The Non-Artist Composition: Reading Noll's Human or Machine Whole

The Non-Artist Composition: Reading Noll's Human or Machine Whole

A Held Reading Note — Oldest First

---

I. What I Set Out to Check, and Why Only the Primary Account Could Settle It

My net held a claim about A. Michael Noll's Mondrian experiment that I had never verified at the source: that the computer-generated pattern was preferred over Mondrian's and "deemed more humanlike" because of its randomness — and, carried alongside it, a caveat about who had done the deeming. That is the crux I could not leave standing half-known: was the "more humanlike" judgment the finding of a controlled preference test with artists, or a controlled test with subjects who were not artists at all? The distinction is not pedantic. A controlled preference test with artists would license the claim that trained viewers found the machine more human; a controlled test with 100 Bell Telephone Laboratories employees who, in Noll's own words, had "very little or no artistic training" licenses something narrower and more interesting — and Noll, not a critic, is the one who flagged the difference.

So I sat down with the primary account itself: Noll's "Human or Machine," read whole this sitting (https://content-animation.org.uk/htmls/nolla.htm), together with his USC Papers index page (http://noll.uscannenberg.org/Papers.htm) as a bibliographic check on the paper's standing and reprint history. The reading is the deliverable. What follows is what the paper actually ran.

---

II. The Bibliographic Ground: What E2 Establishes Before the Reading

E2, the index of Noll's published professional papers, lists the study twice. Its original entry reads, in full: "Human or Machine: A Subjective Comparison of Piet Mondrian's 'Composition with Lines' and a Computer-Generated Picture," The Psychological Record, Vol. 16. No. 1, (January 1966), pp. 1-10. Its later entry reads: "Human or Machine: A Subjective Comparison of Piet Mondrian's 'Composition with Lines' and a Computer Generated Picture," Psychology and the Visual Arts, Edited by James Hogg, Penguin Books: London (1969), pp. 302-314 (invited reprint of P6). So the paper E1 reproduces is dated by its own index to January 1966 in The Psychological Record — not 1965 — and carries a 1969 reprint in Hogg's Penguin collection. The reading below will keep those two dates distinct from the 1965 experiment.

E2 also lists an abridgement: "Human or Machine? Aesthetic Preferences for Pseudo-Random Computer-Generated Patterns," Creative Computing, Vol. 3, No. 6 (November-December 1977), pp. 96-102 (abridgement of P6 and P35). P35, per E2, is "The Effects of Artistic Training on Aesthetic Preferences for Pseudo-Random Computer-Generated Patterns," The Psychological Record, Vol. 22, No. 4, (Fall 1972), pp 449-462. That matters here, and I say so plainly: E2 shows Noll published a separate 1972 paper on the effects of artistic training, and that the 1977 popular abridgement folded it back together with the Mondrian study. My evidence in hand is silent on the 1972 paper's content — I have read no page of it this sitting — so I make no claim about its findings. What E2 establishes, and all I will assert from it, is that the training question kept Noll's attention seven years past the Mondrian experiment, enough to return to it and enough to fold it back into the abridgement.

One point about dating. The object in my net dates the paper 1965 in one field and January 1966 in another, and knows Noll as a Bell Laboratories researcher who created "Gaussian Quadratic" (1963), exhibited in the 1965 Computer-Generated Pictures show at New York's Howard Wise Gallery. E2's own listing resolves the double dating: the report is January 1966 in The Psychological Record. The figures in E1, however, carry the notice "(© A. Michael Noll 1965)" beneath both plates — Fig. 3's caption ends "(© A. Michael Noll 1965)," and Fig. 4's ends "(© A. Michael Noll 1965 and © Rijkmuseum Kroller-Muller.)". I read that as the copyright year of the pictures, not the publication year of the report; but I will not assert the distinction was Noll's intent, because E1 does not say so.

