If serial dependence serves to stabilize perception, then it should generate an online illusion of stability in which a single physically changing object can be misperceived as unchanging. For example, an object that smoothly changes identity could appear “frozen” and unchanging; change blindness in this case would arise from continuously biasing our representations toward the past. The key difference between the predictions from an active serial dependence and that of the passive change blindness–based explanations ( 3 5 ) is that serial dependence predicts that the online perceptual appearance of attended things can be altered and incorrect. Change blindness explanations make no predictions about an ongoing misrecognition of attended objects at each moment in time. Here, we show that the online appearance of a single physically changing object can be made to seem stable through an active mechanism of perceptual serial dependence.

While the limits of perceptual and cognitive processing certainly define the envelope of our awareness, there is an alternative but complementary explanation for why objects in the world appear stable. The hypothesis is that the visual system uses an active stabilization mechanism, a pull toward the past, which manifests as serial dependence in perceptual judgments. Serial dependence causes objects at any moment to be misperceived as being more similar to those in the recent past. This serial dependence has been reported in the appearance of things ( 10 13 ), perceptual decisions about things ( 14 16 ), and the memory for things ( 17 19 ). In all of these examples, serial dependence is found for random or unpredictable sequential images.

The most common modern explanations for the appearance of object stability revolve around some form of change blindness ( 3 5 ) or inattentional blindness ( 6 7 ), in which the capacity limits of visual short-term memory ( 8 9 ) prevent us from being aware of things that change. This class of explanation revolves around the limits of perceptual, decisional, or memory processing: Any fluctuations in the retinal images of objects do not cause the objects to appear to change identity because we simply do not notice those changes. None of these explanations of perceptual stability make any predictions about the online appearance of objects at any particular moment in time. According to these accounts, one may not recall whether an object at this moment has changed identity, but the object at this moment is not misperceived as being some other thing.

Why do objects in the world appear to be so stable despite constant changes in their retinal images? Retinal images continuously fluctuate because of many sources of internal and external noise ranging from retinal image motion, occlusions and discontinuities, lighting changes, and perspective changes, among many other sources of noise. However, the objects do not appear to jitter, fluctuate, or change identity from moment to moment. This problem—why the world seems unchanging over time—is decades, if not centuries, old ( 1 2 ).

RESULTS

In experiment 1, two separate groups of 44 and 45 participants rated on Mechanical Turk the age of a young or old static face embedded in a blue frame (13 and 25.5 years, respectively; Fig. 1A , white square and circle). A third group of 47 independent participants were presented with a movie of a face that morphed gradually, aging from young to old. These observers then rated the age of the old face embedded in a blue frame (20.2 years; Fig. 1A , blue circle). For simplicity, we will refer to the rating of the static face alone as the “reference face” (white data points) and to ratings of the static face preceded by the movie as the “test face” (blue circle). Participants were made aware of the age rating task only after the task-irrelevant movie, when the static face and blue frame appeared; they were not warned, cued, or asked to make any explicit decision about the movie while watching it. In addition, each participant completed one, and only one, single trial across all our experiments. Thus, every experiment consisted of one-shot independent trials, from independent subjects who had no prior knowledge or experience with the stimulus or task.

