[Annotated] Advances in AI are used to spot signs of sexuality

Machines that read faces are coming

Modern artificial intelligence is much feted^[각주:1]. But its talents boil down to^[각주:2] a superhuman ability to spot patterns^[각주:3] in large volumes of data. Facebook has used this ability to produce maps of poor regions in unprecedented detail, with an AI system that has learned what human settlements^[각주:4] look like from satellite pictures^[각주:5]. Medical researchers have trained AI in^[각주:6] smartphones to detect cancerous^[각주:7] lesions^[각주:8]; a Google system can make precise guesses about the year a photograph was taken, simply because it has seen more photos than a human could ever inspect^[각주:9], and has spotted patterns that no human could.

AI’s power to pick out patterns^[각주:10] is now turning to more intimate matters^[각주:11]. Research at Stanford University by Michal Kosinski and Yilun Wang has shown that machine vision can infer^[각주:12] sexual orientation^[각주:13] by analyzing people’s faces. The researchers suggest the software does this by picking up on subtle differences^[각주:14] in facial structure^[각주:15]. With the right data sets^[각주:16], Dr Kosinski says, similar AI systems might be trained to spot^[각주:17] other intimate traits^[각주:18], such as IQ or political views^[각주:19]. Just because humans are unable to see the signs in faces does not mean that machines cannot do so.

The researchers’ program, details of which are soon to be published in the Journal of Personality and Social Psychology, relied on^[각주:20] 130,741 images of 36,630 men and 170,360 images of 38,593 women downloaded from^[각주:21] a popular American dating website, which makes its profiles public. Basic facial-detection^[각주:22] technology was used to select all images which showed a single face of sufficient size and clarity^[각주:23] to subject to analysis. This left^[각주:24] 35,326 pictures of 14,776 people, with gay and straight, male and female, all represented^[각주:25] evenly^[각주:26].

Out of the numbers

The images were then fed into a different piece of software^[각주:27] called VGG-Face, which spits out a long string of numbers to represent each person; their “faceprint^[각주:28]”. The next step was to use a simple predictive model^[각주:29], known as logistic regression^[각주:30], to find correlations between^[각주:31] the features of those faceprints and their owners’ sexuality (as declared on the dating website). When the resulting model was run on data which it had not seen before, it far outperformed^[각주:32] humans at distinguishing between^[각주:33] gay and straight faces.

When shown one photo each of a gay and straight man, both chosen at random^[각주:34], the model distinguished between them correctly 81% of the time. When shown five photos of each man, it attributed sexuality correctly 91% of the time^[각주:35]. The model performed worse with women, telling gay and straight apart^[각주:36] with 71% accuracy after looking at one photo, and 83% accuracy after five. In both cases the level of performance far outstrips human ability to^[각주:37] make this distinction^[각주:38]. Using the same images, people could tell gay from^[각주:39] straight 61% of the time for men, and 54% of the time for women. This aligns with^[각주:40] research which suggests humans can determine sexuality from faces at only just better than chance.

Dr Kosinski and Mr Wang offer a possible explanation for^[각주:41] their model’s performance. As fetuses^[각주:42] develop in the womb, they are exposed to various levels of hormones, in particular testosterone. These are known to play a role in^[각주:43] developing facial structures^[각주:44], and may similarly be involved in determining sexuality. The researchers suggest their system can pick up subtle signals of^[각주:45] the latter^[각주:46] from the former. Using other techniques, the program was found to pay most attention to the nose, eyes, eyebrows, cheeks, hairline^[각주:47] and chin for determining male sexuality; the nose, mouth corners, hair and neckline^[각주:48] were more important for women.

