Deepfake technology poses an existential threat to trust

The claim asks whether deepfakes merely create serious new risks of deception or threaten the continued ability of society to trust audiovisual evidence, information institutions, and other people. The distinction matters: evidence of fraud, uncertainty, or localized distrust can establish substantial harm without establishing an irreversible or society-wide collapse of trust. The assessment therefore turns on the magnitude, persistence, and generality of observed effects, as well as the capacity of technical and institutional responses to preserve authentication and accountability. The strongest support for the claim is that a deepfake can damage institutional trust even when its effect is not confined to persuading viewers that the depicted event occurred. A peer-reviewed experiment found that exposure to a fabricated infrastructure-failure video reduced trust in government, directly demonstrating spillover from false audiovisual content to attitudes toward an institution. Political experiments likewise show that officials can sometimes benefit by falsely labeling authentic unfavorable information as misinformation, indicating that a polluted information environment can weaken accountability even without a successful fake. Institutional assessments place synthetic content within a broader misinformation risk and identify applications in fraud, impersonation, nonconsensual sexual material, evidence manipulation, and influence operations, suggesting that pressure on trust can arise across several domains at once (see Figure 1). A particularly consequential mechanism is the “liar’s dividend”: awareness that audiovisual material can be fabricated gives interested parties a plausible vocabulary for denying genuine evidence. The mechanism is supported both by legal analysis anticipating false denials and by experimental evidence that misinformation claims can reduce accountability under some political conditions. This problem is amplified by evidence that unaided viewers have difficulty distinguishing high-quality deepfakes from authentic videos, often overestimate their ability, and show variable performance across stimuli and study designs (see Figure 3). Public discussion during the Russian invasion of Ukraine also exhibited concern about doubting authentic material and losing confidence in reliable knowledge, showing that the epistemic risk is perceived in an actual conflict setting, although that study did not measure population-level causal effects. The threat is most credible as a compounding force in environments already marked by polarization, platform amplification, weak institutions, or unequal verification capacity. Cross-country election analysis and law-enforcement assessments describe harms arising through existing social and institutional vulnerabilities rather than through audiovisual realism alone. The available assessments support concern that synthetic media can scale established abuses and broaden the number of settings in which identity, records, and testimony become contestable. These sources establish growing operational challenges and multiple harmful uses, but they do not directly quantify the amount or permanence of society-wide trust loss. The decisive challenge to the claim is that the evidence establishes bounded harms and plausible systemic risks, not the destruction of society’s capacity to trust. The experimental findings concern effects that vary by response and context, while the conflict-discourse study measures expressed concern more directly than actual population-level collapse. Expert risk rankings and law-enforcement reports identify serious dangers, but they address broader misinformation categories or prospective criminal uses rather than proving that deepfakes alone produce universal, irreversible distrust. Accordingly, the evidence supports a serious threat to particular trust relationships more strongly than an existential threat to trust as a social capacity. Nor does detection research show categorical helplessness: large-scale experiments found complementary strengths between human crowds and a leading model. People outperformed the model on some videos, and model predictions could improve aggregate crowd judgments, although erroneous machine advice could also mislead. Together with the documented weaknesses of unaided viewing, this evidence favors layered verification over either confidence in casual inspection or the conclusion that verification is impossible. Trust can also be reorganized around authenticated history and institutional procedures rather than resting solely on visual appearance. The C2PA specification provides for cryptographically signed credentials that record provenance and editing history, thereby offering an authentication layer independent of detecting visual artifacts. Such credentials do not prove that the depicted event is true, do not protect unsigned media, and depend on trustworthy implementation and broad adoption. Institutional reports and reviews therefore recommend a portfolio of rapid authentication, trusted messengers, provenance, platform procedures, public education, law, and governance rather than reliance on any single detector. The demonstrated availability of these tools supports adaptability, although their effectiveness at scale is not yet conclusive. Finally, reviews characterize deepfake technology as dual-use and describe beneficial applications alongside impersonation, manipulation, and other harms. That evidence indicates that effects depend on use and context rather than following inevitably from the existence of synthetic-media tools. The evidence is best understood conditionally: deepfakes can produce acute and consequential trust harms, but their severity depends heavily on political context, distribution systems, verification speed, and institutional capacity. Theoretical work and cross-country election analysis identify polarization, platform amplification, weak institutions, and unequal response capacity as conditions that can intensify harm, while the infrastructure experiment confirms that institutional distrust can arise in at least one controlled scenario. The theoretical analysis does not estimate prevalence, and the experiment’s single scenario and study population constrain generalization to durable society-wide distrust. The liar’s dividend may be more resistant to ordinary fake detection than direct deception because the relevant harm is doubt about authentic evidence rather than belief in fabricated evidence. For that reason, defenses that positively authenticate source and editing history address a different problem from classifiers that merely search for generated artifacts. Positive authentication can narrow opportunities for false denial, but it remains incomplete for unsigned material and cannot independently establish that a recorded event is truthful. Generation and detection are better characterized as an evolving contest than as a settled victory for either fabricators or defenders. Human and automated judgments each have weaknesses, while combined judgments can improve performance under some conditions. The defensible long-run characterization is therefore an ongoing authentication and governance challenge, not guaranteed collapse and not guaranteed technical control. The principal evidence gap is the absence of longitudinal, population-level proof linking deepfake exposure to persistent and society-wide collapse of trust. Controlled experiments, thematic analyses, technical specifications, and threat assessments answer different parts of the question, but none alone resolves whether localized effects accumulate, fade, or generalize across institutions and relationships. Evidence is also thin on the real-world effectiveness, adoption, and distributional consequences of layered countermeasures at societal scale. A further structural limitation is that unresolved conflict-of-interest classifications reduce confidence in how some institutional and industry sources should be weighted. Figure 2 cannot be used to establish growth in this synthesis because the frozen evidence bundle contains no admissible evidence record for the Sensity AI or Deeptrace chart; scale claims based on that figure therefore remain outside the audited evidentiary basis. On the current evidence, deepfakes pose a serious and potentially systemic threat to particular forms of evidentiary and institutional trust, but the stronger existential claim—that they threaten to destroy society’s ability to trust—is not established. Confidence in this balanced judgment is high because multiple peer-reviewed experiments, reviews, and institutional analyses converge on real but context-dependent harm while stopping short of demonstrating universal or irreversible collapse. The dominant uncertainty is whether repeated exposure and the liar’s dividend will cumulatively outrun provenance, verification, and institutional adaptation over time, compounded by unresolved conflict-of-interest classifications for some sources.