(A) Experiment 1. Two groups of observers were asked to rate the age of a young or old static face embedded in a blue frame (white square and circle; reference faces). A third group was presented with a face morphing movie gradually aging from young to old and was then asked to rate the age of the old face embedded in a blue frame (blue circle, test face). Although the test face and reference faces are identical, the old test face was rated as much younger than what it actually was (green bracket). White text is shown for illustration. (B) A fourth group was presented with a face morphing movie gradually rejuvenating from old to young and was then asked to rate the age of the young face embedded in a blue frame (blue square; test face). The young test face was rated as much older than what it actually was (green bracket). ***P < 0.0001. (C) Experiment 2. Attraction percentage was computed as age difference between reference faces (e.g., reference face: old) and test faces (e.g., movie: young to old; test face: old) divided by the total age range (e.g., old reference face − young reference face). Increasing (A) and decreasing (B) age directions were equally balanced. White and black circles indicate zero (0%) and full (100%) attraction toward the beginning of the movie. When both movie and test face were presented without noise, test face age ratings were attracted toward 28% of the movie. When both movie and test face were presented with high constant dynamic noise, attraction was around 42% (A and B). When the movie was presented with high dynamic incremental noise and the test face with high noise, attraction was around 48%. (D) Experiment 3. When the age in the face morphing movie increased in gradual steps of 6, 4, and 2, attraction gradually decreased. Incremental noise and high noise were added to the movie and static face (reference or test faces), respectively. Error bars are bootstrapped 95% confidence intervals. Photo credit: Anthony Cerniello. Computer-generated face images were slightly modified for visualization purposes.

P < 0.001; We compared the age ratings between physically identical static faces, either alone (reference face) or with a preceding video (test face). The last frame of the video was identical to the test/reference face. Although the two faces (test and reference) were identical, the old test face, seen after the video, was rated as 5 years younger than the old reference face, seen without the video (20.2 versus 25.5 years;< 0.001; Fig. 1A , green bracket). We propose that, because of serial dependence, the identity of the face is continuously merged over time through the movie, and hence, observers perceive a slower age change (movie S1). As a result, the static test face is misperceived as biased toward the content of the movie seen 12 to 15 seconds ago. This is an online illusion of stability, and it shows that a single physically changing object can be misperceived as unchanging. This serial dependence effect occurs on a perceptual level, as observers were asked to judge the age of a static test face without any prior instructions, no prior decision about the movie, and no memory load.

P < 0.001; To test whether the stability illusion is due to a simple unidirectional bias in age ratings, a fourth group of 45 new participants watched a movie of a rejuvenating face that gradually morphed from old to young. Following the movie, observers rated the age of a young static test face embedded in a blue frame ( Fig. 1B , blue square). The young face was rated as 5 years older than its actual age (18.4 versus 13 years;< 0.001; Fig. 1B , green bracket). This confirms that the stability illusion can cause faces to appear younger or older depending on the previously seen faces.

A) Experiment 4. Experimental design was identical to B to D) Experiment 5. (B and C) Thirteen groups of observers were presented with a movie with a face gradually aging from young to old (black line) and, after an I.S.I. of 1 s (gray line), were asked to rate the age of the static face (test faces; blue lines). The test face was randomly chosen by random sampling the video. As a control, the other 13 observers’ groups were asked to rate the age of a static face with the same ages (reference faces; dashed lines). (D) Standardization of previous graph in terms of the distance of test face rating from the reference baseline (white dots, dashed line). For each test face age, age error in test faces was computed as the difference between reference and test face (red squares). Negative and positive values indicate that the static face was rated as younger and older than what it actually was, respectively. As in P < 0.0001. Photo credit: Anthony Cerniello. Computer-generated face images were slightly modified for visualization purposes. ) Experiment 4. Experimental design was identical to Fig. 1 (A and B) , except for an I.S.I. of 0, 1, 5, 10, or 15 s between the movie and static face. Incremental noise and high noise were added to the movie and static face (reference or test faces), respectively. Attraction percentage was computed with equally balanced increasing and decreasing age directions in the face morphing movie ( Fig. 1, A and B ). Static face age rating was attracted toward the movie at all tested I.S.I. (to) Experiment 5. (B and C) Thirteen groups of observers were presented with a movie with a face gradually aging from young to old (black line) and, after an I.S.I. of 1 s (gray line), were asked to rate the age of the static face (test faces; blue lines). The test face was randomly chosen by random sampling the video. As a control, the other 13 observers’ groups were asked to rate the age of a static face with the same ages (reference faces; dashed lines). (D) Standardization of previous graph in terms of the distance of test face rating from the reference baseline (white dots, dashed line). For each test face age, age error in test faces was computed as the difference between reference and test face (red squares). Negative and positive values indicate that the static face was rated as younger and older than what it actually was, respectively. As in Fig. 1A , when the test face was old, it was rated as much younger than what it actually was (white circles in the center). When the test face was even older (white circles on the right side), attraction gradually decreased. When the static face was younger (white circles on the left side), attraction gradually decreased and flipped to the opposite direction, i.e., a young test face was rated as older than what is actually was. Error bars are bootstrapped 95% confidence intervals. ***< 0.0001. Photo credit: Anthony Cerniello. Computer-generated face images were slightly modified for visualization purposes.