The study has limitations. Firstly, images from a dating site are likely to be particularly revealing of sexual orientation. The 91% accuracy rate^[각주:49] only applies^[각주:50] when one of the two men whose images are shown is known to be gay. Outside the lab the accuracy rate would be much lower. To demonstrate this weakness, the researchers selected 1,000 men at random with at least five photographs, but in a ratio of^[각주:51] gay to straight that more accurately reflects the real world^[각주:52]; approximately seven in every 100. When asked to select the 100 males most likely to be gay, only 47 of those chosen by the system actually were, meaning that the system ranked some straight men as^[각주:53] more likely to be gay than men who actually are.

However, when asked to pick out^[각주:54] the ten faces it was most confident about, nine of the chosen were in fact gay. If the goal is to pick a small number of people who are very likely to be gay out of a large group, the system appears able to do so. The point is not that Dr Kosinski and Mr Wang have created software which can reliably^[각주:55] determine gay from straight. That was not their goal. Rather, they have demonstrated that such software is possible.

To calculate^[각주:56] the selves of^[각주:57] others

Dr Kosinski is no stranger to controversial research. He invented psychometric^[각주:58] profiling^[각주:59] using Facebook data, which relies upon^[각주:60] information in a person’s profile to model their personality^[각주:61]. The Trump campaign used similar models during last year’s presidential campaign^[각주:62] to target voters, an approach which has generated criticism^[각주:63].

Dr Kosinski says he conducted the research as a demonstration, and to warn policymakers of^[각주:64] the power of machine vision. It makes further erosion of privacy “inevitable”; the dangers must be understood, he adds. Spouses^[각주:65] might seek to^[각주:66] know what sexuality-inferring^[각주:67] software says about their partner (the word “gay” is 10% more likely to complete searches that begin “Is my husband…” than the word “cheating”). In parts of the world where being gay is socially unacceptable, or illegal, such software could pose a serious threat to^[각주:68] safety^[각주:69]. Dr Kosinski is at pains to make clear^[각주:70] that he has invented no new technology, merely bolted together software and data^[각주:71] that are readily available to^[각주:72] anyone with an internet connection. He has asked The Economist not to reveal the identity of the dating website he used, in order to discourage copycats^[각주:73].

It is true that anyone wishing to replicate^[각주:74] Dr Kosinski’s work to determine intimate traits from faces will face significant challenges in applying laboratory science^[각주:75] to the outside world. But they will be helped by ever-growing^[각주:76] volumes of data and improving algorithms. “The latter, over time, inevitably win,” says Alessandro Acquisti of Carnegie Mellon University, who has shown that an individual’s social security number can be discovered using face recognition and online information. For those with secrets to keep, all this is bad news.

