The Gilliam Autism Rating Scale (GARS) was developed as a relatively easy, inexpensive aid in the surveillance and diagnosis of autism. This study examined the validity of the GARS when used with a sample of 119 children with strict DSM-IV diagnoses of autism, ascertained from both clinical and research settings. The GARS consistently underestimated the likelihood that autistic children in this sample would be classified as having autism. The sample mean for the Autism Quotient, a hypothesized index of the likelihood of having autism, was 90.10, significantly below the reference mean of 100. Diagnostic classification according to criteria specified by the GARS resulted in a sensitivity of only .48. Limitations of rating scales in general and of the GARS specifically are discussed. It is recommended that clinicians and researchers using or considering using the GARS for autism diagnosis or ratings of autism severity recognize the need for further research regarding its use.