WEBVTT

1
00:00:00.150 --> 00:00:01.230
- This is "Future Tech,"

2
00:00:01.230 --> 00:00:04.230
where each week, we discuss
the good the bad, and the ugly

3
00:00:04.230 --> 00:00:07.560
of where tech is headed
in 2023 and beyond.

4
00:00:07.560 --> 00:00:09.450
We are talking about the drama

5
00:00:09.450 --> 00:00:12.030
that unfolded at CNET
after it was revealed

6
00:00:12.030 --> 00:00:15.930
that the outlet was low-key
using AI to write articles.

7
00:00:15.930 --> 00:00:17.700
We're now joined by Blake Montgomery,

8
00:00:17.700 --> 00:00:19.500
deputy editor at Gizmodo.

9
00:00:19.500 --> 00:00:21.150
And Blake, what went down this week?

10
00:00:21.150 --> 00:00:23.460
- So on Twitter, people started to notice

11
00:00:23.460 --> 00:00:26.040
that some CNET articles
sounded sort of strange

12
00:00:26.040 --> 00:00:30.120
and then they ran them through
chatGPT0 or GPT0, sorry,

13
00:00:30.120 --> 00:00:33.090
which is a detection software

14
00:00:33.090 --> 00:00:34.740
basically for AI-written articles.

15
00:00:34.740 --> 00:00:37.800
And it came back that these
were probably written by AI

16
00:00:37.800 --> 00:00:39.570
and then people started to notice

17
00:00:39.570 --> 00:00:42.810
that the bylines were the CNET money team

18
00:00:42.810 --> 00:00:45.583
or the CNET like had no names involved-

19
00:00:45.583 --> 00:00:47.190
- Very cryptic, right?
- Very cryptic.

20
00:00:47.190 --> 00:00:50.250
And then, when we asked
CNET when we were reporting,

21
00:00:50.250 --> 00:00:52.680
they said, "Yes, this was written by AI

22
00:00:52.680 --> 00:00:54.930
and edited by our human editors."

23
00:00:54.930 --> 00:00:56.700
And it turns out they've
been doing this for months.

24
00:00:56.700 --> 00:00:58.320
There are like 70 stories

25
00:00:58.320 --> 00:00:59.850
that have been written since November.

26
00:00:59.850 --> 00:01:02.700
- And since this whole
revelation came to light,

27
00:01:02.700 --> 00:01:05.700
they've had to do a lot
of corrections, right?

28
00:01:05.700 --> 00:01:07.350
- And pretty major corrections.

29
00:01:07.350 --> 00:01:09.810
There are instances where chatGPT

30
00:01:09.810 --> 00:01:11.340
or whatever AI they're using

31
00:01:11.340 --> 00:01:15.240
got math just egregiously
very obviously wrong

32
00:01:15.240 --> 00:01:16.980
in a way that you'd be surprised

33
00:01:16.980 --> 00:01:18.480
that a computer would get it wrong.

34
00:01:18.480 --> 00:01:20.160
- You know, of course, we're both editors.

35
00:01:20.160 --> 00:01:22.080
Can you imagine having to edit for an AI?

36
00:01:22.080 --> 00:01:23.730
What do you think that would be like?

37
00:01:23.730 --> 00:01:24.870
- I feel like I could.

38
00:01:24.870 --> 00:01:26.460
I think I would just have to fact check it

39
00:01:26.460 --> 00:01:27.780
a lot more intensively.

40
00:01:27.780 --> 00:01:30.870
Like I don't think the AI
doesn't have the backstop

41
00:01:30.870 --> 00:01:33.000
that is like the thing in
your brain that says wait,

42
00:01:33.000 --> 00:01:34.350
that sounds wrong

43
00:01:34.350 --> 00:01:36.690
when you look at two numbers
right next to each other.

44
00:01:36.690 --> 00:01:39.840
And CNET now, unfortunately,

45
00:01:39.840 --> 00:01:41.760
they're reviewing every single article

46
00:01:41.760 --> 00:01:43.080
that's been written with AI

47
00:01:43.080 --> 00:01:45.960
because of the major corrections
that they've had to issue.

48
00:01:45.960 --> 00:01:48.090
I think editing, it would
be like high stakes.