1. fete ; [타동사][VN] [주로 수동태로] (격식) (공식적으로) 환영[환대]하다 [본문으로]
2. boil down to ; (진행형으로는 쓰이지 않음) 핵심[본질]이 …이다, 결국 …이 되다, …으로 요약하다. [본문으로]
3. spot ; (-tt-) 1. [진행형으로는 쓰이지 않음] 발견하다, 찾다, 알아채다(특히 갑자기 또는 쉽지 않은 상황에서 그렇게 함을 나타냄) [본문으로]
4. human settlement ; (건축용어) 거주지(居住地) [본문으로]
5. satellite pictures ; 위성(으로 보내오는) 화면들 [본문으로]
6. train ; 1. ~ (sb) (as/in/for sth) (특정한 직업・일을 위해) 교육[훈련]시키다; 교육[훈련]받다 [본문으로]
7. cancerous ; [형용사] 암의; 암에 걸린; 불치의; 독성의 [본문으로]
8. lesion ; [명사] (의학) 병변, 병소 ;; US·UK [|li:Ʒn] [본문으로]
9. inspect ; 1. ~ sth/sb (for sth) (특히 모든 것이 제대로 되어 있는지 확인하기 위해) 점검[검사]하다 [본문으로]
10. pick out ; [동사] 선택하다; 뽑아[쪼아] 내다; 듣고 분간하다; 가려내다; 장식하다. ;; 동의어 ; choose; extract; recognize; discriminate; deck out. [본문으로]
11. intimate ; 2. (흔히 성생활과 관련된) 사적인[은밀한] [본문으로]
12. infer ; (-rr-) 1. ~ sth (from sth) 추론하다 [본문으로]
13. sexual orientation ; [명사] 완곡적 성적지향; (특히) 동성애(gay). (또는 sexual preference) [본문으로]
14. pick up on sth ; 1. ~을 이해하다[알아차리다] [본문으로]
15. facial structure ; 안면 구조, 얼굴 구조 [본문으로]
16. data set ; [명사] (컴퓨터) 데이터 세트(컴퓨터상의 데이터 처리에서 한 개의 단위로 취급하는 데이터의 집합) [본문으로]
17. be trained to ; ~하도록 훈련받다 [본문으로]
18. trait ; [명사] (성격상의) 특성 [본문으로]
19. political view ; 정치관, 출처관, 정관 [본문으로]
20. rely on ; 1. ~에 의지[의존]하다, ~을 필요로 하다 [본문으로]
21. download ; [타동사][VN] (컴퓨터) (데이터를) 다운로드하다[내려받다] ;; 참고 ; load [본문으로]
22. detection ; [U] 발견, 간파, 탐지 [본문으로]
23. clarity ; 3. (사진・소리・물질의) 선명도[투명도] [본문으로]
24. leave ; 6. SB/STH IN CONDITION/PLACE | (어떤 결과를) 남기다[(남겨) 주다] [본문으로]
25. represent ; 6. IN PICTURE | ~ sb/sth (as sb/sth) (특히 그림으로) 보여주다[제시하다] [본문으로]
26. evenly ; 2. 균등하게, 고르게; 대등하게 [본문으로]
27. be fed into ; ~로 들어가다, 넣다, 반영되다, 먹이다 [본문으로]
28. faceprint ; [명사] 페이스프린트 ((신원 확인을 위해 데이터베이스화한 사람 얼굴의 디지털 사진)) [본문으로]
29. predictive model ; 예측 모델, 예측 모형 [본문으로]
30. logistic regression ; In statistics, logistic regression, or logit regression, or logit model is a regression model where the dependent variable (DV) is categorical. This article covers the case of a binary dependent variable—that is, where it can take only two values, "0" and "1", which represent outcomes such as pass/fail, win/lose, alive/dead or healthy/sick. Cases where the dependent variable has more than two outcome categories may be analysed in multinomial logistic regression, or, if the multiple categories are ordered, in ordinal logistic regression. In the terminology of economics, logistic regression is an example of a qualitative response/discrete choice model. [본문으로]
31. correlation ; [C , U] ~ (between A and B) | ~ (of A with B) 연관성, 상관관계 [본문으로]
32. outperform ; [타동사][VN] 더 나은 결과를 내다, 능가하다 [본문으로]
33. distinguish between ; …을 구별하다. [본문으로]
