When Seeing Is Not Believing: How Synthetic Media Tests Public Trust

Updated October 9, 2026.

A short clip lands in a family, neighborhood, or workplace chat. It appears to show a candidate, school official, journalist, or loved one saying something shocking. The emotional pressure is immediate: forward it before it disappears, warn everyone, or answer with the same anger. The more useful first question is smaller and calmer: what exactly does this clip establish, and where did it come from?

Synthetic media—audio, images, and video made or altered with artificial intelligence—has moved from a specialist concern into ordinary civic and family life. It can be useful for satire, accessibility, translation, and creative work. But when fabricated material is presented as evidence, it can mislead voters, damage people who have little ability to respond, and make authentic reporting easier to dismiss. The challenge is not to panic about every unusual clip. It is to build habits and rules that make public information more trustworthy.

The human stakes go beyond an embarrassing political post. In March 2026, UN News reported on Kim and Peter Sawyer of Australia, who said they lost more than $2.5 million after engaging with a seemingly legitimate online investment advertisement. Their story concerned organised fraud, not an election clip, but it shows why realism matters: a believable digital message can exploit trust, urgency, and the ordinary wish to make a safe decision.[8] No single story proves how often any one kind of synthetic-media harm occurs. It does show why the people discussing this subject are not only technologists or campaign professionals.

Why this is a live public conversation

Hany Farid, a UC Berkeley digital-forensics professor, has become one visible public voice in this discussion. In an April 2026 Berkeley talk, Farid described the shift toward real-time deepfakes that can hold a live conversation and said people are only slightly better than chance at identifying AI-generated material in his research.[6] His point is not that truth has disappeared. It is that human eyes and ears cannot carry the entire burden of verification once convincing material is cheap to make and easy to distribute.

The wider news conversation reflects that concern. Reuters Institute’s Digital News Report 2026 found that concern about fake news rose four percentage points to 62% on average across its surveyed markets, while 45% of respondents still said they preferred news that did not take sides.[7] Those figures do not tell us that every reader has seen a deepfake. They do explain why people are asking harder questions about who created a clip, how a platform labels it, and whether a correction can reach people before the false claim does.

The technology is entering ordinary political communication

A March 2026 Reuters report described AI-generated advertisements appearing in U.S. midterm campaigning. One ad made a candidate appear to repeat inflammatory language; another made an opponent appear to say something about government policy that he had not said. Campaign representatives defended the material as an efficient way to communicate or as satire, while opponents called it misleading. Reuters reported that 28 states had passed laws addressing AI in political advertising, most emphasizing disclosure rather than an outright ban.[1]

That debate matters because the same tool can serve different purposes. A clearly labeled parody that no reasonable viewer would mistake for a recording is not the same as a fabricated statement designed to look like documentary evidence. A campaign may argue that editing a rival’s past words into a new visual format is political criticism. A voter may reasonably ask whether the presentation creates a false impression about what happened, when it happened, or who said it.

The human cost is easy to miss in an argument about tactics. Candidates and public officials may face harassment after a false clip spreads. Local election workers can become targets if a fake image appears to show wrongdoing at a polling place. Citizens may withdraw from public discussion because they assume every side is manipulating the record. Those harms do not depend on a deepfake persuading everyone; confusion and exhaustion can be enough.

Detection is not a magic answer

It is tempting to treat detection software—or a chatbot—as a digital lie detector. Current evidence argues for caution. In a 2025 evaluation plan, the National Institute of Standards and Technology (NIST) described testing both systems that generate synthetic content and systems that try to distinguish it from human-made content. The point of the program is to measure strengths and weaknesses, including when humans or machines can detect generated material.[2] The fact that such testing is needed is itself a reminder that detection is an evolving measurement problem, not a settled capability.

A 2026 Brennan Center test makes the limitation more concrete. Researchers asked six major chatbots to assess AI-generated images connected to elections. Most incorrectly treated some generated images as authentic, including images made by their own providers. Only one chatbot consistently checked embedded origin data in the test. The center’s conclusion was direct: a chatbot’s answer is not verification.[3]

There are two reasons this matters. First, a detector may fail when a model, editing program, screenshot, or social platform removes the clues it relies on. Second, an accurate label does not prove that the surrounding claim is true. An image can be genuinely photographed but falsely captioned, or AI-generated yet used in an obvious joke. Authenticity and meaning are separate questions.

The opposite problem is also serious. If people learn that some recordings are fake, they may begin dismissing real evidence as fabricated. Researchers and commentators call this the “liar’s dividend”: the benefit a wrongdoer gets when genuine proof can be waved away as AI. In that environment, demanding a perfect visual tell can weaken accountability rather than strengthen it.

What safeguards can and cannot do

One policy approach is provenance: attach information that records how a file was created or edited. The Brennan Center reported that California’s AI Transparency Act, effective in August 2026, requires covered AI platforms to embed origin records and offer a visible indicator and a way for consumers to access those records.[3] These measures can help journalists, platforms, and investigators trace material. They are less useful when content is re-recorded, cropped, compressed, or moved to a service that does not preserve the information.

Another approach is disclosure in political advertising. A clear, prominent notice can help a viewer understand that a scene is simulated. But disclosure is not a cure for false claims. A small label that appears briefly may be missed; a large label may still leave a vivid fabricated statement in memory. Rules also differ by jurisdiction, and many focus on campaigns rather than influencers or anonymous accounts that repost material.

