How to Tell If Text or Images Were Made by AI (2026 Guide)

AI-generated content is everywhere in 2026 — embedded in news articles, product reviews, academic submissions, social media posts, and stock photo libraries. The problem is that most people can not tell the difference by instinct alone. Commercial detectors are unreliable, visual artefacts are increasingly subtle, and a confidently phrased paragraph can fool a careful reader just as easily as it fools a machine. How to tell if text or images were made by AI is no longer a niche concern: it is a core media-literacy skill for editors, educators, HR teams, and curious readers alike.

This guide walks you through a practical, layered approach to spotting AI-generated text and images. By the end you will know which textual tells to scan for, which image artefacts to examine closely, how to use metadata and reverse search as independent provenance signals, and — critically — where all these methods break down so you never over-trust any single test. The AI-likeness checker at humanizily.com is a useful starting tool: it scores how machine-like a passage reads and highlights the specific phrases dragging the score up — all free for 2,000 words with no card required.

What you’ll need

• humanizily.com — AI-likeness checker, 2,000 words free for 15 days, no card required
• Google Images or TinEye for reverse image search (also use Yandex for broader coverage)
• A metadata viewer: Jimpl.com (browser) or ExifTool (command-line)
• Content Credentials verifier at contentcredentials.org/verify (free, browser-based)
• A modern browser with zoom capability (200% is your friend for image inspection)

Step 1 — Read for textual tells: the linguistic fingerprints of LLMs

Large language models are trained to be helpful, comprehensive, and safe. That training produces consistent stylistic fingerprints that are surprisingly easy to spot once you know what to look for.

Hedge stacking. Phrases like “it is important to note that,” “it is worth mentioning,” “in today’s rapidly evolving landscape,” or “as we navigate this complex terrain” almost never appear in natural professional writing. Spot two or three in a single paragraph and your suspicion should sharpen considerably.

Symmetrical pros-and-cons structure. AI text often delivers “advantages and disadvantages” in tidy, equal pairs that feel balanced rather than argued. Real opinion pieces have uneven arguments — the writer actually holds a view and weights some points more heavily than others.

No specific memories. Human writers drop names, dates, places, and personal anecdotes. AI text stays abstract: “many users report” instead of “Sarah Chen, a UX researcher in Berlin, told me in March.” The absence of specificity is itself a signal.

The em-dash overload. Current GPT-4-class and later models use the em-dash — exactly like this — as a stylistic connector at a rate well above normal human prose. One or two per article is fine. Six or more in a 600-word post is a flag worth investigating.

Predictable paragraph length. AI tends to produce paragraphs of similar length, often three to five sentences each. Human writing is messier — very short paragraphs next to long ones, one-sentence punches, fragments for emphasis.

Paste any suspicious passage into the AI-likeness checker at humanizily.com to get a scored reading alongside a breakdown of which phrases pulled the score upward.

Analyse the following paragraph for signs of AI generation. List every hedge phrase, every symmetrical pros-and-cons structure, and any sentence that contains no specific factual claim. Then give an overall likelihood score from 0 to 100 that this was written by an LLM, where 100 means almost certainly AI:

[PASTE SUSPECT TEXT HERE]

Step 2 — Apply the honest-limits caveat: ESL writers and false positives

This is the most important step most guides skip entirely — and skipping it causes real harm.

AI text detectors produce false positives at a disproportionately high rate for non-native English writers. A student from Algeria, Morocco, or any country where English is a third language will naturally avoid contractions, use formal connectors (“furthermore,” “consequently”), and structure arguments symmetrically. That pattern looks “AI-like” to every detector tested to date. A Stanford study from 2023 found false-positive rates on ESL essays reaching 61% in some tools — and the accuracy has not improved meaningfully with 2025-era detectors.

Practical rule: never use a single detector score as evidence of dishonesty for a non-native English writer. Treat detector output as a prompt for a deeper conversation with the person, not as a verdict. Ask for something the text could not fabricate: a specific lived memory, an opinion about a local event, or a follow-up detail about a person mentioned in the piece.

I suspect this text may have been AI-generated, but the author writes English as a second or third language. Without using AI-detection scores, identify three things a human who actually wrote this would almost certainly know — specific local references, personal experience markers, or domain knowledge that requires lived context. Then suggest three follow-up questions I could ask the author to distinguish a genuine human account from a plausible AI fabrication.

TEXT: [PASTE HERE]

Step 3 — Spot AI image artefacts with a trained eye

Even the best image generators in 2026 leave characteristic artefacts. Train yourself to examine these zones first rather than making a snap judgement on overall composition quality.

Hands and fingers. Count every finger and check that joints bend in the right direction. Generators have improved but they still produce fused digits, extra knuckles, and fingers that end mid-air — especially when hands are partially hidden, gloved, or shown at an unusual angle.