(A) Two groups of observers were asked to rate the gender of a male or female static reference face embedded in a blue frame (white square and circle; reference faces). A third group was presented with a face morphing movie gradually changing gender from male to female and was then asked to rate the gender of the test female face embedded in a blue frame (blue circle; test face). The female face was rated as much more masculine than what it actually was (green bracket). (B) A fourth group was presented with a face morphing movie gradually changing gender from female to male and was then asked to rate the gender of the male face embedded in a blue frame (blue square; test face). The male face was rated as much more feminine than what it actually was (green bracket). Different static faces ratings between (A) and (B) are due to incremental noise [(A) low-high; (B) high-low]. (C) Test face gender ratings were attracted 52% of the way toward the starting point of the movie. Photo credit: Mauro Manassi.

Fig. 4 . Experiment 7. Three groups of observers were asked to rate the age of a young, middle, or old face embedded in a blue frame (white circle, red diamond, and white square; reference faces). A fourth group was presented a movie with a face gradually aging from young to middle age and were then asked to rate the age of the test middle face embedded in a blue frame (light blue diamond; test face). The middle test face was rated as younger than it actually was. A fifth group was presented a movie with a face gradually rejuvenating from old to middle age and were then asked to rate the age of the test middle face embedded in a blue frame (dark blue diamond; test face). The middle test face was rated as older than it actually was. ***P < 0.0001. Photo credit: Anthony Cerniello. Computer-generated face images were slightly modified for visualization purposes.

It might be argued that our results are due to a central tendency bias, i.e., the tendency to rate test faces as being close to middle age, independent of movie content. To test this hypothesis, experiment 3 replicated the same conditions as in Fig. 1 (A and B) , but instead of a linear increase or decrease in the age of the face, the age of the face was morphed in different staircase functions ( Fig. 1D ). This left intact the starting and ending points of the movies (young and old). Attraction gradually decreased with decreasing the number of age steps in the movie, thus showing that our illusion is not only due to a simple response or central tendency bias but also it strongly depends on the whole content of the face morphing movie ( Fig. 1D ). As a further proof that integration occurs across the entire movie range, we computed the attraction with the last 6, 18, and 30 seconds of the video preceding the test face. Attraction linearly increased with increasing video duration, thus showing that the attraction effect involves all parts of the preceding video (experiment 11; fig. S3A).

P < 0.001 in all conditions; If our illusion is due to an active mechanism of perceptual serial dependence, then it should occur on a broad temporal range in accordance with previous literature ( 10 12 ). Accordingly, in experiment 4, we measured the temporal strength of our illusion with an interstimulus interval (I.S.I.) of 0, 1, 5, 10, and 15 seconds between the movie and test face (movie S4). Test face age ratings were attracted toward the movie at all intervals, thus showing that our illusion extends across a large period of time (< 0.001 in all conditions; Fig. 2A ). These results further show that, without intervening trials, serial dependence magnitude extends over a larger period of time than previously shown ( 10 ).