Platforms face a genuine trade-off. Removing every manipulated clip would risk suppressing satire, commentary, and legitimate journalism. Leaving everything up can reward the most deceptive actors, especially during the short period before an election when corrections travel more slowly than sensational material. A reasonable system should distinguish context: label altered media, preserve access to credible corrections, act more quickly when content impersonates a person or interferes with voting, and explain decisions publicly.

A better public habit: verify the claim, not the pixels

Individuals cannot solve a structural problem alone, but a short pause can reduce avoidable mistakes. Before sharing a dramatic clip, use a three-check test:

Before you share: the three-check pause

1. Name the claim. Write down what the post is actually asserting. Is it claiming that a person said something, that an event happened at a specific place and time, or that someone needs immediate help? A real image can still carry a false caption.

2. Find the source. Look for the full recording, original account, date, location, and any missing context. A repost, cropped screenshot, or confident chatbot answer is not an original source.

3. Seek independent confirmation. Search for reporting or an official statement that is not simply repeating the same clip. For an election question, check the relevant election office. For an urgent call involving money or personal information, the Federal Trade Commission advises calling the person back using a number you already know and verifying the story through another route.[9]

This is not a demand that every reader become a forensic analyst. It is a way to refuse the false choice between instant belief and permanent cynicism. When the evidence is incomplete, the responsible action may simply be to wait.

It also helps to separate three judgments: “Was this file generated or altered?” “Does it accurately represent what happened?” and “What does it prove?” An image may pass the first test and fail the second. A real speech may be clipped so that it reverses the speaker’s meaning. A genuine document may be old and irrelevant to the current claim.

Families, schools, and newsrooms can teach these habits without turning every reader into a forensic analyst. UNESCO’s 2025 discussion of deepfakes argues that education should build “epistemic agency”—the ability to navigate uncertainty and make responsible judgments—not merely train people to spot visual glitches.[4] That is a useful shift. Trust should rest less on whether a person can detect a strange hand and more on whether institutions show their sources, methods, corrections, and limits.

The choice is not technology or trust

Synthetic media will not disappear, and banning every use would discard legitimate creative and civic applications. The stronger response is layered: provenance standards, prominent disclosures, focused laws against impersonation and election interference, platform accountability, independent journalism, and public education. Each layer has weaknesses. Together, they make deception harder and correction more visible.

The most important goal is not to make citizens suspicious of everything. It is to make skepticism proportional and constructive. Public trust does not mean believing every video; it means having reliable ways to check important claims and institutions willing to show how they know. In an age when seeing is no longer enough, that shared practice—verification, transparency, and a willingness to correct the record—may be the most durable civic technology we have.

A public face for media literacy

That shift is already visible in the culture around misinformation. Alex Mahadevan, a journalist and media-literacy educator who directs Poynter’s MediaWise project, is one public-facing example. His work on generative AI, verification, and online misinformation translates a technical problem into repeatable questions ordinary people can ask before they share something.[5]

The relevance is not that one educator can decide whether a particular video is real. He cannot, and neither can a chatbot. It is that public trust depends on habits people can use: checking the original source, looking for independent confirmation, noticing when a claim has been stripped of date or context, and being willing to pause before amplifying a clip that flatters an existing belief. That is cultural literacy, not a digital lie detector.

References

[1] Joseph Ax and Helen Coster, Reuters, “AI deepfakes blur reality in 2026 US midterm campaigns,” March 28, 2026. https://www.reuters.com/business/media-telecom/ai-deepfakes-blur-reality-2026-us-midterm-campaigns-2026-03-28/

[2] George Awad, Hariharan Iyer, Seungmin Seo, Peter Fontana, and Yooyoung Lee, National Institute of Standards and Technology, “2025 NIST GenAI (Pilot) Evaluation Plan for Image Discriminators,” March 14, 2025. https://www.nist.gov/publications/2025-nist-genai-pilot-evaluation-plan-image-discriminators

[3] Abdiaziz Ahmed, Brennan Center for Justice, “Chatbots Still Bad at Spotting AI-Generated Images Despite New Law,” October 5, 2026. https://www.brennancenter.org/our-work/analysis-opinion/chatbots-still-bad-spotting-ai-generated-images-despite-new-law

[4] Nadia Naffi, UNESCO, “Deepfakes and the crisis of knowing,” October 1, 2025; updated October 27, 2025. https://www.unesco.org/en/articles/deepfakes-and-crisis-knowing

[5] Alex Mahadevan, official profile and MediaWise work, https://alexmahadevan.com/ and https://www.poynter.org/mediawise/; public profile: https://x.com/AlexMahadevan (accessed October 9, 2026).

[6] UC Berkeley News, “Berkeley Talks: The erosion of shared reality in the age of deepfakes,” April 6, 2026. https://news.berkeley.edu/2026/04/06/berkeley-talks-hany-farid-on-deepfakes/

[7] Reuters Institute for the Study of Journalism, Digital News Report 2026: Executive Summary, June 16, 2026. https://reutersinstitute.politics.ox.ac.uk/digital-news-report/2026/dnr-executive-summary

[8] UN News, “Deepfakes, voice cloning and weaponised AI: Global wake-up call to organised fraud,” March 16, 2026. https://news.un.org/en/story/2026/03/1167144

[9] Federal Trade Commission, “Fighting back against harmful voice cloning,” April 8, 2024. https://consumer.ftc.gov/consumer-alerts/2024/04/fighting-back-against-harmful-voice-cloning