Artificial Intelligence* During the past few yeare pattern recognition and cluster analysis procedures)? have been increasingly applied to the computer-aided interpretation cf spec- troscopic data of organic compounds. The user of such methods, however resolved to remain on a matiemati- cally sccure footing, is not insured against runniag afoul of spectroscopic irrelevancy or nonsense through mis- application, This is especially true if the objects to be classificd do not outnumber the features used by a considerable margin. Recently such methods were enlisted to demonstrate a relationship between mass spectra and pharmacologi- cal activity® for a group of 66 drugs. For this purpose an algorithm was developed that pigeonholes any com- pound in question as cither a tranquilizer or a sedative from the intensities at 30 selected mass-to-charge ratios. The application of this algorithm to six further drugs was purported to compound the significance of the cor- relation found. A relationship between mass spectra and pharmacological activity would seem to be an intriguing possibility, but unfortunately the results cited shove are irrelevant to it’, To show how data can be moulded to lend apparent support to an absurde hypothesis we have constructed the following example: The same sct of drugs as used in the cited article was broken down into two classes: those with names made up of an even or odd number of characters, The cor- responding mass spectra were taken from a collection compiled at the mit MS-Laboratory in coorperation with Committee VI of the American Society for Mass 4. LEDERBERG Fg, 3 197% Separatum of Chimia 27 (1973) fase. 12 Monatsschrift des Schwelzerischen Chemiker-Verbandes Sauerlinder AG, CH-5001 Aarau Printed in Switzeriand ’ Spectrometry (one of the 66 spectra [Thioperazine] was not available). A simple learning machine® using the intensities at 30 selected mass-to-charge ratios produced a decision vector capable of classifying the 65 compounds as having even or odd names with an accuracy of better than 95%. The six test compounds were all correctly assigned. The reader is welcome to interpret this as an indica- tion of a real correlation linking the parity of the name of a compound to its mass spectrum! J.T.Cierc, P. NAgcELI and J.Semu** Department of Organic Chemistry Swiss Federal Institute of Technology CH-8006 Ziirich (Switzerland) * Received November 13, 1973. *° This article was submitted in August 1973 as a letter to the Editor of Science, where the paper® appeared, which ie subject to our - comment. Since we were not able to get a definite notice of ac- eeptance for publication by Science up to now, we have with- drawn the letter and submitted it to Chimia (this journal), ~- T.L.tarnroua and (£0. Juas, Asal. Chom.43 (1971) (10) 290A, and references cited therein. B.R.Kowarskpand C.F. Benner, Anal. Chem. 44 (1972) 1405, Kat-Lr H.Tinc, R.C.T. Lee, G.W.A.Mitnge, M. SHAPIRO and A.M. Guarino, Science 180 (1973) 417. «Toute bonne théorie doit remplir deux conditions: 1. I faut qu'elle s’accorde avec l’experience. 2. Il n'est pas moins nécessaire qu'elle soit philosophiquement vraie... Un principe condamné par Je sens commun est philosophiquement faux et ne peut étre qu'une erreur scientifique.» A.-S.Couren, Ann. Chim. ot Physique, 3° serie, 53 (1858) 469. 5 N.J.Nizsson, Learning Machines, Mc Graw-Hill, Now York 1965.