How to spot AI-generated content: signals and limits

Most advice on spotting AI-generated content says look closely: count the fingers, read the background signs, notice prose that feels too smooth. That advice is ageing badly. The tells people learn are flaws, flaws get fixed, and the tools sold to catch AI text have a record of accusing the wrong people.

A better question is not "does this look AI-made?" but "where did this come from, and who stands behind it?" This explainer covers why surface signals are weak, why text detectors misfire, and a short checklist built on provenance and context. It pairs with our guide to why AI hallucinations happen, which deals with the related problem of checking what an AI tells you.

Why visual tells are weak and getting weaker

In a 2022 PNAS study, Sophie Nightingale and Hany Farid asked 315 participants to classify faces as real or AI-synthesised. Average accuracy was 48.2%, against 50% for guessing. A second group, trained on rendering artefacts and given feedback, reached 59.0% and did not improve over the session. A third rated the synthetic faces slightly more trustworthy than real ones.[1]

The authors had removed images with obvious glitches, reasoning that anyone using a fake face to deceive would do the same, and suggest that "some synthetic faces simply do not contain perceptually detectable artifacts."[1] Those faces came from a 2020-era generator; our explainer on how AI image generators work covers how the models have moved on since.

Text follows the same pattern. Stylistic tells, such as a favourite phrase or tidy three-part lists, describe one model at one moment, and a person can edit them out in minutes. Treat any oddity as a reason to check, never as a finding.

Why AI text detectors misfire

Many detectors rely on perplexity: how predictable each next word is to a language model. AI output tends to be predictable, but so is plenty of human writing.

The developer's own verdict. OpenAI released a classifier for AI-written text in January 2023. On its own challenge set it caught 26% of AI-written text while wrongly labelling human writing as AI 9% of the time, and the company warned that it "should not be used as a primary decision-making tool." In July 2023 it added a note: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy."[2] The page's list of limits still reads as a guide to detectors generally: unreliable on short texts, worse outside English and evadable by editing.[2]

Bias against non-native writers. Weixin Liang and colleagues at Stanford ran seven widely used detectors on human-written essays: 91 TOEFL essays by non-native English writers and 88 by US eighth-graders. The US essays were classified accurately; the TOEFL essays were flagged as AI at an average false-positive rate of 61.3%. At least one detector flagged 97.8% of them. When the researchers used ChatGPT to enrich the essays' vocabulary, the false-positive rate fell to 11.6%. When they asked it to rewrite AI-generated college essays in more literary language, detection rates "plummeted to near zero."[3]

The tools penalise plain writers and miss lightly polished AI text. The authors' first recommendation:

“we strongly caution against the use of GPT detectors in evaluative or educational settings, particularly when assessing the work of non-native English speakers.”

— Liang, Yuksekgonul, Mao, Wu and Zou, Patterns (2023).[3]

These are 2023 results; our comparison of AI detectors vs. Content Credentials explains what each approach can and cannot tell you.

What provenance can and cannot tell you

Provenance means the recorded history of a file. Two technologies matter here, and both are opt-in.

Content Credentials (C2PA). The Coalition for Content Provenance and Authenticity publishes an open standard for attaching signed, tamper-evident records to images, video, audio and documents. A record can state when and how the file was made, what edits followed and whether AI was used.[4] The standard's explainer is candid about limits: "provenance information alone cannot tell you whether the digital content is true, accurate or factual." Records can be stripped, and history is incomplete if a file passes through a tool that doesn't support the standard. Should you distrust media without credentials? The explainer answers: "Maybe."[4] Our explainer on what Content Credentials can show goes further.

Watermarks such as SynthID. Google DeepMind says SynthID embeds imperceptible watermarks in AI-generated images, audio, text and video from its products, and that Gemini, Search and Chrome can check uploads for one.[5] A watermark exists only if the generator added it, so finding none says nothing about other tools. For text, Google's developer documentation says the watermark is less effective on factual responses, and that "detector confidence scores can be greatly reduced when an AI-generated text is thoroughly rewritten, or translated to another language."[6]

A valid credential or watermark is useful evidence about origin. Its absence is not evidence of anything.

A four-step checklist

These steps adapt Mike Caulfield's SIFT method (Stop, Investigate the source, Find better coverage, Trace to the original context).[7] Its first move matters most: "Don't read it or share media until you know what it is."

AI Chronicle

1. Reverse image search. Search with the image or a video keyframe. Google's "About this image" shows when Google may first have seen similar versions and which pages used them earlier, exposing old photos recycled as new.[8]
2. Follow the source trail. Find the earliest post, the account behind it and what the original caption said. A screenshot of a screenshot has lost its context; go back to the first upload.
3. Check Content Credentials and metadata. Open the file in a Content Credentials viewer, or check the "how the image was made" details where a platform shows them. Google cautions that credit and source-type metadata "can be modified with image generation and editing tools, so it may not be accurate."[8]
4. Read laterally. Leave the page and search for the account, outlet or claim in new tabs. Sam Wineburg and Sarah McGrew watched 10 fact checkers, 10 PhD historians and 25 Stanford undergraduates evaluate live websites. Historians and students were often fooled by official-looking logos and domain names; fact checkers left the site quickly, checked it elsewhere and reached "more warranted conclusions in a fraction of the time."[9]

For text, steps 2 and 4 do most of the work; check the claims themselves with our method for verifying AI-generated research.

A worked example (hypothetical): a flood photo in a local group

This example is fictional. The newsroom, photo, accounts and findings were invented to illustrate the checklist. No tool was tested and no results were captured.

A social media editor at a regional news site sees a photo of floodwater around a town's clock tower, shared thousands of times overnight with the caption "This morning." The reflections look slightly odd; she treats that as a reason to check, not a verdict.

A reverse image search finds no copies older than a day. The earliest post traces to a two-week-old account with no other local content. The file has no Content Credentials, which proves nothing alone. Lateral reading settles it: the council's flood updates report no flooding near the tower, and no resident or outlet has a second angle.

Decision: don't publish it as news; ask readers for original footage instead. A real flood at a landmark would normally produce eyewitnesses and an official notice within hours.

Questions readers commonly ask

Can an AI detector prove a student used AI?

No. OpenAI withdrew its own classifier for low accuracy, and research found high false-positive rates on non-native English writing.[2][3] A score can prompt a conversation about drafts and notes; it is not evidence.

If an image has no Content Credentials, is it fake?

No. Credentials are optional and can be stripped. The C2PA explainer warns against judging trustworthiness purely on whether they are present.[4]

Can I ask a chatbot whether an image is AI-generated?

Only for a specific check. A chatbot's opinion of how an image looks shares your weakness. Google says Gemini can check uploads for a SynthID watermark, which only finds content from tools that applied it.[5]

Is it still worth learning visual tells?

As a trigger, yes. An extra finger is a good reason to start the checklist. Its absence is not a reason to stop.

The next step

The core habit is simple: judge the history, not the look. Next, read Content Credentials explained: what provenance can show, which explains how signed records work and how to read one.

Sources and evidence note

This explainer relies on peer-reviewed research, OpenAI's retired classifier page, the C2PA standard's explainer, Google and Google DeepMind documentation, and Mike Caulfield's published SIFT method. Detection figures describe the tools and generators tested at the time and are not current rates. Google's descriptions of SynthID are vendor statements, not independent evaluations. The worked example is fictional. No detector, watermark checker or search tool was tested for this article. All sources accessed 16 October 2026.

Related reading

• How AI video generators work
• AI training and copyright: the important distinctions