AI Image Detection vs. Text Detection: Why They're Completely Different Problems

An AI text detector and an AI image detector share a marketing category and almost nothing else. Text detection measures the statistical predictability of word choices; image detection looks for pixel-level generation artifacts, frequency-domain fingerprints, and — increasingly — cryptographic provenance metadata like C2PA Content Credentials. Different signals, different failure modes, different accuracy math. A vendor's reputation in one says nothing about its competence in the other.

That last sentence is the practical takeaway, and it cuts both ways. A company with a respected text detector has no automatic claim to image expertise, and the fluent image-forensics labs have no automatic claim on text. This post explains what each kind of detection actually measures, why the image side has tools the text side structurally can't have, what "AI image detector" products really check under the hood, and how to read accuracy claims in each domain without being fooled by the shared branding.

One scope note before anything else, stated plainly because our editorial policy requires it: HumanFlow detects text only. We don't analyze images, and nothing below is a pitch for a product we don't make. It's a map of a neighboring territory that our readers keep asking about — usually because a vendor's image claims made them wonder about its text claims, or vice versa.

What text detection actually measures

Text detectors work on a single deep idea: language models generate text by repeatedly picking probable next words, so model-generated text is, on average, more probable than human text.

The two workhorse measurements are perplexity — how surprising each word is to a reference language model — and burstiness — how much sentence length and structure vary across a passage. Human prose lurches. It mixes a five-word sentence with a forty-word one, reaches for the odd word, backtracks, over-explains one point and under-explains the next. Model output is smoother: consistently mid-length sentences, consistently probable words. A detector scores your document on these properties and compares the result to a vendor-chosen threshold. Above the line: "likely AI." Below: "likely human." (The threshold is a business decision, which is why the same essay gets different verdicts from different tools — we walk through the mechanics in how AI detectors work.)

Notice what this method cannot have: hard evidence. There is no residue in a paragraph that proves a machine made it, because text is just a sequence of symbols anyone could have typed. The detector is making a statistical inference about style, and every consequence follows from that. Short passages defeat it (too little signal). Editing defeats it (mixed signal). Non-native English trips it (human writing that happens to be statistically smooth — the documented disaster we cover in the false-positive literature, and in its strangest form in what happens to translated text). Even OpenAI couldn't make the inference reliable for its own models' output, which is why it retired its classifier in July 2023 at a 26% catch rate — a story worth reading in full.

Text detection, in short, is stylometry: judgment about statistical style, performed on a medium that carries no physical trace of its origin.

What image detection actually measures

Images are different in kind, not degree. An image is a dense grid of millions of correlated pixel values, and the process that generated those values leaves physical-layer traces that have no analogue in text. Image detectors hunt three broad families of evidence.

Generation artifacts. Generative models make characteristic mistakes. The famous ones are semantic — six fingers, mangled written signage in the background, earrings that don't match, hair merging into collars, shadows disagreeing about where the sun is. Detectors (and trained humans) can catch these. The subtler ones are textural: skin that's too uniformly smooth, backgrounds with a particular synthetic blur, repeating micro-patterns in hair or foliage. These artifacts are real evidence in a way no awkward sentence ever is — but they shrink with every model generation, and the semantic ones (the hands) have largely been fixed by current tools.

Frequency-domain fingerprints. This is the family with no text equivalent at all. Transform an image into the frequency domain — a mathematical representation of how fine detail is distributed — and generated images have historically shown telltale spectral patterns: grid-like peaks from GAN upsampling layers, unnatural distributions of high-frequency energy from diffusion models' denoising process. The camera pipeline (lens, sensor, demosaicing, JPEG) leaves its own distinct spectral signature that generators imitate imperfectly. Forensics models trained on these patterns can identify generated images with high accuracy when they've seen that generator family before — a condition doing a lot of work, as the benchmark numbers below show.

Provenance metadata. The third family isn't inference at all — it's paperwork, done cryptographically. More on it in a moment, because it's the biggest structural difference between the two fields.

The evidence on how well the inference-based approaches perform comes from benchmarks like GenImage (NeurIPS 2023), a dataset of roughly 2.7 million real and generated images across eight generators including Midjourney and Stable Diffusion. The headline pattern: detectors trained and tested on the same generator exceed 98.5% accuracy — near-perfect. Move to a generator the detector wasn't trained on and it collapses; a ResNet-50 trained on Stable Diffusion 1.4 managed only 54.9% on Midjourney images — a coin flip — and averaged about 66.9% across generators. Degrade the image the way the internet does and things get worse: JPEG compression at quality 30 dropped that detector to 51.2%, and low-resolution versions to 57.4%. Some architectures held up much better under compression, but the generalization gap is the field's defining open problem.

If that shape sounds familiar — excellent in-distribution, brittle everywhere else — it should. It's the same shape as text detection's results on unedited versus edited output. The two fields fail in structurally similar ways for structurally similar reasons: both are classifiers chasing a moving target that's actively optimized to look like the real thing.

