AI assistance and error communication in deepfake evidence evaluation: A systematic review of human authenticity judgments

In the modern world, a video recording is often treated as an unchangeable record of truth. If a camera captures a person speaking, the assumption is that the person was there, saying those words, at that moment. This belief forms the backbone of criminal investigations and legal judgments. However, a new technology called deepfake generation has disrupted this certainty. These systems can create videos where a person's face and voice are replaced or altered so convincingly that the event never actually happened. The problem for investigators is not just spotting these fakes; it is figuring out how to use computer tools that claim to detect them. When a human judge or police officer is shown a video alongside a computer's opinion that says "this is fake," how does that advice change their mind? Does the computer help them see the truth, or does it trick them into making mistakes they would not have made alone?

This question is the focus of a recent systematic review by researcher Hardianto Djanggih. The study did not test a new computer program or run a single experiment. Instead, it gathered and analyzed seven existing studies that looked specifically at how humans react when they receive help from artificial intelligence while judging whether a video or image is real. The researcher wanted to know if this assistance improves the accuracy of human judgment, how the way the advice is presented matters, and whether these findings can be trusted in a courtroom setting. The review found that the answer is not a simple "yes" or "no." The value of the AI depends entirely on whether the advice is correct and how it is shown to the human.

The review examined experiments where people watched videos or looked at faces and decided if they were authentic or manipulated. In some cases, the participants were given a prediction from a computer model before making their final decision. In others, they saw visual clues, such as highlighted areas where the computer thought a fake was hiding, or they were told how reliable the computer usually is. The results showed that when the computer gave correct advice, it often helped people get the right answer more often. For example, in one study involving video clips, showing the computer's prediction raised the correct identification rate from 66 percent to 73 percent. However, the study also revealed a dangerous side effect: when the computer gave the wrong advice, people tended to follow it, even when it led them to the wrong conclusion. The assistance did not just make people smarter; it also made them more likely to make specific errors if the tool itself was flawed.

How the information was presented mattered just as much as the information itself. One study compared simple text labels with visual displays that amplified the tiny, unnatural artifacts often left behind by deepfake creation. The visual displays made people more confident in their judgments and more likely to accept the computer's assessment. Yet, this confidence did not always mean they were right. In another experiment, researchers told participants that a computer was very accurate, but the computer's actual performance did not match that claim. The participants trusted the computer more when they were told it was better, even though the underlying technology was the same. This suggests that the way an error rate is communicated can change how much a person relies on the tool, regardless of whether the tool is actually performing as promised.

The review also looked at whether the source of the advice mattered. In one experiment, participants received the exact same guidance, but it was sometimes labeled as coming from an artificial intelligence and other times from a human. The accuracy of the final decisions did not change based on who the participants thought was giving the advice. This finding challenges the idea that people automatically trust machines more than humans, or that they are naturally suspicious of computers. Instead, the quality of the advice and the correctness of the underlying data were the true drivers of success or failure. The researchers noted that in many of these studies, the advice given to the participants was perfect or simulated, meaning the experiments did not fully test what happens when a real-world detector makes a mistake.

A critical limitation of this research is that none of the included studies tested how these tools work in a real criminal trial. The experiments involved ordinary people making quick judgments in a lab or online, not judges or lawyers deciding the fate of a defendant. The review found no evidence that using these tools leads to more accurate verdicts in actual legal cases. The studies also did not measure whether people could correctly identify who created a fake video or whether the person in the video actually committed a crime. The research only addressed the narrower question of whether a video looks real or fake. This gap is significant because a video can be technically authentic but still be used to tell a misleading story, or a fake video might be used for a harmless purpose.

The author concludes that the relationship between humans and AI in this field is complex. Assistance, trust, and accurate judgment are three different things. A person might trust a tool and agree with it, but still be wrong. The review suggests that for these tools to be useful in criminal evidence, the communication about them must be very precise. Experts need to explain exactly what conditions the tool was tested under and what it cannot do. They should avoid making broad claims like "this recording is fake" and instead stick to specific, testable statements like "this file contains features associated with a specific type of manipulation." The study also highlights the need for transparency. If a tool highlights a suspicious area in a video, the original video must remain available so that others can check if the highlighting tool itself is working correctly.

Ultimately, this review serves as a cautionary guide rather than a celebration of technology. It shows that while AI assistance can improve human detection of deepfakes, it also introduces new risks if the advice is incorrect or if the human becomes too dependent on the machine. The findings suggest that in a legal setting, the focus should not be on how much a tool increases confidence, but on whether it helps people distinguish between helpful advice and misleading errors. Until these tools are tested in real legal scenarios with real consequences, their ability to ensure justice remains an open question. The path forward requires careful testing, clear communication of limits, and a human who remains capable of checking the work, rather than just accepting the computer's word.