Artificial intelligence may be failing thanks to human error according to a new study. That's because data AI models use to learn how to identify images is not always correct to start with.

The problem affects neural networks, which are designed to work in a similar way to the human brain, considering multiple possibilities at the same time. The idea is to get the benefits of human thought but with the speed and reliability of computers.

In principle, training these AI models is a straightforward process. Rather than humans creating a set of rules for the models to follow, they simply give them a large data set of labeled images and let them figure out their own rules for how to identify different objects. Similar processes work on other data such as audio clips and passages of text.

Ambiguous Pictures

The problem, according to the MIT research, is that these data sets often have errors in the human-created labels. They calculated that 3.4 percent of the data labels they examined were either flat-out wrongly labeled or were questionable in some way. (Source: theregister.com)

One example of the latter is an image showing a bucket filled with baseballs and labeled with only one word. Either "bucket" or "baseballs" would make sense to a human but the choice could affect the lessons the AI learns from the information.

Other cases came down to whether people took a literal approach or concentrated on what was significant about an image. For example, an image looking down at a set of circular steps prompted a divide about whether it's primary label should be "lighthouse" or "coil". (Source: labelerrors.com)

Manual Review Flawed

Ironically tech company attempts to deal with such inaccuracies may be making things worse. The Register notes that some companies used low-paid outsourced workers to review the data sets and spot errors.

The problem was that the system they used to assess the performance of these workers assumed that those who picked up a lot of apparent errors were themselves either getting things wrong or deliberate trying to sabotage the system. The workers figured this out and became much more likely to "agree" the original label was correct rather than say what they actually believed.

What's Your Opinion?

Is this a big problem for AI? Have you come across similar mislabeling? Is there a simple answer or is this an inevitable issue with multiple ways to interpret an image?