The provenance layer: what images have that text never will

Here's where the fields genuinely diverge, and where "current status" matters.

The Coalition for Content Provenance and Authenticity (C2PA) — whose steering committee now includes Adobe, Amazon, BBC, Google, Meta, Microsoft, OpenAI, Sony, TikTok, and Truepic — publishes an open standard called Content Credentials: a cryptographically signed manifest attached to a media file recording where it came from and what was done to it. The C2PA site describes it as a nutrition label for digital content, and the current specification family (version 2.3+ as of this writing, per c2pa.org) is in active deployment. Adobe has shipped Content Credentials workflows in Photoshop, Premiere Pro, and Lightroom for years. OpenAI attaches C2PA metadata to its generated images. Google has been rolling out Content Credentials verification across its surfaces alongside SynthID, DeepMind's watermarking system that embeds an imperceptible signal directly into generated pixels. Several camera manufacturers have shipped or announced in-camera signing for professional newsroom bodies, though the supported-model lists move often enough that any specific list here would be wrong by the time you read it.

When provenance works, it changes the epistemics completely. A valid Content Credential is not a statistical guess; it's a signed statement — "this file was produced by DALL·E," or "this photo was captured on this camera and cropped in Lightroom" — that either verifies or doesn't. No text detector can offer anything comparable for a paragraph pasted into a submission box, because paragraphs pass through human retyping, and retyping strips everything. (Text watermarking exists — SynthID has a text variant — but it requires the generator's cooperation and degrades under paraphrase, which is why it hasn't rescued text detection.)

The honest caveats are just as important. Provenance metadata routinely dies in transit: screenshots, platform re-encoding, exports, and ordinary uploads can strip or break the manifest, and a stripped credential proves nothing — a file with no Content Credential might be an unsigned camera photo, a generated image that lost its label, or anything in between. Absence of provenance is not evidence of anything. The system also only covers cooperating tools; nobody signing their output is nobody caught by it. Provenance is a floor being built under honest actors, not a net catching dishonest ones. Still, it's a floor. Text detection doesn't have one.

What "AI image detector" tools actually check

Consumer-facing image checkers — the upload-a-picture-get-a-verdict sites — are typically doing some mix of four things, and the mix matters more than the marketing.

A tool leaning on metadata checks is fast and cheap and defeated by a screenshot. A tool running a trained classifier over pixels can catch what metadata misses, but inherits the generalization gap — strong on last year's generators, weaker on last month's, and often silent about which generators it was trained on. A tool checking watermarks can only find marks that cooperating generators embedded. The better products layer all of these and report which layer fired; the worse ones report a single confident percentage and let you assume it means more than it does. If that move sounds familiar from text-detector marketing, it's the same move. A percentage with no stated methodology is decoration in either medium — the questions to ask are the ones we lay out for reading any accuracy claim: tested on what, generated by what, degraded how, at what false-positive rate?

Why reputation doesn't transfer between the two

It should be clear by now that the two problems share almost no machinery. Text detection expertise is language-model expertise: token distributions, stylometry, threshold calibration on prose. Image detection expertise is signal processing and computer-vision forensics: sensor pipelines, spectral analysis, compression behavior, plus the provenance-standards world of certificates and signing chains. Different training data, different research literature, different failure modes, different engineering staff. A vendor excellent at one starts from roughly zero on the other.

Yet the market keeps bundling them, because "AI detector" is one search term and adding an image tab is one product decision. Some multi-modal vendors are genuinely competent in both; some have a serious flagship and a checkbox afterthought — and the accuracy study they publicize is nearly always about the flagship. So evaluate each mode as if from a separate company, because functionally it is. Ask for the image-side methodology specifically. If the vendor's evidence for its image detector is its text detector's reputation, you have your answer.

We'll say our own version plainly, since we just told you to demand this of everyone: HumanFlow's AI detector analyzes text, sentence by sentence, and nothing else. We don't detect images, we don't plan to pretend otherwise by shipping a thin wrapper, and — same policy as on the text side — we publish no accuracy percentage without published methodology, and we don't promise our humanizer beats any detector, because nobody can honestly promise that.

How to actually verify an image today

Detection tools are one input, not a verdict — same sermon as the text side, different tactics.

Check provenance first, because it's the only step that can end the inquiry: Content Credentials inspection (the C2PA ecosystem's verify tools let you upload a file and read any attached manifest) either shows you a signed history or shows you nothing. Reverse-image search comes second — a generated "news photo" usually has no earlier, larger, credited version anywhere, while a real one usually does; provenance-by-context predates provenance-by-cryptography and still works. Then look at the picture like an editor: written text in the background, physics of shadows and reflections, jewelry symmetry, where hair meets skin. Then, if you use an automated checker, weight it by what you now know — heavily compressed, re-uploaded social-media images sit exactly where benchmarks show classifiers are weakest, so a confident score on a screenshot-of-a-screenshot deserves very little.

And in the other direction: if you make images with AI tools, leave the Content Credentials on. The system only builds trust if the honest majority carries the label.