49
00:01:48.090 --> 00:01:49.890
You'll probably be editing a lot more

50
00:01:49.890 --> 00:01:53.280
because the AI can write a
lot faster than a human can.

51
00:01:53.280 --> 00:01:56.040
- Yeah, and I think, I mean,
I edited one of the stories

52
00:01:56.040 --> 00:01:58.440
that one of our writers wrote about this

53
00:01:58.440 --> 00:02:01.650
and I, as the human editor,
editing a human's piece

54
00:02:01.650 --> 00:02:05.610
about bad edits from an AI piece,

55
00:02:05.610 --> 00:02:09.030
I definitely did my due
diligence with that piece.

56
00:02:09.030 --> 00:02:11.910
But you know, I think for me,

57
00:02:11.910 --> 00:02:14.160
if I were to have to ever edit an AI,

58
00:02:14.160 --> 00:02:16.909
at least you wanna feel bad
for like rewriting a lot of it?

59
00:02:16.909 --> 00:02:17.742
- I guess so.

60
00:02:17.742 --> 00:02:19.740
You know that it can
rewrite the whole draft

61
00:02:19.740 --> 00:02:20.870
within three seconds if you want.

62
00:02:20.870 --> 00:02:22.230
- Be like, "Hey, we're cutting this.

63
00:02:22.230 --> 00:02:23.858
They're not gonna like
take it personally."

64
00:02:23.858 --> 00:02:25.050
(Blake laughs)

65
00:02:25.050 --> 00:02:27.090
- It's never gonna feel like you told it

66
00:02:27.090 --> 00:02:28.710
that it did a bad job.

67
00:02:28.710 --> 00:02:30.870
- I mean, what do you think is next for,

68
00:02:30.870 --> 00:02:33.960
I mean, AI-generated news articles or just

69
00:02:33.960 --> 00:02:36.030
the AI-generated content in general?

70
00:02:36.030 --> 00:02:38.160
I mean, is there, I don't know,

71
00:02:38.160 --> 00:02:41.130
like plagiarism concerns?

72
00:02:41.130 --> 00:02:42.510
I mean, if they're parsing the internet

73
00:02:42.510 --> 00:02:44.280
and taking pieces of other people's stuff?

74
00:02:44.280 --> 00:02:47.790
- So far, we haven't
seen outright plagiarism

75
00:02:47.790 --> 00:02:50.640
where an AI will just take
whole sentences or paragraphs

76
00:02:50.640 --> 00:02:52.290
from something that already exists.

77
00:02:52.290 --> 00:02:55.470
The way that AI-generated text works

78
00:02:55.470 --> 00:02:58.740
is that it guesses what
word should come next

79
00:02:58.740 --> 00:03:01.830
based on like a body of trillions of words

80
00:03:01.830 --> 00:03:04.530
and hundreds and thousands,
millions of sentences.

81
00:03:04.530 --> 00:03:07.950
So, that's why you get
sometimes AI-generated sentences

82
00:03:07.950 --> 00:03:09.060
that sounds so strange,

83
00:03:09.060 --> 00:03:11.670
like it guesses wrong of
what word should come next.

84
00:03:11.670 --> 00:03:14.700
But in that sense, like
because it's guessing,

85
00:03:14.700 --> 00:03:16.620
it is random in the way that you and I

86
00:03:16.620 --> 00:03:19.740
think of random writing,
rather than copy pasting.

87
00:03:19.740 --> 00:03:21.810
- Yeah. It's kind of like
when you and your friend

88
00:03:21.810 --> 00:03:23.190
are in a conversation,
it's like oh my gosh,

89
00:03:23.190 --> 00:03:24.870
it's like we can just finish each other's-

90
00:03:24.870 --> 00:03:26.250
- Sentences.
- Right.

91
00:03:26.250 --> 00:03:27.840
And sometimes you can,
sometimes you can't.

92
00:03:27.840 --> 00:03:29.250
- And sometimes you get it wrong.

93
00:03:29.250 --> 00:03:30.510
Sandwiches.
- Yeah and sometimes

94
00:03:30.510 --> 00:03:33.193
you issue a lot of
corrections as a result.

95
00:03:33.193 --> 00:03:35.693
(light music)