WEBVTT

00:00:00.000 --> 00:00:02.090
[Soft instrumental music begins and continues under narration.]

00:00:02.190 --> 00:00:10.390
I spent an afternoon pulling apart an eight minute BBC film frame by frame, and the evidence changed the story I thought I was writing.

00:00:11.310 --> 00:00:19.270
The BBC World Service film explains AI 2027, a speculative paper about an AI race that can end with humanity gone.

00:00:19.870 --> 00:00:24.870
The obvious criticism would be that a major newsroom used generated video and hoped nobody noticed.

00:00:27.020 --> 00:00:28.660
That criticism is wrong.

00:00:31.020 --> 00:00:33.760
The BBC disclosed the generated footage three ways.

00:00:34.320 --> 00:00:37.280
The presenter explains the experiment at thirty seconds.

00:00:38.000 --> 00:00:40.660
The YouTube description names the use of generative AI.

00:00:41.760 --> 00:00:51.620
Individual scenario shots carry captions naming Sora, Runway, Kling, Hailuo, and Veo.

00:00:52.860 --> 00:00:56.000
Most of the scenario sequence also carries a future year stamp.

00:00:56.780 --> 00:00:58.400
The policy question is less simple.

00:00:58.900 --> 00:01:03.980
One sentence in the BBC guidance contains an exception when AI is the subject and its use is illustrative.

00:01:04.740 --> 00:01:11.620
The surrounding guidance also says AI media must not materially mislead, and it requires an editorial referral.

00:01:13.540 --> 00:01:15.200
I cannot see that internal process.

00:01:16.040 --> 00:01:19.060
The public evidence does not establish full compliance or a breach.

00:01:19.760 --> 00:01:21.920
What it does establish is a timing problem.

00:01:22.720 --> 00:01:28.400
The BBC tool caption reaches roughly five minutes and twenty two seconds, then does not return.

00:01:28.400 --> 00:01:35.080
After that point, about thirty five seconds of generated footage used again appears between real interviews.

00:01:36.000 --> 00:01:42.200
The label was reliable for five minutes and absent where synthetic shots and filmed interviews sit closest together.

00:01:42.660 --> 00:01:45.640
The protest sequence contains the clearest visual defect.

00:01:46.400 --> 00:01:49.660
A generated correspondent holds a microphone with the word NEWS.

00:01:50.800 --> 00:01:53.780
The marks above NEWS do not form a network name I can read.

00:01:54.780 --> 00:01:56.480
Nearby placards contain non-words.

00:01:57.220 --> 00:01:58.580
I can observe the artifacts.

00:01:59.580 --> 00:02:02.580
A single frame cannot prove that a person or a network does not exist.

00:02:03.500 --> 00:02:05.380
The deeper issue is epistemic status.

00:02:06.180 --> 00:02:16.140
The narration makes sixteen explicit attributions of scenario claims in one thousand and three words, an average interval of about twenty one seconds from the first to the last.

00:02:16.740 --> 00:02:22.680
The film later shows a forecast distribution, shows both endings, and calls the scenarios fictional.

00:02:23.400 --> 00:02:24.700
That qualification exists.

00:02:25.290 --> 00:02:30.340
I found no simultaneous shot-level overlay telling a viewer that the event on screen is hypothetical.

00:02:33.420 --> 00:02:35.160
That gives me three practical rules.

00:02:37.680 --> 00:02:39.420
Label the pixels and the claim separately.

00:02:40.220 --> 00:02:43.720
Keep the label on every generated shot, including footage used again.

00:02:44.160 --> 00:02:47.360
Read every sign, logo, and screen before the edit ships.

00:02:47.940 --> 00:02:55.960
The full article includes the timestamps, arithmetic, captured sources, contrary evidence, and the finding I tested and withdrew.

00:02:57.660 --> 00:02:59.340
That last part matters.

00:02:59.940 --> 00:03:05.280
A credible case study should show where the evidence changed the author, not only where it supported the headline.

00:03:05.330 --> 00:03:10.620
[Music fades out.]