Deep neural nets, nearly 10 years after their celebrated rebirth, are still utterly fooled by images like the one shown.

Judged as of 2021-11-15. True: deep neural nets had their widely recognised revival about 10 years earlier (around 2011–2012, e.g. AlexNet), and evidence from 2021 shows that leading deep image classifiers were still routinely fooled by adversarial or unusual images of the kind shown. Mostly true: the core claim holds with a small slip, such as a revival date one or two years off from 'nearly 10 years'. Mixed: the models were still fooled by some types of images but by 2021 resisted the specific type shown fairly well. Mostly false: the vulnerability was real only for narrow, older or undefended models, and the claim presents it as general. False: evidence shows that 2021 state-of-the-art vision models were robust to such images.

Set before the research began.

1. E1 Natural Adversarial Examples ar5iv (arXiv paper 1907.07174, Hendrycks et al.)

   “On ImageNet-A a DenseNet-121 obtains around 2% accuracy, an accuracy drop of approximately 90%, and its out-of-distribution detection performance on ImageNet-O is near random chance levels.”

2. E2 Rethinking Natural Adversarial Examples for Classification Models ar5iv (arXiv paper 2102.11731, Li et al., Tsinghua/Huawei)

   “A detection model based on the classification model EfficientNet-B7 achieved a top-1 accuracy of 53.95%, surpassing previous state-of-the-art classification models trained on ImageNet”

3. E3 Rethinking Natural Adversarial Examples for Classification Models ar5iv (arXiv paper 2102.11731, Li et al., Tsinghua/Huawei)

   “Figure 9 shows some examples that EfficientNet-B7 with AdvProp [9], one of the best DNN models nowadays, still failed to recognize.”

4. E4 How Did AlexNet Transform AI? Explore the Groundbreaking Source Code IEEE Spectrum (republishing a Computer History Museum blog post by Hansen Hsu) · 21 March 2025

   “In partnership with Google, the Computer History Museum (CHM) has released the source code to AlexNet, the neural network that in 2012 kickstarted today’s prevailing approach to AI.”

E4 dates the deep-learning revival to AlexNet in 2012, so "nearly 10 years" by November 2021 fits. E1 reports that an ImageNet classifier fell to about 2% accuracy on unmodified natural adversarial images. E3, a 2021 paper, says EfficientNet-B7 with AdvProp, described as one of the best models of the time, still failed to recognise such examples. E2 gives 53.95% top-1 accuracy for a strong model, but the quote does not name the benchmark, so it shows only indirectly how often leading models still failed. The evidence does not show the specific image Marcus posted, so we can't confirm that models were fooled by exactly that type of image, which keeps the verdict at mostly true.