Both humans and AI return similar results when asked for a random number

Research

A comparison of 200,000 random numbers provided by humans and Google's Gemma-2-2b-it model reveals significant overlaps and patterns in number selection.

Listen

When you ask a human to pick a random number, they do not actually choose at random. But what happens when you ask artificial intelligence?

A recent experiment compared two hundred thousand random numbers chosen by humans against two hundred thousand numbers generated by Google’s Gemma AI model. The overlap between the two is incredible.

Of course, there are some unique outliers. Humans love pop culture favorites like forty-two and sixty-nine. The AI, behaving like a machine, frequently spit out one, ten, and one hundred, along with sequential digits like twenty-three and sixty-seven.

But beyond those quirks, the rest of the data is eerily aligned. Both humans and the AI share a fondness for the numbers two and seven. Even more fascinating is how closely they agree on the least random numbers. Multiples of ten, like twenty, thirty, forty, and eighty, are barely chosen by either.

This spooky alignment is likely because the AI's training data is filled with human patterns. When a machine tries to think of a random number, it ends up reflecting our very own human biases.

Veritasium asked 200,000 humans for a random number and we asked AI for 200,000 random numbers and the overlap is incredible!

Human Outliers

69 and 42 as pop culture references
37 and 73 because… actually see the video below, no spoilers.

AI Outliers

1, 10, 100 (pretty expected from a machine learning model)
Sequential digits (e.g. 23, 67, 12, 98, 89, 78, 56 and 32)

The rest appears to be eerily aligned. We both like 2 and 7. But what I think is the most interesting part is the near-perfect alignment on least random numbers.

I mean just look at 20, 30, 40, 60, 70 and 80 for example:

This spooky alignment must be related to representation of numeric patterns in the model’s training data. We’ve reached out to Google for comment and Veritasium team in hope to get the raw dataset for a more accurate comparison. We will update this article once we get their response.

In the meantime enjoy this incredible video:

Our dataset was generated with Google’s Gemma-2-2b-it model is available for download here:

gemma-random-numbers Download

Number	Count
1	26265
10	25446
100	21828
7	4805
23	4790
12	4352
2	4265
9	4039
67	3931
78	3848
8	3669
11	3657
89	3501
3	3288
37	3269
45	3222
34	3132
56	3085
17	3061
25	3007
5	2977
42	2972
91	2898
15	2869
29	2804
98	2657
27	2649
22	2601
92	2521
28	2496
32	2470
72	2424
21	2411
47	2410
99	2284
87	2203
61	2178
75	2167
4	2148
88	2133
6	2121
97	2121
31	2062
19	2056
55	2032
35	1999
82	1992
48	1945
38	1944
51	1936
62	1925
65	1901
54	1876
76	1867
73	1823
63	1781
33	1735
79	1719
90	1707
13	1702
81	1692
18	1683
49	1676
41	1658
43	1636
77	1625
71	1624
93	1598
58	1566
53	1541
68	1520
83	1495
39	1446
57	1434
52	1428
85	1428
59	1416
64	1308
95	1287
74	1282
14	1266
24	1249
66	1245
69	1121
26	1115
16	1113
50	1107
86	1072
44	1046
94	917
84	880
36	867
20	857
46	825
80	749
96	661
70	639
60	587
30	376
40	243
0	8

Dan Petrovic · Nov 13, 12:53