Text inside the image. AI generators hallucinate text. Shop signs, book covers, T-shirt slogans, and name badges in AI images typically contain convincing-looking but entirely nonsensical letter sequences. Zoom in on any visible text before trusting the image.

Background repetition errors. Walls, tiles, brickwork, and fabric patterns in AI images often repeat too perfectly in the centre and then warp or break near the frame edges. The eye glosses over this at normal viewing size; it becomes obvious at 200% zoom.

Ear, jewellery, and accessory geometry. Earrings are a consistently weak point — they rarely match between left and right, or they partially merge with hair. Glasses often have one arm that disappears into the temple or face. Watch straps frequently have the wrong number of holes or connect at impossible angles.

Skin texture uniformity. Diffusion models produce unnaturally smooth, luminous skin. Real photographs show pores, slight redness, blemishes, and texture variation across the face. If the skin looks airbrushed to clinical perfection without any visible filtering, treat it as a yellow flag.

I am going to describe an image to you. Based on my description, identify the three zones most likely to contain AI generation artefacts and tell me exactly what visual irregularity to look for in each zone. Then suggest a reverse-image search query that might surface the original source if this image was drawn from a training dataset.

Image description: [DESCRIBE THE IMAGE HERE]

Step 4 — Check metadata and Content Credentials for provenance signals

Metadata is the most under-used tool in media literacy. Every genuine photo taken on a smartphone or dedicated camera carries EXIF data: the make and model of the device, GPS coordinates if location services were enabled, and the exact timestamp down to the second. AI-generated images carry none of this — or carry metadata that was manually added afterward, which is itself a suspicious signal when combined with visual artefacts.

How to check EXIF: Upload the image to Jimpl.com or run exiftool filename.jpg in a terminal. Look for the fields Make, Model, GPS Latitude, and DateTimeOriginal. An image with no camera make is not automatically AI-generated — many people strip metadata for legitimate privacy reasons — but when absent EXIF combines with the visual tells from Step 3, the case strengthens considerably.

Content Credentials (C2PA standard). Adobe, Microsoft, and a growing number of camera manufacturers now embed cryptographically signed provenance records into images at the moment of creation. Visit contentcredentials.org/verify and drag the image in. If it was created with a C2PA-compliant tool such as Adobe Firefly, you will see a signed badge confirming it was made with AI. The absence of credentials does not confirm the image is authentic — the majority of AI generators do not implement C2PA yet — but the presence of a valid credential is a reliable positive signal in the direction it points.

Step 5 — Use reverse image search as a definitive provenance check

Reverse image search does two things simultaneously: it finds the origin of a genuine photograph that someone is misrepresenting, and it exposes AI images by surfacing their appearance on prompt-sharing communities like Civitai, Midjourney’s member gallery, or Leonardo.ai.

Upload the image to Google Images (click the camera icon at images.google.com), TinEye, and Yandex Images. Use all three: Google favours indexed web pages, TinEye tracks exact pixel matches across time, and Yandex has a strong index of social media and Eastern European platforms. If the same image — or a near-identical variant with slight seed changes — appears on a Midjourney showcase page or a Stable Diffusion community post, you have definitive provenance.

The DETECT Framework

Here is the complete method condensed into a six-step checklist you can run through in under five minutes for any suspicious piece of content:

Best AI tools for media literacy and detection

Common mistakes to avoid

• Treating any single detector score as proof. No text detector achieves reliability high enough to serve as standalone evidence of AI generation. Always combine at least two independent signals before drawing a conclusion.
• Ignoring the ESL false-positive problem. Flagging a non-native English writer as dishonest on the basis of a detector score alone is both statistically likely to be wrong and genuinely harmful to real people. Add context before acting.
• Only checking hands in images. Generators have improved dramatically. A convincing hand does not mean the image is real — also check text, jewellery, background repetition, and skin texture uniformity.
• Forgetting that EXIF can be stripped or injected. Missing EXIF is not proof of AI generation. Present EXIF data can be added manually after the fact. Use it as a corroborating signal, not a definitive one.
• Assuming detectors remain accurate over time. AI models are updated constantly. A tool trained on GPT-4 outputs struggles with Claude 3 or Gemini 2.5 prose patterns. Benchmark your chosen tools against known outputs every few months.

Pro tips & power moves

Stack your signals — three beats one every time. The strongest conclusion comes from three independent indicators pointing the same direction: a high AI-likeness score from humanizily.com, an absent camera model in EXIF, and a reverse-image hit on a prompt-sharing forum. Any one of those in isolation is weak. Three converging is compelling. Treat each signal as one vote, not a verdict.