If consistent with previous serial dependence literature on face stimuli, our illusion should be determined by face feature similarity [face identity: ( 21 22 ); face expression: ( 23 24 )], and hence, it should occur only when the face morphing movie and test face are similar. Crucially, and unlike previous passive change blindness–based explanations ( 3 5 ), any modulation of the illusion by feature similarity would be consistent with serial dependence and would allow us to make predictions about the apparent age of the test face.

P < 0.01; 18, 19, In experiment 5, we presented a movie of a face that morphed from young to old (as in Fig. 1A ), and after an interval of 1 second, we varied the age of the static test face by making it younger or older than the original test old face ( Fig. 2, B and C , and movie S5). On the basis of the known tuning of serial dependence for face similarity ( 23 ), we formulated three predictions. First, our illusion should occur only with faces similar in age (and thus identity) to the test face and not between dissimilar faces. We found that the old test face was rated as younger (attraction effect) only for a few similar identities that were most similar to the old face; the attraction disappeared for more dissimilar identities (< 0.01; Fig. 2D , red squares with asterisks). Second, as the “old” test face was perceived as being ~20 years old after watching the movie ( Fig. 1A ), we predict that, when a reference face that is 20 years old is used as a test face after the movie, the degree of attraction for that face should be zero. We found no attraction for a test face of 20 years of age, meaning that the 20-year-old face is the age actually perceived at the end of the movie ( Fig. 2D , purple dot). Third, test faces younger than ~20 years old should be perceived as older, because the movie content contains older identities across the duration of the morph movie and, hence, should bias test face perception toward older ages. When the test face was younger, it was rated as older than it actually was ( Fig. 2D , white circles on the left side). Our results and predictions were very well captured by a two-parameter derivative of Gaussian model, in accordance with previous results ( 10 25 ), and ideal observer models proposed in the serial dependence literature ( 20 ).

Our illusion is not restricted to the specific identity we tested. We replicated the effect with a different set of identities (see fig. S1). As a further confirmation of the featural tuning of our illusion, we found that attraction toward the past was higher when the movie and static test face had the same identity compared to when they had different ones (see fig. S1D).

~50 participants (see Materials and Method for more details) rated on Mechanical Turk the gender (male/female) of static reference faces with low and high dynamic noise, embedded in a blue frame. Ratings were on a scale from 0 to 100% (percent female). Male faces were rated as 46 to 35% (low and high noise, respectively), and female faces were rated as 74 to 71% (low and high noise, respectively; P = 0.002; P = 0.07; P < 0.001; It might be argued that the illusion of stability here is restricted to judgments of age. Conversely, if serial dependence induces apparent stability more generally, then the effect should hold true across different stimuli and tasks. To address this, in experiment 6, we tested whether our stability effect generalizes to a new set of stimuli and a gender rating task ( Fig. 3 and movie S6). Four separate groups of50 participants (see Materials and Method for more details) rated on Mechanical Turk the gender (male/female) of static reference faces with low and high dynamic noise, embedded in a blue frame. Ratings were on a scale from 0 to 100% (percent female). Male faces were rated as 46 to 35% (low and high noise, respectively), and female faces were rated as 74 to 71% (low and high noise, respectively; Fig. 3, A and B , white squares and circles; reference faces). A fifth group of 53 participants viewed a face morphing movie that gradually changed gender from male to female. After this, they were asked to rate the gender of the female test face embedded in a blue frame ( Fig. 3A , blue circle; test face). A sixth group of 50 participants watched a face morphing movie that gradually changed gender from female to male and then rated the gender of the male test face embedded in a blue frame ( Fig. 3B , blue square; test face). The static test face was pulled toward the previously viewed movie and rated as more masculine (55.4% versus 71%;= 0.002; Fig. 4A , green brackets) or feminine (56% versus 46%;= 0.07; Fig. 4B ; green brackets) than it actually was. Collapsed across both conditions, static test face ratings were biased more than halfway toward the starting point of the movie (< 0.001; Fig. 3C ). Hence, our effect is not limited to age judgments or particular stimuli.