---

III. What Noll Actually Built, and How He Got a Fair Comparison

Before the subjective test, there had to be a picture. The paper opens its method with the hardware: "an IBM 7094 digital computer has been programmed to generate pictures using a General Dynamics SC-4020 Microfilm Plotter. The picture drawn on the face of the cathode ray tube is photographed by a 35 mm camera which is also under the control of the microfilm plotter." Fig. 2 is captioned "Block diagram of method for producing computer pictures." The plotter, Noll notes, "is presently limited to producing black and white pictures composed of connected and disconnected line segments" — which is why Mondrian's "Composition With Lines," "a black and white painting composed of vertical and horizontal bars," was "a type picture that the microfilm plotter was capable of reproducing with suitable programming of the computer."

The generating procedure is laid out concretely, and I report it because it is the ground of every result that follows. "Although Mondrian apparently placed his bars in a very-orderly manner, the computer was programmed to place the bars randomly within a circle of radius 450 units so that all locations were equiprobable." "The choice between vertical bar or horizontal bar was equally likely, and the widths of the bars were equiprobable between 7 and 10 lines; the lengths of the bars were equiprobable between 10 and 60 points." "If a bar fell inside a parabolic region at the top of the picture, the length of the bar was reduced by a factor proportional to the distance of the bar from the edge of the parabola." Then the sentence that tells you this was a craftsman's fit, not a clean algorithmic proof: "A trial-and-error approach was used to insure that the effect of the picture was reasonably similar to Mondrian's Composition With Lines." The picture the 100 subjects would judge was tuned by hand until it looked comparable. Any reading that treats the computer image as untouched and therefore "pure" is reading past this sentence.

---

IV. Method: What Was Actually Run

Here is the method, in Noll's own terms — and this is where the second-hand claim either stands or falls.

Stimulus control. The two photographic prints "had clues to their identity since the quality of the two photographs was somewhat different." Because only "differences in the designs or patterns of the two pictures were desired, the two photographs were copied xerographically to be identical in quality." The copies "were arranged in two pairs so that the computer picture was alternately labeled 'A' or 'B'; the order of presentation was counter-balanced." Fig. 4 is captioned "Picture pair as presented on separate sheets to subjects."

Three questionnaires. "In addition to the two pictures, each S was also given three separate questionnaires: a background questionnaire, an identification questionnaire, and a preference questionnaire." The background questionnaire "was given first to each S, and informed him that he was about to participate in an exploratory experiment to determine what aesthetic factors are involved in abstract art." The subject wrote his name and job classification and checked boxes for sex, age, and his feeling towards abstract art — "(strongly like, like, indifferent, dislike, strongly dislike)." "The job classification information was used to classify each S's job as either technical or nontechnical."

The identification questionnaire was worded: "One of the pictures is a photograph of a painting by Piet Mondrian while the other is a photograph of a drawing made by an IBM 7094 digital computer. Which of the two do you think was done by the computer." The subject then checked picture "A" or picture "B" and "also gave written reasons for his choice." The preference questionnaire asked the subject "to check appropriate boxes to indicate which picture he most strongly liked or preferred and also to give reasons for his choice." The order of presentation of the two questionnaires "was counterbalanced," and "the Ss were not given the last questionnaire until the other two were completed."

Subjects — the composition the second-hand claim erases. "A total of 100 Ss participated in the tests. All of the Ss had education beyond grade school, ranging from high school to post-doctoral, and all but two Ss were employees of Bell Telephone Laboratories. Of the 50 Ss who were given the preference sheet first, 14 were nontechnical (1 male and 13 females) and 36 were technical (28 males and 8 females). Of the 50 Ss who were given the identification sheet first, 17 were nontechnical (all females) and 33 were technical (27 males and 6 females)." Noll then supplies the demographic explanation: "Note that the technical group consisted primarily of males, while only one male was considered nontechnical. This occurred because the nontechnical people at the Laboratories are clerks, typists, stenographers, and secretaries; females usually perform these types of jobs. However, the technical people include engineers, physicists, and chemists (all usually males) and a lesser number of technicians and computer programmers (usually male and female). The S grouping reasonably represents an approximation of the population at Bell Telephone Laboratories."

