In Experiment 1, we examined the strength of memory for Facebook status updates (i.e., messages posted to friends). We compared the memorability of such microblogs to that for sentences from published books (Exp. 1a) and to the memorability of human faces (Exp. 1b).
Thirty-two University of California, San Diego (UCSD) undergraduates (age: M = 21.00 years, SD = 2.14; 27 female, five male) were randomly assigned to one of two conditions, with half in the Facebook condition and half in the book condition.
Five undergraduate research assistants, who were blind to the hypothesis, gathered 200 posts written by others (from, collectively, over 3,000 friends) from their own Facebook feeds. Each post was the most current for its author, and none was connected with games, links, or photos (see the Appendix for examples).
A total of 200 sentences were also selected from books on amazon.com, using the “Last 30 days” option under the “New Releases” section, including both fiction and nonfiction. For each of these books for which the “Look Inside” feature was available, we used the “Surprise Me!” option, and within that random page, a single sentence was randomly selected. Sentences with quotations, single-word sentences, and sentences that contained more than 25 words were not selected (see the Appendix for examples).
Of the 200 stimuli in each condition, 100 were randomly selected as targets for each participant to memorize during the study phase. The other 100 were reserved as lures for the testing phase, to assess whether the participants could identify stimuli they had seen before: (a different combination of targets and lures per participant). Instructions and stimuli were displayed via E-Prime (Psychology Software Tools Inc., Sharpsburg, PA; www.pstnet.com) on a 22-in. monitor in 18-point Courier font.
The participants were informed from the outset that they were participating in a memory experiment. After a practice trial, the study phase began. During study, 100 targets were individually presented for 3,000 ms, followed by a blank-screen interstimulus interval (ISI) of 250 ms. Immediately after the study phase, participants took a self-paced recognition test that consisted of the 100 targets randomly intermixed with 100 lures. Participants indicated their confidence that each post had been previously seen (old) or not previously seen (new) on the study list, using a 20-point rating scale that is shown in Fig. 1 (Mickes, Wixted, & Wais, 2007). A keypress of “1” indicated that they were 100 % certain that the post had not appeared in the study list, and a keypress of “20” represented that they were 100 % confident that an item had appeared. Thus, a correct response for a lure would be between 1 and 10, and a correct response for a target would be between 11 and 20.
Participants in the Facebook condition responded generally with the highest levels of confidence and were highly accurate. The average d' (an unbiased measure of memory strength) in the Facebook condition (M = 2.48, SD = 0.69) was significantly greater than the average d' in the book sentence condition (M = 1.72, SD = 0.54), t(30) = 3.47, p = .002. The percent correct measure (calculated as the number of hits plus correct rejections, divided by 200 trials), while potentially biased (Macmillan & Creelman, 2005), is intuitively appealing, and it reveals exactly the same story: significantly greater memory for Facebook posts (M = 85 %, SD = 7.2) than for sentences from books (M = 76 %, SD = 6.2), t(30) = 3.72, p = .001.
To generate the receiver-operating characteristic (ROC) plot, hit and false alarm rate pairs were computed for each level of confidence. For example, in the Facebook condition, 56 % of the targets and 6 % of the lures received a confidence rating of 20 (i.e., a hit rate of 56 % and false alarm rate of 6 % for ratings of 20). Next, another hit and false alarm rate pair was obtained by computing the percentages of targets and lures that received ratings of 19 and 20. The ratings were cumulated in this manner until we had generated 19 separate hit and false alarm rate pairs, which are plotted in the ROC. The farther the operating points that make up the ROC curve are from the diagonal, or “chance” line (i.e., the closer they are to the upper left corner), the greater the discriminability between targets and lures (Macmillan & Creelman, 2005). Figure 2 displays the ROC data, and it clearly illustrates that individuals in the Facebook condition discriminated targets from lures much more easily than did those in the book sentence condition.
