Every week, image desks at news organisations around the world field photographs that were never taken by a human photographer. Some arrive via social media during breaking news events. Others slip into press releases or wire feeds. The question is no longer theoretical: how should a newsroom approach the task of spotting AI-generated images before they reach readers?

The honest answer is that no single method is foolproof, and detector accuracy should never be overstated. What we have instead is a combination of visual literacy, metadata inspection, and a longer-term infrastructure shift toward provenance that deserves more attention than it typically receives.

Visual signals worth knowing

Trained editors can catch a significant share of AI-generated images by looking carefully at specific areas where current generative models still produce characteristic errors. These are not guarantees of authenticity or inauthenticity, but they are useful starting points for triage.

  • Hands and fingers: Generative models have historically struggled with human hands, producing extra digits, fused fingers, or anatomically implausible joints. Newer models have improved here, so absence of hand errors does not clear an image.
  • Text within the image: Logos, signs, banners, and captions rendered inside a generated image often contain garbled, nonsensical, or inconsistent lettering. Real photographs capture real text.
  • Ears, teeth, and eyes: Look for asymmetry that goes beyond normal human variation, reflections in eyes that do not match the apparent light source, or teeth that appear smoothed into a uniform texture.
  • Background coherence: AI images frequently produce backgrounds that repeat patterns, merge objects illogically, or blur in ways inconsistent with real optical depth-of-field.
  • Jewellery and accessories: Earrings that differ between ears, watches showing impossible faces, or glasses with distorted frames are common tells.
  • Lighting and shadows: Multiple or contradictory light sources, shadows falling in implausible directions, and hair that appears to merge with the background remain frequent artefacts.

These visual checks take seconds once an editor internalises them. They are most useful during fast-moving news events, when a quick visual triage can flag an image for deeper review before publication.

Detection tools: useful, but limited

A range of commercial and open-source tools now claim to identify AI-generated images. Some analyse noise patterns and frequency artefacts in image data. Others use classifiers trained on known outputs from specific generative models. As we have explored in our broader analysis of why AI detectors disagree with each other, classifier-based tools are only as reliable as their training data, and generative models update faster than most classifiers can track.

For images specifically, the accuracy problem is compounded by compression. When an AI-generated image is screenshotted, shared via messaging apps, or re-uploaded to social platforms, the artefacts that detection tools rely on are often degraded or removed entirely. A tool that performs well in a controlled environment may produce unreliable results on a JPEG that has passed through several platforms. Newsrooms should treat detector outputs as one input among several, not as a verdict.

Understanding how detection technology works at a technical level is useful context for anyone evaluating these tools. The underlying mechanics explain both their power and their limits.

Why provenance is the more durable approach

If detection is unreliable, the more promising long-term approach is provenance: cryptographically binding metadata to an image at the moment of capture or creation, so that any downstream user can verify where an image came from and how it was processed.

The Coalition for Content Provenance and Authenticity, known as C2PA, has developed an open technical standard that embeds a tamper-evident manifest into image and video files. When a camera, editing tool, or platform implements the standard, the manifest records the device, the timestamp, any AI-assisted processing, and a chain of custody through subsequent edits. Major camera manufacturers including Canon, Nikon, and Sony have announced or shipped implementations. Adobe's Content Credentials system, built on C2PA, is already present in several of its tools.

Our detailed guide to how AI content detection works across modalities covers where provenance fits into the broader detection infrastructure. The key practical point for newsrooms is this: C2PA does not detect AI after the fact. It establishes a record at the point of creation, which is a fundamentally stronger guarantee than any post-hoc analysis.

The limitation today is adoption. An image without a C2PA manifest tells us nothing on its own, because the standard is not yet universal. During the transition period, newsrooms need both the visual literacy described above and a clear policy on how to handle images that arrive without verifiable provenance.

Building a practical newsroom workflow

Drawing on the approaches documented by organisations including the Reuters Institute, the Dart Center, and verification-focused outlets, a workable triage process looks roughly like this:

  • Run a reverse image search as a first step. If the image exists in an earlier context that contradicts the claimed origin, that is a red flag independent of AI.
  • Check EXIF and metadata fields. Legitimate photographs from modern cameras carry rich metadata. Stripped or absent metadata on a supposedly fresh news photograph warrants scrutiny, though metadata can be removed legitimately for privacy reasons.
  • Inspect for the visual signals listed above, paying particular attention to text, hands, and lighting.
  • Run detection tools as a supplementary check, treating results as probabilistic indicators rather than conclusions.
  • Check for C2PA Content Credentials where the manifest is present. Adobe's free Content Credentials Verify tool and the C2PA open-source validator can read manifests from supported sources.
  • When in doubt, contact the claimed source directly. Social engineering is a faster route to verification than any automated tool.

The comparison of leading detection platforms in our earlier coverage is a useful reference for understanding what tools exist in the broader detection landscape, even though image-specific tools are a distinct category.

Detection of AI images will remain an arms race for the foreseeable future. Our newsrooms are better served by building habits of provenance verification and visual literacy than by placing uncritical trust in any single tool. The infrastructure to support that is being built. Our task now is to understand it well enough to use it properly.

Sources

  • Coalition for Content Provenance and Authenticity (C2PA): C2PA Technical Specification, c2pa.org
  • Adobe: Content Credentials overview and Verify tool, contentcredentials.org
  • Reuters Institute for the Study of Journalism: Journalism, media, and technology trends and predictions, reutersinstitute.politics.ox.ac.uk
  • Hany Farid, UC Berkeley: published research on image forensics and AI detection limitations
  • Dart Center for Journalism and Trauma, Columbia University: verification guidance resources