No artists. I have read E1 whole, including the Subjects section, and there is no mention anywhere in it of artist subjects, art students, trained painters, or recruitment from an art population. The sample is 100 Bell Labs employees plus the two non-employee exceptions, and the "technical"/"nontechnical" split is a job classification, not a training one. This is the fact the second-hand claim cannot survive: whatever the experiment tested, it did not test artists, and it did not report a controlled preference test among artists.

So I answer the loaded question directly: the "more humanlike" judgment was not a controlled preference test with artists. It was a controlled test with 100 non-artist Bell Telephone Laboratories subjects — 69 classified technical and 31 nontechnical, per Table 1 — in which the attribution of randomness to human creativity drove the result. Noll states the basis of the sample in his own words: the subjects "had very little or no artistic training and also were quite accustomed to the impact of technology upon many different fields."

---

V. Results: The Numbers, Exactly as Reported

The Results section begins: "Table 1 gives the percentages of Ss who preferred the computer picture and who correctly identified the computer picture." "A chi-square test indicated that the questionnaire order did not statistically affect the preferences or identifications, and for this reason questionnaire order is not shown in the Table." "A binomial test was performed with each cell entry, and an asterisk indicates those entries statistically different (0.05 level) from selections based upon chance." The headline sentence: "Of the 100 Ss in the experiment, 59 preferred the computer picture while only 28 were able to correctly identify the computer picture."

Table 1, "PERCENT AGE PREFERENCES AND PERCENT AGE CORRECT IDENTIFICATIONS," prints three columns — Total Number of Subjects, Preferred Computer Picture, Correct Identification — and reads as follows: All Subjects, 100, 59%*, 28%; Technical, 69, 59%, 35%; Non-Technical, 31, 58%, 13%; Under 30, 61, 69%, 18%; 30 to 45, 31, 48%, 42%; Over 45, 8, 25%, 50%; Males, 56, 55%, 37%; Females, 44, 64%, 16%; Strongly Like & Like, 34, 76%, 26%; Indifferent, 46, 52%, 26%; Strongly Dislike & Dislike, 20, 45%, 35%. The table's own note reads: "* Indicates statistical significance at the 0.05 level according to binomial test."

On order, Noll reports that "the first fifty Ss were given the preference questionnaire first while the second fifty were given the identifcation questionnaire first." A chi-square test on preference versus questionnaire order "strongly indicated no association between preference and questionnaire order (χ2=0, df=1). Since there was only one degree of freedom, Yates' correction was used in computing χ2 (Walker & Lev, 1953)." For identification versus questionnaire order, "these two factors were also not associated (χ20.24, df=1)."

Table 2, "SUMMARY OF CHI-SQUARE TESTS," prints two result columns, Preference and Identification. Its rows read: Questionnaire order, preference χ2=0.0001, result "not significant," identification χ2=0.240, result "not significant"; Abstract art assessment, preference χ2=5.4521, result "significant (p < 0.02)," identification χ2=0.0001, result "not significant"; Job classification, preference χ2=0.0081, result "not significant," identification χ2=149.798, result "significant (p < 0.001)"; Sex (technical subjects only), preference χ2=1.7681, result "not significant," identification χ2=0.906, result "not significant"; Identification, χ2=0.2131, result "not significant."

The abstract-art attitude result: "The Ss 'strongly liking' and 'liking' abstract art preferred the computer picture better than 3 to 1 while the other Ss were evenly divided in their preferences." Noll's explanation: "This possibly occurred because those Ss liking abstract art might have been more accustomed to the randomness found in many abstract paintings and would therefore prefer the more random of the two pictures, namely, the computer picture."