Before investigating deeper conceptual explanations for the advantage that Facebook posts seem to enjoy, we ruled out various more superficial possibilities. First, we tested whether the length of the post or sentence was a critical factor. The Facebook posts (M = 11.49, SD = 6.54) had significantly, albeit slightly, more words than the book sentences (M = 10.27, SD = 4.00), t(199) = 6.14, p < .001. To adjust for this, we did a median split on the numbers of words for the Facebook posts and book sentences (median = 10.00 for both), and then compared the d' scores for the shorter Facebook posts and the longer sentences. The average d' for the Facebook posts (M = 2.67, SD = 0.98), was still significantly higher than that for the book sentences (M = 1.77, SD = 0.92), t(102) = 6.86, p < .001, even with the length confound reversed, suggesting that the Facebook posts’ advantage was not due to length.Footnote 1
Another possibility is that Facebook posts may capitalize on perceptual matching of surface-level differences (e.g., Mandler, 1979), since they are littered with irregular typography. To test this, we separated posts that contained emoticons, multiple exclamation points, all letters capitalized, or multiple, repeated letters (e.g., “Del Mar Opening Day is on my birthday this year!!! :) Hellooo HATS”) from posts that did not. A total of 86 posts did not include any of these components. Memory was still significantly higher for these orthographically regular posts (d': M = 2.57, SD = 1.18) than for the book sentences (d': M = 1.95, SD = 1.07), t(284) = 4.32, p < .001.
Memory was substantially higher for Facebook posts than for book sentences. Explanations for this advantage that were based merely on surface-level differences were ruled out: The difference was not due to posts containing emoticons, unique characters, or many or few words; the advantage persisted when all such posts were removed. The posts culled from Facebook showed remarkable memorability; memory for book sentences, on the other hand, reflected more typical memory performance. For example, the memory for sentences found here seems similar to results reported by Belmore (1982) using similar recognition tests for sentences with roughly the same number of targets and lures as in our experiment.
Clearly participants recognized Facebook posts better than ordinary published sentences. To get a further sense of the memorability of such posts, we next compared them to memory for faces. A region of the brain, the fusiform face area, is dedicated to face processing (e.g., Kanwisher, McDermott, & Chun, 1997), suggesting that the brain is specially designed to process and store facial information. While many factors can influence the memorability of a set of faces, faces nonetheless can provide some calibration for the magnitude of Facebook’s memorability, measuring whether memory for Facebook posts is particularly strong or memory for sentences from books is particularly weak. In Experiment 1b, accordingly, participants completed a memory task for faces and for Facebook posts.
Sixteen UCSD undergraduates (age: M = 20.25 years, SD = 1.34; 11 female, five male) participated for course credit.
Four undergraduate assistants used their own Facebook accounts to find a new sample of 200 posts by othersFootnote 2 (see the Appendix). Using a new set of Facebook posts allowed an independent replication of their memorability. For the face memory task, 200 neutral faces (frontal view only) were selected from the Color FERET database (http://face.nist.gov/colorferet/), with 100 of each stimulus type randomly chosen to be targets and 100 to be lures (a different combination of targets and lures per participant).
The procedure was identical to that of Experiment 1a, with two changes: Faces were used instead of book sentences, and a within-subjects rather than between-subjects design was used. Task order was counterbalanced; half of the participants completed the face memory task first, and half completed the Facebook memory task first. Again, the presentation time for each target was 3,000 ms, with a 250-ms ISI. Immediately after the study phase, participants took a self-paced recognition test that consisted of 100 targets randomly intermixed with 100 lures, shown one at a time, and the participants provided old/new responses on a 20-point rating scale.
We found no order effects, so the data from the two orders were combined for analyses. Accuracy was much higher for Facebook posts than for faces (d': M = 2.51, SD = 0.37, vs. M = 0.95, SD = 0.67, respectively), t(15) = 11.70, p < .001. Figure 3 shows the ROC data, and as in Experiment 1a, participants easily discriminated targets from lures when the stimuli were Facebook posts. Facebook posts are particularly memorable as compared to faces.