A video call freezes for less than a second. The candidate’s smile stalls while the voice continues. A doctor’s mouth jumps forward to catch up with the sound. A financial adviser’s face briefly dissolves into blocks.

Everyone knows the person on screen did not choose this. No important sentence may have been lost. The glitch is a property of networks, compression and devices, not character.

And yet the person is judged for it.

A large programme of research published in Nature found that minor audiovisual faults lowered hiring interest, reduced trust in a telehealth provider and weakened willingness to work with a financial adviser. In records of real Kentucky parole hearings, glitches were also associated with a substantially lower release rate.

The researchers’ explanation is more specific than simple annoyance. A video call works partly by making another person feel present. Distort their face or break the timing between mouth and voice, and the almost-natural interaction can suddenly feel eerie. That uncanniness spills into the judgment of the human being represented by the faulty stream.

The research followed the effect across ten studies

Melanie Brucks of Columbia Business School, Jacqueline Rifkin of Cornell University and Jeff Johnson of the University of Missouri-Kansas City combined five main experiments, three supplementary studies and two archival datasets.

Some participants watched recorded job candidates, healthcare professionals or advisers. Others took part in live conversations. The team also examined naturally occurring video calls and hearing transcripts rather than relying only on laboratory simulations.

That breadth matters because each design answers a different question. Randomized videos can isolate what a glitch does to an evaluation. Real calls show whether the pattern survives outside a controlled stimulus, but they leave more room for alternative explanations.

Across the experiments, the researchers altered audiovisual delivery while holding the communicator and message constant. They used freezes, choppy movement, pixelation, audio delays, echoes, distortion and lost sound.

The team has made the experimental datasets and analysis code available through its ResearchBox project. The interview and parole transcripts remain restricted to protect identities.

The strongest claim is therefore not that every bad connection ruins every call. It is that seemingly minor transmission faults can change interpersonal judgments even when viewers know, in principle, that technology caused them.

A healthcare adviser lost trust without changing the advice

One experiment placed 497 participants in a simulated telehealth setting. They viewed the same healthcare professional giving advice, but some saw a version containing glitches.

In the clean-video condition, 77% chose to meet with the featured provider rather than other providers. In the glitch condition, 61% did so, a difference of 16 percentage points.

The professional had not become less knowledgeable, less careful or less clear in substance. The transmission changed. Participants nevertheless reported more uncanniness and were less likely to choose that provider.

A conceptual replication involving 486 people used a financial adviser. Willingness to work with him averaged 4.33 in the control condition and 3.59 after glitches, on the study’s rating scale. Uncanniness again statistically mediated part of the difference.

Mediation is not a scan of the mind and does not prove that one feeling caused every decision. It means the measured pattern was consistent with glitches increasing uncanniness, which in turn tracked lower evaluations.

The title’s reference to a patient’s trust is therefore grounded in a randomized scenario, not a review of treatment outcomes. The study measured people’s provider choice and trust-related judgment after watching advice; it did not test whether real patients ignored care or became less healthy.

The hiring penalty survived when the words arrived intact

The researchers built a taxonomy of ten common glitch subtypes and tested them in a multi-wave hiring study involving 3,092 participants. Fleeting problems lasted from one to 2.5 seconds; sustained versions lasted four seconds.

Collapsing across the ten types, glitchy candidate videos felt more uncanny and reduced hiring interest relative to an uninterrupted control. The standardized difference in hiring interest was modest but clear, with a Cohen’s d of minus 0.42.

Not all glitches worked in the same way. Losing audio reduced comprehension, so viewers had an obvious informational reason to judge the interview less favourably.

Most other glitches did not measurably reduce comprehension. The candidate’s answer still arrived, yet ratings fell. The relationship between uncanniness and evaluation remained after the analysis accounted for comprehension and both objective and perceived disruptiveness.

Some especially brief faults could be surprisingly unsettling. Duration alone did not explain the penalty. What mattered more was how sharply the stream violated the expectation that a living face and voice should move together.

This is why “tiny” should not be read as “invisible.” Participants noticed enough temporal or visual wrongness to feel it, even when they did not miss the message.

The screen had to be standing in for a person

A dropped frame is irritating in any media. To test whether the result was merely irritation, the researchers compared direct, face-to-face video with a shared video presentation.

