PA does not detect deepfakes, that forged images can be signed and that trusted status depends on trust lists.
- CONTEXT C2PA Specifications. Coalition for Content Provenance and Authenticity. The technical specification for Content Credentials, version 2.2.
Ledger updated 4 Nov 2026. Open this ledger as its own page · The P(n) Board · Method
Themes: Surveillance and scoring · AI and machine minds
