Watermarking AI Images

Executive summary

OpenAI's decision to embed invisible watermarks and C2PA metadata into every AI-generated image it produces has landed at exactly the right moment — and the public knows it. Nearly 7 in 10 Americans surveyed responded positively to the move, with close to half calling it outright necessary, as synthetic imagery has quietly become a daily fixture of online life.

The urgency is hard to overstate. More than 8 in 10 respondents say they encounter images they suspect are AI-generated every single day. Yet widespread exposure hasn't built confidence — it's built anxiety. Respondents lean toward believing future AI images will become indistinguishable from real photographs, and trust in tech companies to label their content honestly is fragile at best.

OpenAI's watermarking initiative is broadly welcomed, but the data flags real headwinds: a skeptical fifth of the public is still unmoved, and more socially connected users are actually less enthusiastic about the change. The companies that make provenance tools work in the real world — not just on paper — will define what trust in AI-generated imagery looks like for years to come.

Context

OpenAI announced in May 2025 that it would attach two layers of provenance to every image produced by its tools: C2PA metadata, an open standard that logs creation details into a file's header, and Google DeepMind's SynthID invisible watermark, which encodes an imperceptible signal directly into pixel patterns. Together, they're designed to let platforms, journalists, and ordinary users verify whether an image came from an AI system — without relying on visual cues alone.

The announcement arrived against a backdrop of fast-moving concern. Synthetic images have become a standard instrument in disinformation campaigns, financial scams, nonconsensual intimate imagery, and political manipulation. Regulators in the EU and several U.S. states have begun mandating disclosure rules for AI-generated content, and the Coalition for Content Provenance and Authenticity (C2PA) — which includes Adobe, Microsoft, and major news organizations — has been pushing for exactly the kind of infrastructure OpenAI is now deploying.

To gauge public reaction, this pulse survey collected 154 responses across four questions: one measuring sentiment toward the watermarking move, one probing the frequency of AI image encounters, and two open-ended questions capturing concerns about AI imagery and trust in tech companies to label it accurately. Respondents also completed personality inventories (OCEAN and Prism), enabling correlations between individual traits and attitudes toward the initiative.

The timing matters. This is not a theoretical question about a future technology — it's a reaction to a specific corporate decision made right now, as the visual internet tips toward a reality where synthetic and authentic images coexist in every scroll. The survey captures a public that is already living inside that reality and forming firm opinions about who should be responsible for keeping it legible.

Findings

Nearly 7 in 10 Welcome the Watermark — But the Skeptics Are Real

When asked directly how they felt about OpenAI's announcement, 46.8% of respondents chose the strongest positive option: "Very positive — this is needed." Another 22.1% were "somewhat positive," putting combined support at 68.9%. That's a clear majority, and it suggests that whatever friction watermarking creates for creators or platforms, the public sees the tradeoff as worth it.

But the 20.1% who said they were "neutral or unsure" shouldn't be dismissed as soft support waiting to be won over. In combination with the 8.4% who called the initiative unnecessary, nearly 30% of respondents are either unconvinced or actively opposed. For a provenance system to function as a trust signal, it needs near-universal adoption — and that means the skeptical third matters as much as the enthusiastic majority.

One counterintuitive finding from personality data: respondents who score higher on sociability and extraversion are less favorable toward the watermarking initiative. More socially oriented users may be more attuned to the practical gaps in invisible watermarks — the fact that a watermark stripped during a screenshot or crop offers no protection at all in the feeds where they spend most of their time.

AI Images Are Everywhere. People Know It.

The frequency data is striking in its uniformity. A full 81.2% of respondents say they encounter images they suspect are AI-generated on a daily basis. Another 13.6% see them weekly. Only 3.2% say they rarely or never encounter suspected AI imagery online.

That near-total saturation reframes the entire watermarking debate. This isn't a precautionary policy for a threat that might materialize — it's a response to something already happening in nearly every user's daily feed. The sheer volume of exposure may itself be driving support for labeling: people who see synthetic images constantly have a practical, self-interested stake in being able to tell them apart from real ones.

Moderately trusting users are especially exposed: respondents who said they "somewhat trust" tech companies to label AI content were more likely to report daily encounters with suspected AI images, suggesting that the most engaged and aware users hold nuanced rather than binary views on corporate accountability.

The Deeper Fear: You Won't Be Able to Tell

Open-ended responses revealed a concern that runs deeper than any single policy fix. Respondents consistently expressed worry that AI images will eventually become indistinguishable from real photographs — and the scored data backs this up. Among the 75 respondents whose free-text answers were analyzed on a detectability dimension, the mean score was -0.40 on a -1 to +1 scale, a statistically significant lean toward the view that the gap between synthetic and authentic imagery will close entirely.

One respondent put it plainly: "As the technology improves, it could conceivably become very difficult to tell AI from genuine images — which could be used in harmful ways." That anxiety runs through the data like a current. It's not that respondents distrust OpenAI specifically — it's that they don't believe any technical solution will keep pace with the generative models producing the images in the first place.

This fear feeds directly into low trust. Free-response answers to the trust question were dominated by skepticism, and personality data found a meaningful negative correlation between neuroticism scores and trust in tech companies — meaning anxiety-prone users are the least likely to believe corporate labeling commitments will hold. For a system built on brand assurance rather than independently verifiable enforcement, that's a structural problem.

Harm Is Real, but the Public Is Split on How Serious

Respondents are not of one mind about whether AI-generated images constitute a genuine societal threat. Among the 110 respondents whose concern-focused free-text answers were scored on a harm perception dimension, the mean score was -0.37 — a modest lean toward viewing AI imagery as seriously harmful, but with enough variance to call the distribution polarized rather than consensus.

The harms people named were concrete: deepfakes used as false evidence, images used to embarrass or harass, scams built on fabricated visual proof, and the creeping erosion of trust in any photograph. "They could be used to fake evidence" and "deep fakes are a concern" appeared as recurring themes. But a meaningful minority reported no concerns at all — a segment whose skepticism about harm may overlap with the 8.4% who called OpenAI's watermarking initiative unnecessary.

On artist impact, 42 respondents whose answers touched on creative economics scored a mean of -0.21 — a moderate lean toward the view that AI imagery threatens artists' livelihoods and devalues human skill. For this group, watermarks are less a neutral technical feature than a minimum-viable safeguard against an industry-reshaping force. Their frame is defensive, not celebratory.

Conclusion

OpenAI's watermarking move is the right step in the wrong vacuum. Public support is real and broad — but it rests on a foundation of anxiety rather than confidence. People back labeling not because they trust the system to work, but because they're already drowning in synthetic imagery and have no other tools to reach for.

The signals to watch next are adoption and enforcement. C2PA metadata is only as durable as the platforms and apps that read and display it. SynthID watermarks are only as useful as the detection infrastructure built around them. If watermarks become a check-the-box compliance gesture rather than a genuinely accessible signal — stripped by screenshots, ignored by social platforms, invisible to the average user — the skeptics in this data will be proven right.

The practical implication for OpenAI and the broader provenance ecosystem is clear: invest in the last mile. Third-party verification, public detection tools, and platform-level commitments to surface provenance labels are what converts a technical standard into actual public trust. The public has signaled it wants this to work. The question is whether the infrastructure will catch up before the images become truly indistinguishable.