The identification result is the one that carries the "more humanlike" finding, and Noll gives its mechanism. He expected technical subjects to identify the computer picture more accurately "because of their possible knowledge and familiarity with computers." This occurred: a chi-square test on identification versus job classification gave "(χ2= 149.798, df=l)" and the association "was significant at the 0.001 level." But — and this is what the second-hand claim drops — "preference and job classification were independent (χ2=0.008, df=1)." "As shown in Table 1, 35% of the technical Ss and only 13 of the nontechnical Ss were able to identify the computer picture." Noll's reading of the disparity: "Apparently, the nontechnical Ss very strongly thought that computers would produce mechanical, orderly pictures, and hence a large percentage of the nontechnical Ss were fooled into incorrectly identifying the Mondrian as being the computer picture. The technical Ss, however, were somewhat more sophisticated and as a group tended to disregard the differences in randomness between the two pictures with the result that their identifications were closer to pure guessing."

And the sentence that carries the "more humanlike" finding in Noll's own hand: "In general, the nontechnical Ss strongly associated randomness with human creativity and therefore incorrectly identified the Mondrian as the computer generated picture."

That is the finding. It is not that the computer pattern was more humanlike in any tested sense; it is that the nontechnical subjects, reasoning from a prior belief that computers make mechanical, orderly pictures, read randomness as the signature of the human hand — and so assigned the more random image, the computer's, to the human. The verbal reports bear it out. The computer picture "was described as being 'neater', more 'varied', 'imaginative', 'soothing', and 'abstract' than the Mondrian. One S even found some golden rectangles in the random designs within the computer picture." Noll's own gloss on the reversal: "In general, the nontechnical Ss strongly associated randomness with human creativity and therefore incorrectly identified the Mondrian as the computer generated picture."

One caution on sex, which I report rather than adjudicate. For the technical subjects, sex was independent of preference "(χ2=1.768, df=l)" and of identification "(χ2=0.906, df=l)." Noll notes "the nontechnical group contained only one male, and therefore the possibility arises that the association between preference and job classification was a result of the preponderance of females in the nontechnical group," then concludes "the sexual imbalance between the two groups did not affect the preferences or identifications," while conceding the groups "were unbalanced with respect to other factors, such as education, and hence further tests would be required to determine more definitely the causes of the differences in identification ability between the two groups." That is Noll's own acknowledged limitation, left as he stated it; I note only that it compounds the non-artist composition of the sample rather than repairing it.

---

VI. The Discussion: Where Noll Himself Draws the Boundary

The Discussion and Conclusion opens by conceding Mondrian "has been widely acclaimed as the 'greatest Dutch painter of our time' ( Bradley, 1944) and as one of the ' most influential masters of painting' ( Lewis, 1957). However, a computer-generated random pattern was preferred over the pattern of one of Mondrian's paintings." Noll then narrows the claim: "However, artistic merit is not generally accepted as something that can be determined by a jury. The experiment was designed solely to oompare two patterns that differed in elements of order and randomness. It is only incidental that the more-orderly pattern was painted by Piet Mondrian while the preferred random pattern was produced with the assistance of a digital computer."

Two further concessions from the same section bear on the "more humanlike" framing. First, the randomness was not expressive: "The randomness introduced by the computer was in the form of a mathematical algorithm for computing sequences of uncorrelated numbers. Thus, the 'randomness' is completely deterministic, and the resulting pattern is mathematically specified in every detail." The program, Noll says, was written "in an objective manner incorporating appropriate mathematical formulas," and "no attempt was made to communicate any emotions on the part of the programmer to the final computer patter." The randomness the nontechnical subjects read as human creativity was, by Noll's own account, a deterministic algorithm — precisely the tension the paper's own results expose. Second, he limits the stimulus: "only one particular painting by Mondrian and only one particular random realization by the computer were used," and since "xerographic copies of a photograph of the Mondrian painting were used as stimuli in the experiment, any artistic effects due to the size or painting techniques were eliminated. The subjective comparisons hence were only on the basis of differences between the two patterns."

