Can AI Detect a Photoshopped Diploma? Yes. Here Is How.

Every HR analyst, admissions officer, and background-check investigator we talk to asks the same question on the first call: can AI catch a Photoshopped diploma? The answer is yes, with one nuance. The detection methods are stable, well understood, and have been deployed in forensic labs for fifteen years. What changed in 2024 is the arrival of generative AI, which produces edits more sophisticated than a manual Photoshop pass. The forensic toolkit kept pace.

The methods below are the five most reliable. Each catches a different edit class. A production forensic engine runs all five in parallel and combines the scores into a single confidence verdict.

ELA is the workhorse. The technique was formalized by Neal Krawetz around 2007 and is the basis of FotoForensics. The principle is simple. JPEG compression is lossy: each save discards information. A region that has been edited and re-saved has lost more information than the surrounding pixels. If you re-save the whole image at a known quality (typically 95 percent) and diff against the original, the edited region jumps out as bright noise.

ELA works beautifully on a single Photoshop pass and a fresh save. It weakens as the image gets re-saved repeatedly, because the whole frame eventually converges to the same error level. For diplomas, this is rarely a problem: a forger usually edits once and exports once. Multi-edit workflows do exist, and they push the analysis to the next method.

JPEG ghost detection generalizes ELA. Instead of saving at one quality, the engine saves the image at every quality from 50 to 100 and watches how each region responds. An authentic image has one underlying quantization signature across the whole frame. A composited image has at least two: one from the original document, one from the pasted region. The two signatures produce a visible “ghost” at the quality factor that matches the pasted region’s history.

For diploma forgeries, JPEG ghosts often appear in three hotspots: the candidate name, the conferred date, and the GPA or honors row. The forger typically pasted text from a template into those fields and resaved the composite once.

If the document is a PDF rather than a flat image, metadata is the highest-yield first check. Every PDF carries a producer string, a creator string, and an XMP block that catalogs every editing tool that touched the file.

An authentic registrar-produced diploma carries a producer like Ellucian Banner, Workday Student, PeopleSoft, or, for older documents, a high-end scanner’s software. A producer of “Adobe Photoshop CC 2025” or “Microsoft Word” on a document claiming to be a 1998 diploma is decisive. Forgers often strip metadata to hide it; absent metadata is itself a signal, since an authentic diploma carries a full XMP trail.

The XMP block also records incremental update history. A PDF that has been edited inherits a chain of edit timestamps and tool signatures. The forensic engine reads the chain and surfaces anything inconsistent with a single-issuer workflow.

Pasted text rarely matches the original document’s font rendering. The differences are subpixel: kerning variance, baseline drift, anti-aliasing pattern. They are invisible to the eye and conclusive under measurement.

The engine extracts every glyph in suspect fields (name, date, GPA, honors row), measures kerning between pairs, and compares the variance against a reference distribution from the original document body. A spike in variance on the name line, against a flat distribution elsewhere, is a conclusive tell.

The same method catches font substitution. A document that uses Trajan Pro in the body and Arial in the candidate name has been edited, because no registrar tool produces that mix.

Copy-move detection finds duplicated pixel regions. The classic case is a seal cloned from one document to another: the same vector pattern, the same noise profile, at two different locations. A correlation pass across all small-window patches surfaces the duplicate.

Inpainting detection finds regions with a noise profile different from a single camera or scanner. Generative inpainting tools fill regions with statistically plausible content but introduce a distinctive smoothness that does not match the host image. The engine measures local noise statistics and flags any region that diverges from the rest of the frame.

Three attacks weaken the digital methods. Recognize them so the workflow degrades gracefully when one applies.

• ·Print and scan. Printing a fake to paper and re-scanning destroys ELA, JPEG ghost, and metadata signals. The scan inherits a single, uniform compression history. Font anomaly detection survives partially; copy-move survives fully. The decisive method becomes registry cross-reference: a perfectly-rendered fake from an institution that does not exist is still fake.
• ·Heavy re-encoding. Saving the JPEG five or six times at varying qualities homogenizes the error level. ELA degrades. JPEG ghost detection still finds the composite, since the underlying quantization signatures remain distinct.
• ·Fully generative documents.A diploma generated end-to-end by a diffusion model has no pasted regions and no edit history. ELA and ghost detection are irrelevant. The decisive methods become GAN/diffusion fingerprint detection (which catches the model’s characteristic frequency-domain signature) and registry cross-reference.

The takeaway: no single forensic method is a silver bullet. The methods compose. Layered, they catch the long tail.

Three browser tools run ELA at no charge:

• ·FotoForensics (the original Neal Krawetz reference implementation). Solid ELA, JPEG quality estimation, metadata extraction. The interface is utilitarian; the math is sound.
• ·29a.ch Photo Forensics. Browser-side ELA, noise analysis, level sweep. Useful for a quick spot check.
• ·fakeimagedetector.com. Hosted ELA, with a friendlier UI. Skip the ML-based “is this AI-generated?” verdict, which is less reliable than the underlying ELA.

Those tools are excellent for one-off checks. They do not combine methods, do not score a verdict, and do not run registry cross-reference. For a production workflow, the forensic engine has to do all three. Our AI document verification primer describes the production pipeline that bundles the methods above with registry checks.