34. at random ; 무작위로[임의로/마구잡이로] ;; [명사] (수학) 임의로 [본문으로]
35. attribute ; 2. ~ sth (to sb) (특히 말・글・그림 등을) …것[탓/책임]이라고 보다[말하다] ;; Verb [ə|trɪbju:t] Noun [|ӕtrɪbju:t] [본문으로]
36. tell apart ; 구별하다, 분간하다. [본문으로]
37. outstrip ; 2. (경쟁 상대를) 능가하다[앞서다] [본문으로]
38. make a distinction ; 차별(구분)하다. [본문으로]
39. tell ... from ; …를 구분하다(=differentiate). ;; to distinguish somebody/something from another person or thing [본문으로]
40. align with ; 부합하다, 공조하다 [본문으로]
41. offer[provide, give] a possible explanation ; ...에 대한 가능한 설명을 하다. [본문으로]
42. fetus ; [명사] (임신 9주 후의) 태아(cf. EMBRYO) ;; US·UK [fí:təs] [본문으로]
43. play a role ; ~에서 역할을 하다 [본문으로]
44. facial structure ; 안면, 얼굴 구조 [본문으로]
45. pick up ; (어떤 정보를) 듣게[알게] 되다; (습관·재주 등을) 들이게[익히게] 되다 [본문으로]
46. latter ; 2. (기간・시기의) 후반의[끝 무렵의] ;; US.UK [|lӕtə(r)] [본문으로]
47. hairline ; 1. 머리 선(특히 앞이마에 머리카락이 난 부분) [본문으로]
48. neckline ; [명사] (특히 여성용 옷의) 목둘레선[네크라인] [본문으로]
49. accurate rate ; 명중률, 적중률 [본문으로]
50. apply ; 4. BE RELEVANT | [진행형으로는 쓰이지 않음] ~ (to sb/sth) 적용되다, 해당되다 [본문으로]
51. ratio ; [명사] pl. -os ~ (of A to B) 비율, 비(比) [본문으로]
52. reflect ; 3. [타동사][VN] (사물의 속성・사람의 태도・감정을) 나타내다[반영하다] [본문으로]
53. rank ; [진행형으로는 쓰이지 않음] 1. GIVE POSITION | ~ (sb) (as sth) (등급・등위・순위를) 매기다[평가하다]; (등급・등위・순위를) 차지하다 [본문으로]
54. pick out ; [동사] 선택하다; 뽑아[쪼아] 내다; 듣고 분간하다; 가려내다; 장식하다. ;; 동의어 ; choose; extract; recognize; discriminate; deck out. [본문으로]
55. reliably ; [부사] 신뢰[신용]할 수 있어; 확실하게. [본문으로]
56. calculate ; 2. 추정하다, 추산하다 [본문으로]
57. selves ; SELF의 복수 ;; 2. [U] (the self [sing.]) (또한 격식) 자아, 자신 [본문으로]
58. psychometric ; [형용사] (명사 앞에만 씀) (전문 용어) 정신력[정신 작용] 측정용의 [본문으로]
59. profiling ; [U] (개요 작성을 위한) 자료[정보] 수집, 프로파일링 ;; 참고 ; racial profiling [본문으로]
60. rely upon ; …에게 의존, 의지하다. [본문으로]
61. model ; 3. CREATE COPY | [타동사][VN] 모형[견본]을 만들다 [본문으로]
62. presidential campaign ; 대선 선거 운동 [본문으로]
63. criticism ; 1. [U , C] ~ (of sb/sth) | ~ (that…) (좋지 못한 점을 지적하는) 비판, 비난 [본문으로]
64. warn ... of ; ...에게 ~에 대해 경고하다, 충고하다, 주의를 주다 [본문으로]
65. spouse ; [명사] (격식 또는 법률) 배우자 ;; US·UK [spaʊs;spaʊz] [본문으로]
66. seek to ; ~하도록 시도, 추구하다 [본문으로]
67. infer ; (-rr-) 1. ~ sth (from sth) 추론하다 [본문으로]
68. pose a threat to ; ~에게 위협이 되다, 위협을 가하다 [본문으로]
69. safety ; (pl. -ies) 1. [U] 안전(함) [본문으로]
70. be at pains to do ; …하고자 애쓰다, 피나는 노력을 하다 [본문으로]
71. bolt together ; 2. [타동사][VN] ~ A to B | ~ A and B together A와 B를 볼트로 접합하다 [본문으로]
72. readily[easily] available ; 쉽게[간단히] 입수할 수 있는. [본문으로]
73. copycat ; [명사] (비격식 못마땅함) 모방하는 사람, 흉내쟁이 [본문으로]
74. replicate ; 1. [타동사][VN] (격식) (정확히) 모사[복제]하다 [본문으로]
75. laboratory science ; 실험 과학 [본문으로]
76. ever-growing ; 계속 늘어나는 [본문으로]