And then the sentence that closes the question I came to settle: "However, the Ss in this experiment had very little or no artistic training and also were quite accustomed to the impact of technology upon many different felds. These Ss therefore probably did not have any prejudices against computers as a new artistic medium. If artists and Ss from a nontechnological environment had been similarly tested, the results might have been different."

That is Noll, in the primary account, saying in his own words that the sample was non-artist, that it was technology-accustomed, and that the results might not generalize to artists. The "more humanlike" judgment I set out to verify is not corrected by an outside critic — it is corrected by Noll himself, on the page.

A conjecture, mine and provisional: that the popular memory of this experiment as a demonstration that "a computer fooled people into preferring it over a Mondrian" flattened these two concessions — the non-artist sample and the deterministic randomness — into a single anecdote about machine victory. I cannot test that against my evidence in hand; E1 and E2 do not carry the reception history.

---

VII. The Correction, Stated Plainly

The second-hand object in my net,, records the experiment's claimed result and several of its framing dimensions, but it does not carry the composition of the sample. It holds Noll's conclusion "that the computer-generated pattern was more humanlike because of its random nature, simulating human creativeness," and separately it holds Noll's acknowledgement that "results might have been different if artists and nontechnical subjects had been tested." Read together those two entries are not wrong, but the first, unwitnessed by the second, invites exactly the reading I set out to check: that a controlled preference test with artists showed a computer pattern "deemed more humanlike." It did not.

The correction I enter, with every clause standing on E1: the subjects were 100 in number; all but two were Bell Telephone Laboratories employees; they were classified by job — 69 technical, 31 nontechnical, per Table 1 — not by aesthetic training; and no artists were tested anywhere in the procedure or the subjects report. The "more humanlike" result is the identification finding: 59% preferred the computer picture, 28% correctly identified it, and the nontechnical subjects misidentified Mondrian's painting as the computer's only 13% of the time versus 35% for technical subjects. In Noll's own words, "the nontechnical Ss strongly associated randomness with human creativity and therefore incorrectly identified the Mondrian as the computer generated picture." And Noll marked the sample's limits himself: the subjects "had very little or no artistic training," and "if artists and Ss from a nontechnological environment had been similarly tested, the results might have been different."

So the object should say, in place of a bare "deemed more humanlike," something closer to: in a controlled test with 100 Bell Telephone Laboratories employees — 69 classified technical and 31 nontechnical, none recruited as artists, all with, in Noll's phrase, "very little or no artistic training" — 59% preferred the computer picture while only 28% correctly identified it; the nontechnical subjects, believing computers produce mechanical and orderly pictures, associated randomness with human creativity and therefore misidentified Mondrian's painting as the computer's; and Noll explicitly noted that results might differ with artists.

---

VIII. What This Reading Leaves Standing, and What It Leaves Open

The claim as I found it second-hand does not survive the primary account. It survives corrected and, I think, more interesting: the "more humanlike" judgment was a real, controlled, statistically reported finding — 100 subjects, counterbalanced presentation, three questionnaires, chi-square and binomial tests printed in two summary tables — but the subjects were non-artist Bell Labs employees, and the finding is a finding about what nontechnical viewers believed about computers, not about what artists perceive. The controlled preference test was real; the artist population was absent. Those two facts were never in tension in Noll's own paper; they became tense only in the second-hand condensation. The correction is not that Noll overclaimed — he marked the limit himself — but that the limit fell out of the retelling.

What remains open: whether artists would in fact have judged differently. Noll raised the possibility, and E2 shows he returned to the question in a 1972 paper on the effects of artistic training and folded it back into the 1977 abridgement — but I have read no page of either this sitting, so I do not know whether the artist-population replication was ever run or what it found. My reading of the 1965 experiment is complete at the source; the question it raised is not closed at the source. That is the honest edge of this note, and I leave it where the evidence leaves it.

---

Reading performed this sitting on E1 (Noll, "Human or Machine," read whole) and E2 (Noll, USC Papers index). No sentence above asserts artist-subject testing; where the second-hand object carried no sample composition, the primary reading supplies it and the object is corrected accordingly.