Read the last paragraph first. LLMs reliably front-load their best content and trail off into vague, summarising language. If the closing paragraph of a blog post reads like a generic wrap-up (“In conclusion, as we have seen…”) with no new information or specific recommendation, that is more suspicious than the polished opening lines.

Know the writing styles of the major models. ChatGPT tends toward numbered lists and em-dashes. Claude uses longer, more nested sentences with careful hedging. Gemini often produces very clean parallel structure. Comparing a piece against the stylistic fingerprint of the specific model it might have come from is a faster triage tool than a generic detector. Explore the differences yourself — you can get both ChatGPT and Claude through clickdz.ai, which provides official subscriptions payable in Algerian dinar (DZD) via CIB, EDAHABIA, or BaridiMob, with instant activation. See also this detailed ChatGPT vs Claude 2026 comparison for a side-by-side stylistic breakdown.

Ask for something the model cannot fabricate. If you can speak to the author, ask a question requiring lived experience: “What was the room like when you interviewed them?” or “Which version of the software were you running when that happened?” AI-polished content rarely has consistent answers to experiential follow-up questions.

Zoom every image to 200% before judging. Artefacts that disappear entirely at normal viewing size — a slightly wrong tooth, a misaligned wall tile, a shadow that goes the wrong direction — become obvious at 200%. Most people share images at display resolution and most viewers never zoom. You are not most viewers.

You are a forensic media analyst. I am going to give you a text passage. Your job is to:
1. Identify every phrase that matches typical LLM output patterns, explaining why each matches.
2. Identify every phrase that sounds like a human with genuine knowledge or lived experience.
3. Rate the passage on a 0–10 scale where 0 = almost certainly human and 10 = almost certainly AI.
4. Explain your rating in 2–3 sentences, naming the two most decisive signals.

Passage: [PASTE HERE]

FAQ

Are AI text detectors reliable enough to use in academic or HR settings?

Not as standalone evidence. The best commercial detectors operate at roughly 75–85% accuracy on fluent native-English text, but false-positive rates spike sharply for ESL writers. Academic institutions and HR teams that use them responsibly treat a high score as a trigger for a direct conversation with the person — never as a finding of misconduct in itself. If your institution has an AI policy, check whether it specifies how detectors should and should not be used.

Can someone defeat these detection methods deliberately?

Yes, with effort. A determined person can rewrite AI output manually, strip metadata before uploading, and avoid letting AI images get indexed. But the vast majority of AI-generated content is produced at volume and published quickly — it is not carefully laundered. The DETECT framework reliably catches most casual cases. Sophisticated deliberate forgeries require a more investigative, multi-source approach beyond what any automated tool can currently provide.

Do all AI image generators leave visible artefacts in 2026?

Decreasingly so. Midjourney v7, FLUX, and the latest Stable Diffusion checkpoints produce images where the classic tells from 2022 — badly mangled hands, obvious background warping — are far less pronounced. The field is moving quickly enough that specific artefact advice ages within months. That is why the provenance-based signals — metadata, C2PA credentials, reverse search — are the most durable part of the DETECT framework: they are independent of how good the generator has become.

Your action checklist

• ✅ Bookmark humanizily.com and paste any suspicious text through the AI-likeness checker before drawing conclusions
• ✅ Check the author’s background before applying detector output to a non-native English writer
• ✅ For any suspicious image, zoom to 200% and specifically examine: hands/fingers, any visible text, background patterns, jewellery, and skin texture
• ✅ Run the image through Jimpl.com to check for EXIF camera data — missing Make/Model plus visual artefacts strengthens the case
• ✅ Drag the image into contentcredentials.org/verify to check for C2PA provenance badges
• ✅ Run the image through Google Images, TinEye, and Yandex Images — a hit on a prompt-sharing community is the strongest single signal
• ✅ Apply the full DETECT framework only after gathering at least three independent signals — never reach a conclusion from a single test

Conclusion

Spotting AI-generated content is no longer a question of running one detector and trusting the number. It requires layered evidence: textual pattern analysis (with an honest acknowledgment of its limits for ESL writers), image artefact inspection across multiple zones, metadata verification, and reverse-image search used in parallel across three platforms. The DETECT framework above gives you a repeatable, five-minute checklist you can apply to any piece of content.

The honest truth is that no single method is perfect and the gap between AI output and human output is narrowing with every model update. The most durable skill is not knowing which specific tells to look for — those change quarterly — but knowing how to weigh multiple weak signals into a defensible conclusion. To understand the models generating the content you are analysing, the top AI models in 2026 overview is an excellent starting point for mapping the landscape. And if your own writing is getting flagged by detectors, humanizily.com will show you exactly which phrases to change — any language, any niche, 2,000 words free to try.