The easiest way to fool artificial intelligence Feedback is our weekly column of bizarre stories, implausible advertising claims, confusing instructions and more 10 March 2021 Josie Ford This is not an apple Another week, another artificial intelligence going decidedly off-piste. We recently mentioned the autonomous video-interviewing system that appeared to be grading candidates on the strength of the bookcase behind them (27 February). Now a paper published on the website of the company OpenAI reveals how CLIP, a neural network system that learns to recognise visual concepts through being fed verbal descriptions of them, can be spoofed simply by overlaying an image with text declaring it to be something else. Stick a sticker on an apple declaring it to be a different apple product, an iPod, and the AI says it is an iPod 99.7 per cent of the time. Plaster dollar signs on a picture of anything, from a poodle to a chainsaw to a horse chestnut, and, with a charmingly artless naivety, CLIP mostly returns the answer "piggy bank".

We recently mentioned the autonomous video-interviewing system that appeared to be grading candidates on the strength of the bookcase behind them (27 February). Now a paper published on the website of the company OpenAI reveals how CLIP, a neural network system that learns to recognise visual concepts through being fed verbal descriptions of them, can be spoofed simply by overlaying an image with text declaring it to be something else. Stick a sticker on an apple declaring it to be a different apple product, an iPod, and the AI says it is an iPod 99.7 per cent of the time. Plaster dollar signs on a picture of anything, from a poodle to a chainsaw to a horse chestnut, and, with a charmingly artless naivety, CLIP mostly returns the answer "piggy bank". This suggests an excellent way to defy privacy-violating face-recognition systems when on nefarious business: simply attach a sheet of paper about your person declaring yourself to be your favourite frenemy or privacy violating tech guru.

Artificial intelligence might not be as smart as we think. University and military researchers are studying how attackers could hack into AI systems by exploiting how these systems learn. It's known as "adversarial AI." In this encore episode, Dina Temple-Raston tells us that some of these experiments use seemingly simple techniques. For more, check out Dina's special series, I'll Be Seeing You.

Artificial intelligence might not be as smart as we think. University and military researchers are studying how attackers could hack into AI systems by exploiting how these systems learn. It's known as "adversarial AI." In this encore episode, Dina Temple-Raston tells us that some of these experiments use seemingly simple techniques. For more, check out Dina's special series, I'll Be Seeing You.

Last week, at the International Conference on Machine Learning (ICML) in Stockholm, a group of researchers described a turtle they had 3D printed. Most people would say it looks just like a turtle, but an artificial intelligence (AI) algorithm that can normally recognize turtles saw it differently. Most of the time, it thought the turtle was a rifle. Similarly, it saw a 3D-printed baseball as an espresso. These are examples of "adversarial attacks"--subtly altered images, objects, or sounds that fool AIs without setting off human alarm bells.