Published February 25, 2025 · By CampaignsLive · Production
A market has emerged for tools that claim to detect AI-generated imagery. The pitch is straightforward: brand teams, journalists, social platforms, and regulators all want to know whether a given image was AI-generated, and a tool that could answer the question reliably would be valuable. Several products targeting this market are commercially available by 2025.
Most of them do not work. The reasons they do not work are technical, structural, and probably permanent. The implication for brand teams is that AI detection cannot be a load-bearing piece of any production process; the alternative — content provenance from the production side rather than detection from the consumption side — is where the working solutions actually live.
This is a working assessment.
Why detection does not work
Three structural reasons.
The cat-and-mouse dynamic is asymmetric in favor of generation. AI image generation improves continuously. Detection tools have to keep up with each improvement. When detection tools are trained against the current generation of generative models, they perform well on output from those models. When the generative models update — which happens every few months — the detection tools degrade until they are retrained against the new generation. The detection side is always behind, by varying amounts.
Compression, post-processing, and screen capture defeat most detection signals. Detection tools typically look for statistical artifacts that AI-generated imagery leaves behind: pixel-level patterns, frequency-domain signatures, metadata residues. Almost any post-processing — JPEG re-compression, social-platform resizing, screen capture, format conversion — degrades these signals to the point where they are not reliably detectable. Most AI-derived imagery in the wild has been through enough post-processing that the original detection signals are gone.
False positives at production volume are unacceptable. A detection tool that produces a 5% false positive rate looks acceptable in a demo but is unusable at production volume. A brand running 10,000 creative variants per quarter cannot have 500 of them flagged as AI-derived when they are not. The tools that achieve usable false-positive rates do so by missing more of the actual AI-derived content; the trade-off is unavoidable and structural.
The cumulative effect is that AI detection tools, in 2025, produce results that are reliable enough to be interesting and unreliable enough to be unusable as the basis for production decisions.
What does work: provenance from the production side
The working alternative to detection is content provenance: documentation, attached to the asset at the time of production, that records what tools were used and what inputs produced the output. The C2PA standard (Coalition for Content Provenance and Authenticity), launched in 2021 and significantly matured through 2023 and 2024, is the dominant infrastructure for this approach.
How C2PA actually works. A C2PA-compliant production tool embeds cryptographically-signed metadata into the output file. The metadata records the production history — which tool, which version, which inputs, when, by whom — and is signed by the producing entity. Downstream consumers (platforms, brands, regulators) can verify the signature and read the production history.
Why this scales where detection does not. The provenance is attached at production time, when the production history is known. Detection has to infer the history after the fact, which is the harder problem. As long as the signed metadata is preserved through downstream handling, the provenance question can be answered definitively rather than probabilistically.
The weakness of C2PA. The metadata can be stripped. Many social platforms still strip image metadata on upload. Many production tools do not preserve C2PA metadata through their output pipeline. The standard is well-designed; the ecosystem support has been uneven.
The trajectory through 2024 and into 2025 was toward broader adoption. Adobe’s Content Credentials feature implements C2PA across the Creative Cloud stack. Microsoft, Google, and OpenAI have committed to C2PA support in their generative output. Several social platforms, including LinkedIn and Meta on specific surfaces, have begun displaying Content Credentials on uploaded media. The infrastructure is converging on a working standard.
What brand teams should do
Three working positions, by 2025.
Do not rely on AI detection tools for production decisions. The tools exist, they are interesting, they should not be load-bearing. A workflow that depends on detection to identify AI-derived content reliably will be wrong frequently enough to create real production problems.
Maintain provenance from your own production side. For every piece of creative that goes through brand-side production, document the tools used, the inputs provided, the outputs produced, and the timestamps. This documentation is the brand’s authoritative record of what was AI-generated and what was not. Production-grade AI tools maintain this automatically; brands that use less rigorous tools end up reconstructing the provenance manually.
Adopt C2PA where the production stack supports it. Tools that emit Content Credentials produce outputs that downstream consumers can verify. The infrastructure is mature enough by 2025 that this should be a procurement consideration: tools that support Content Credentials are preferable to tools that do not, all else equal.
What this implies about the broader trust environment
The shift from detection to provenance is the same pattern that other categories of verification have followed historically. The early stage of any new media category — photography in the late nineteenth century, broadcast in the twentieth, the internet in the late nineties — relied on detection-of-fakes as the verification mechanism. The mature stage of each category shifted to provenance-of-originals: standards, credentialing, attribution practices that made the question “is this real?” answerable from the production side rather than the consumption side.
Generative AI is going through the same shift now. The detection-side tools are visible because they fit the existing trust-verification mental model. The provenance-side infrastructure is becoming the working solution because it is the only approach that scales.
For brand teams, the practical implication is to build production-side provenance infrastructure into the standard creative workflow, regardless of whether the immediate compliance environment demands it. The infrastructure is becoming a basic operational hygiene requirement for any brand running significant creative production. Brands that build it early absorb the operational cost gradually; brands that retrofit it later absorb it all at once.
For the related discussion of compliance requirements that intersect with provenance, see The EU AI Act, Translated for Brand Teams. For the cultural moment that made the trust environment shift visible, see The Pope in a Puffer Jacket.