Why Strong AI Governance Could Shape Public Acceptance of Deepfakes

A new study suggests that the public's relationship with deepfake technology is more complicated than simple fear or rejection. People may recognize serious risks while still accepting beneficial uses, provided they believe the rules are credible.

Published in the MDPI journal Information, the study, "Public Trust in Generative AI: Risk Perceptions, Regulatory Safeguards, and the Acceptance of Deepfake Technology," was authored by Cathrine Linnes, Giulio Ronzoni, Joseph Lema, Babu George and Jerome Agrusa. Drawing on survey responses from 924 participants across several countries, the researchers examined how perceived benefits, privacy concerns, misinformation risks and demographic differences shape trust in deepfakes under assumed regulatory safeguards.

When respondents evaluated deepfake technology in a scenario where privacy-protecting laws and guidelines were assumed to exist, perceived benefits mattered more for trust than perceived risks. It means people appeared more willing to focus on potential value once they believed institutions could manage the danger.

Public Anxiety Is Overwhelming, but Rejection Is Not Absolute

The survey captures the deep contradiction at the heart of generative AI. A majority of respondents, 54.3%, said deepfakes can be useful. Yet 84.8% believed the technology can mislead people, while 82.7% said it puts personal data at risk through unauthorized use of images or voices.

This is a picture of people making a risk-benefit calculation in real time. Respondents could see value in education, digital content creation and creative media, but they also worried about fraud, blackmail, political manipulation, reputational damage and the growing difficulty of distinguishing authentic content from fabrication.

The qualitative responses make the stakes especially clear. Participants described deepfakes as a threat to reputations, democratic processes, businesses and personal well-being. They worried that manipulated content could remain online long after being debunked, making reputational recovery difficult for both individuals and organizations. They also linked deepfakes to propaganda, political warfare, online harassment and social isolation.

What matters is that most respondents did not call for blanket prohibition. Instead, they treated legitimacy as conditional. They were open to education, entertainment, accessibility and creative applications, but only when accompanied by consent, labeling, legal protection and responsible use.

Regulation Appears to Shift Attention From Danger to Value

The study's most consequential result comes from its regression analysis. The model explained 36.3% of the variation in trust toward regulated deepfake technology. Belief that deepfakes could improve online learning was the strongest positive predictor of trust, followed by the view that they could positively transform digital content creation.

On the other hand, concerns about misinformation, personal-data risk and privacy did not remain statistically significant once other variables were considered. The authors describe this pattern as the "regulatory absorption of risk." In practical terms, respondents seemed willing to delegate part of the risk-management burden to institutions. Once safeguards were assumed, they focused less on whether harms existed and more on whether the technology offered meaningful benefits.

The study does not prove that regulation causes trust. Its cross-sectional design and hypothetical scenario cannot establish causality. The findings show an association consistent with the idea that credible governance may reduce the independent weight of perceived risk in public judgments.

Even with that caution, the policy lesson is substantial. Public acceptance may depend less on persuading people that deepfakes are safe, an increasingly implausible claim, and more on proving that institutions can contain misuse, enforce accountability and preserve legitimate benefits.

Trust, in other words, may attach not to the technology itself but to the system around it.

Demographic Divides Show Trust Will Not Be Built Uniformly

The study also challenges simplistic assumptions about who fears or accepts generative AI. Women reported stronger concern about misinformation, privacy, criminal misuse and personal exposure, but they also expressed greater trust when privacy protections were assumed.

Respondents aged 45–59 showed the strongest support for educational applications and the highest trust in regulated deepfakes. Those aged 60 and above were the most skeptical, reported the greatest difficulty distinguishing authentic from manipulated media and showed the lowest trust even under safeguards.

Education produced another nuanced pattern. Respondents with higher educational attainment tended to recognize both the benefits and risks of the technology more strongly, while also showing greater trust under regulation. The effect sizes were generally small, but the pattern suggests that greater awareness does not necessarily produce either enthusiasm or alarm. It may instead produce a more differentiated judgment.

For policymakers, this means communication cannot be one-size-fits-all. Older adults may need stronger evidence that safeguards work in practice, not merely assurances that rules exist. Groups more exposed to identity abuse, non-consensual content or reputational harm may place greater value on consent, remedy and enforcement.

Digital literacy programs should therefore move beyond generic warnings and address the specific vulnerabilities faced by different communities. It has particular relevance for developing countries. Where regulatory capacity is limited, legal remedies are slow and media literacy is uneven, the technology could generate deeper harm and lower trust. Yet those contexts may also benefit from educational, linguistic and accessibility applications.

The challenge is not to choose between innovation and regulation, but to build institutions capable of making innovation trustworthy.

The Real Policy Test Is Whether Governance Can Become Credible

Synthetic voices, AI assistants, multimodal systems and automated content tools will raise the same basic question: can the public trust the institutions governing technologies that are useful precisely because they are powerful? The answer will depend on whether safeguards are visible, understandable and enforceable. Governments need clear rules on consent, disclosure, identity protection, liability and redress. Platforms and developers need provenance tools, watermarking, authentication systems and rapid-response mechanisms, but they must also communicate honestly about the limits of detection.

The study stresses that because detection will remain imperfect, transparency about uncertainty is itself part of trust-building. Businesses should also treat deepfakes as a governance and reputational risk, not only a cybersecurity problem. A fabricated video can damage investor confidence, employee morale, consumer trust and institutional credibility before technical verification catches up. Civil society organizations and media institutions will need stronger verification capacity, while educators should demonstrate responsible uses rather than presenting the technology only as a threat.

The research has important limitations. The sample was concentrated in Norway and the United States, several variables relied on single-item measures, and the survey does not represent the world. The study also produced an unusual negative relationship between general perceived usefulness and trust. The authors suggest respondents may have interpreted "usefulness" as technological capability, including capability for misuse. That result needs replication.

That said, the broader message is relevant. Deepfakes are not merely a technical problem requiring better detection. They are an institutional test. The public appears willing to accept beneficial uses, but only when governance is credible enough to make risk feel manageable.