Glitches undermined willingness to work with the communicator when the screen simulated an interpersonal encounter. The effect did not appear in the same way when viewers were focused on shared presentation content.

That contrast supports the idea of disrupted social presence. A slide deck is already an object on a screen. A human face is being treated as a person across the table. When its movement becomes almost human but not quite, the perceptual error acquires a social meaning.

The paper borrows the language of the uncanny valley, usually applied to robots or computer-generated faces. Here the unsettling object is not artificial. It is a real person passing through an imperfect digital channel.

The researchers also tested obvious repairs. Having an adviser acknowledge the glitches did not rescue the evaluation and slightly worsened interest in working with him in one experiment. A joke softened the loss a little, but every glitch condition still rated below the clean control.

That does not mean people should pretend a broken call is fine. It means an apology cannot be assumed to erase a perceptual response that has already attached itself to the interaction.

Natural conversations showed a smaller but real social shadow

The first archival test used the CANDOR corpus, a dataset of more than 1,600 get-to-know-you video conversations recorded from January through November 2020.

Participants had rated how connected they felt and reported technical difficulties. Researchers reviewed calls with reported problems to verify the glitches and classify whether they involved audio, video, one participant or both.

People in verified glitchy conversations liked their partners less and felt less shared reality. The effect was small. For liking, the standardized difference was about minus 0.24, and the pattern appeared across glitch types and regardless of whether one or both partners experienced the problem.

These were organic calls, not randomly assigned faults. A poor conversation could make a technical problem more salient in memory, or other features could influence both reported connection and call quality. Verification improves the record, but it does not turn an archive into an experiment.

Taken beside the randomized studies, however, the archive shows that the mechanism is not confined to actors in a single scripted interview.

The parole finding is the most serious and the easiest to overstate

The team analysed transcripts from 472 virtual parole hearings held in Kentucky between January and April 2021. A custom dictionary searched for language indicating audio or video disruption.

Glitches were identified in 32.6% of the hearings. Parole was granted in 48% of glitchy calls, compared with 60% of calls in which no glitch was detected. The 12-percentage-point gap remained after adjustment for recorded characteristics of the incarcerated person and the crime.

Because people inside the same institution used institutional internet, connection trouble was not simply a consumer choice or a signal that one person had purchased cheaper broadband. The person seeking parole also had little practical control over the infrastructure framing the hearing.

Still, this part was observational. The researchers did not randomly damage parole calls, and a transcript dictionary can detect only disruptions that someone verbalized. Unrecorded features of hearings may differ between the glitch and no-glitch groups.

It is therefore accurate to say glitches were associated with a lower probability of parole in this dataset. It is too strong to say the paper proved that a bad connection caused any particular person to remain incarcerated.

That boundary does not make the result unimportant. When a 12-point difference appears beside a technical condition outside the applicant’s control, a justice system should need more than perfect causal certainty before treating connection quality as harmless.

Bad bandwidth can become an invisible credential

The inequality risk sits underneath every domain in the paper. People do not enter video calls with equal broadband, devices, lighting, private rooms or freedom to reconnect.

Lower-income households and rural communities often face worse internet access. A candidate using an old phone on unstable Wi-Fi can therefore arrive with a hidden disadvantage before saying a word. The study did not directly estimate that socioeconomic penalty, but it identifies a mechanism capable of producing one.

Connection quality can begin to function like an invisible credential. Smooth video reads as effortless presence; broken video makes the same person feel less trustworthy or less desirable to work with. Viewers may sincerely believe they are judging only the content.

The practical response belongs at the institutional level. Employers can avoid final judgments after a visibly degraded interview and offer an equivalent retry or audio channel. Telehealth systems can distinguish technical quality from provider quality. Courts and parole boards can stop and resume rather than allowing a failed stream to become part of the applicant’s presentation.

Platforms can also make network faults unmistakably external, though the acknowledgement experiment warns that a simple apology banner may not be enough. Better error handling should preserve the person rather than leave their face frozen in an unnatural expression.

For an individual caught in a glitch, reconnecting may be wiser than pushing through while apologizing repeatedly. But the burden cannot rest only on the person with the weakest connection. In a high-stakes call, the evaluator holds the power to separate signal quality from human quality.

A bad stream does not reveal a bad candidate, an untrustworthy clinician or an undeserving prisoner. The research matters because people can understand that sentence perfectly and still behave as though the pixels said otherwise.