The two fields are converging on the same lesson

Strip away the mathematics and both fields have arrived, expensively, at one conclusion: inference about origin is fragile; records of origin are not. The text world learned it when the maker of ChatGPT couldn't reliably identify ChatGPT's own output. The image world is learning it through the generalization gap — and is responding by building the provenance infrastructure that text, passed through human hands and keyboards, can never fully have. Image detection will likely keep its structural advantages: physical-layer evidence and a workable signing standard. Text detection will keep its structural burden: pure statistical inference, no floor beneath it.

Which is exactly why the accuracy conversation has to be had per-medium, per-tool, per-condition — never at the level of a brand name. That's the discipline this whole accuracy pillar exists to teach.

FAQ

Are AI image detectors more accurate than AI text detectors? Under lab conditions, image classifiers post higher numbers — 98%+ on benchmark images from generators they trained on. But cross-generator accuracy fell to roughly 55–67% in the GenImage benchmark, and heavy JPEG compression pushed some detectors near coin-flip. The honest answer: image detection has stronger evidence available (artifacts, provenance) but is not reliably "more accurate" in the wild.

Can the same tool detect both AI text and AI images? A vendor can ship both, but they are separate systems with separate training data, methods, and failure modes — text tools read word-choice statistics, image tools read pixels and metadata. Evaluate each mode on its own published methodology, not on the vendor's overall reputation.

What are C2PA Content Credentials? An open standard from the Coalition for Content Provenance and Authenticity — steered by Adobe, Google, Microsoft, Meta, OpenAI, Sony, TikTok, the BBC, and others — that attaches a cryptographically signed history to a media file: what created it and what edited it. A valid credential is direct evidence, not a statistical guess; but credentials are easily stripped by screenshots and re-uploads, and a missing credential proves nothing.

Does a screenshot really defeat AI image detection? It defeats the metadata layer completely — screenshotting creates a new file with no Content Credentials. Pixel-based classifiers can still fire on a screenshot, but re-encoding and rescaling measurably reduce their accuracy, which is why heavily reshared social-media images are the hardest case.

Why is there no C2PA equivalent for text? Text survives retyping; cryptographic signatures don't. A paragraph copied by hand into a submission box carries no metadata, so signing schemes can't follow it. Text watermarking (like SynthID's text variant) exists but requires the generator's cooperation and weakens under paraphrase.

Do image generators like DALL·E label their output? Increasingly, yes. OpenAI attaches C2PA metadata to generated images, Google embeds SynthID watermarks in its image outputs and is rolling out Content Credentials verification across its products, and Adobe's tools have shipped Content Credentials for years. The label's survival, though, depends on platforms not stripping it in transit.

Does HumanFlow detect AI images? No. HumanFlow analyzes text only — sentence-level statistical detection, nothing else. We'd rather say that plainly than bolt on an image checkbox we couldn't stand behind.

Key facts

• Text detectors infer origin from perplexity and burstiness — statistical style — because text carries no physical trace of its generator; image detectors can use pixel artifacts, frequency-domain fingerprints, watermarks, and signed metadata.
• The GenImage benchmark (NeurIPS 2023; ~2.7 million images, 8 generators): detectors exceeded 98.5% accuracy on generators they trained on, but a ResNet-50 trained on Stable Diffusion 1.4 scored 54.9% on Midjourney and averaged ~66.9% cross-generator.
• The same benchmark showed JPEG compression (quality 30) cut that detector to 51.2% — image detection degrades sharply on the compressed, re-uploaded images that dominate social media.
• C2PA Content Credentials (spec family 2.3+, per c2pa.org) is in production: Adobe apps, OpenAI image outputs, Google surfaces, and in-camera signing on bodies like the Nikon Z9/Z8 — but credentials are stripped by screenshots and many platform uploads, and absence proves nothing.
• SynthID (Google DeepMind) embeds imperceptible watermarks in generated media; watermarking requires generator-side cooperation and weakens under transformation — structurally different from statistical detection.
• OpenAI retired its text classifier in July 2023 at a 26% true-positive rate — the clearest public marker of how much harder inference is on text than on in-distribution images.

Sources

1. GenImage: "A Million-Scale Benchmark for Detecting AI-Generated Image," NeurIPS 2023 Datasets and Benchmarks. openreview.net/pdf?id=GF84C0z45H
2. Coalition for Content Provenance and Authenticity (C2PA) — c2pa.org and spec.c2pa.org (steering-committee membership, Content Credentials, specification 2.3+; fetched at time of writing).
3. OpenAI. "New AI classifier for indicating AI-written text," January 31, 2023; retirement update July 20, 2023. openai.com
4. Google DeepMind — SynthID documentation (watermarking for AI-generated media).
5. Weber-Wulff, D., et al. "Testing of detection tools for AI-generated text." International Journal for Educational Integrity 19, 26 (2023).
6. Coalition for Content Provenance and Authenticity — specification and adopter list, c2pa.org.