Your social media posts are revealing more than where you’ve been this summer and who you’ve hung out with. New research out of the University of Vermont and Harvard, using Amazon’s Mechanical Turk to crowd source data, shows when a social media user is depressed it's revealed in the photos they post on Instagram.

According to the findings published in the journal EPJ Data Science Tuesday, the researchers involved used Amazon’s Mechanical Turk to recruit volunteers to provide their Instagram feeds as well as mental health records. Of the 166 volunteers chosen, who provided more than 43,000 photos total, about half of them had been clinically depressed in the previous three years, according to the UVM.

Read: How AI Could Turn Social Media Into Stock Market Gold

Those photos were then analyzed by a different set of workers recruited through Mechanical Turk. They rated 20 random photos chosen from the nearly 44,000 included in the study. Those analyzing the photos rated them on a scale of zero to five for how interesting, happy, sad and likeable each individual photo seemed to them. They were given no information about where the photo was taken from, who took it or why they were asked to rate it.

This method of rating the photos was far less reliable than a general practitioner's success rate of unassisted diagnosis, according to the study. And a general practitioner's unassisted diagnosis rate was still less reliable than the computer’s analysis of color and filters and the frequency of faces in posts for depression identification.

Read: Why Certain Women Are More Prone To Depression Than Others

On average the depressed participants posted photos that were more frequently dark, and more blue and gray than their control counterpart participants. The results of the study found that depressed participants were less likely than control participants to use a filter. But if the depressed participants did use a filter, the most common one was “Inkwell,” a filter that makes photos black and white. The control participants, who had no record of depression, used the bright filter “Valencia” most frequently.

In addition to hue and level of color in the photos, the results showed depressed participants were more likely to post photos with faces in them but on average they had fewer faces in the photos they posted.

The research has implications for the way depression could be diagnosed in the future. It would not be able to replace current diagnosis methods but would help speed up the process or help give doctors more insight to their patients than self-reported data.