DeepFake Check
Back to Blog
DeepFakeCheck Team 8 min read

Best Free AI Image & Deepfake Detectors (Tested Oct 2026)

What was actually tested

This is a spot check of four images, not a leaderboard. On 27 October 2026 the same four files were scored for Hive, AI or Not, Sightengine, Is It AI, and TruthScan. Three files were labeled AI: canton-tower, a screenshot of a Google Gemini image; zootopia, a GPT image; and camera-woman, a Flux image. The fourth, man-portrait, was unlabeled. The notes call it likely real. It is not counted below as a confirmed photograph, and a tool that called it AI is not counted as a measured false positive.

An older article on this site, what each deepfake tool checks, did not run a shared benchmark. This one does, at this small size. It does not crown a winner.

A separate pixel-model bake-off, 477 images on 28 October 2026, is on the accuracy page. That test did not include these vendors, and it is not a full-system score for DeepFakeCheck. Full-system accuracy testing is still in progress.

How each tool did on the three labeled images

Generator name, when the tool offered one:

  • Hive matched the generator on 3 of 3: gemini3 at 99.7%, gptimage2 at 99.9%, and flux at 89.9%. On the unlabeled portrait it returned AI 6.5%, recorded in the scoreboard as a real call.
  • AI or Not matched 2 of 3. The GPT image was 4o at 98%. The Flux image was Flux at 86%. The Google Gemini screenshot was called AI at 86%, but the family was Stable Diffusion 47% / Nano Banana 19%, which did not match the label. The unlabeled portrait was called real, with a deepfake score of 15%.
  • Sightengine matched 1 of 3. The GPT image was GPT at 99%. The screenshot was called Wan at 76%. The Flux image was Stable Diffusion 82% / Flux 18%. The unlabeled portrait was AI 3%.
  • Is It AI matched 1 of 3. The GPT image was ChatGPT 4o at 99%. The screenshot was Wanx at 78%. The Flux image was Stable Diffusion at 84%. The unlabeled portrait was human at 97%.
  • TruthScan returned AI 99%, AI 99%, and AI 97% on the three labeled images and did not name a generator. It also returned AI 75% on the unlabeled portrait. The scoreboard marked 75% as an outlier. Without a confirmed label, that call stays out of the false-positive count.

deepfakecheck.io's own result on this set will be added after a re-test on the current production version.

On generator names, Hive was ahead in this set. On a plain AI-or-not call, TruthScan flagged all three labeled AI images, and so did the others except where a label was simply the wrong family. Four files cannot say what happens on the next generator, on a screenshot from a different app, or on a real camera roll. Treat a single wrong family name as a real failure mode: a confident "Wan" on a Gemini screenshot is not a mild miss.

Prices read on 29 October 2026

Prices below are what each vendor's own page said that day. They are not a quote for tomorrow. Where a page did not show a number we could read, the number is omitted.

DeepFakeCheck, from the product code rather than a scraped page: a new website visitor starts with 2 analyses. Anonymous successful free analyses are also capped at 2 per IP per UTC day. Paid packs are one-time, not a subscription: $2.99 for 50, $7.99 for 200, and $14.99 for 500. Paid analyses last 1 year. The pricing page is the public list.

Hive (pricing page): the fetch was titled "Pricing | Deploy". It lists AI Image + Deepfake Classification and AI Video Detection on a usage-based Developers plan, and a custom Enterprise plan. A per-image price for those two products was not in the text we could read, so none is repeated here. The page did not state a no-signup consumer allowance.

AI or Not (pricing page): the free plan lists $5 in free credits, 1 million words of AI text detection free every month, 20 AI image checks, and an API key. Pro is $5/month with $10 in credits, and the page says it unlocks text, image, deepfake, voice, music, and video, listing 2 million words, 500 image checks, and 1,000 seconds of video. A $100/month plan includes $100 in credits and, in their words, 25% less on image, audio, and video checks than Pro.

Sightengine (pricing page): free is 2,000 operations a month, at most 500 a day, including AI image detection. Their comparison table does not include AI video detection on the free plan. Their FAQ says the free plan has no time limit. Starter is $29/month for 10,000 operations, then $0.002 per extra operation, and includes AI image detection, AI video detection, and deepfake detection. The next listed plan is $99/month for 40,000 operations, with the same overage.

Is It AI (plans page): the headline says 3 free checks without signup, plans from $1.99/month, and credit packs from $0.99. The same page also says the free plan's 5 credits a month cover images only. Both sentences were on the page. We did not ask them which allowance a new visitor actually receives. An image check is 1 credit. A video is 6 credits per started 10 seconds. Bought credits last 90 days. Paid-plan credit counts were not in the text we saved, so they are omitted.

TruthScan (pricing page): free is 25 results a month, where one result is one image or one PDF page. Their FAQ says that needs no credit card and has no time limit. Starter is $24/month for 1,000 results, then $0.03 each. Professional is $83/month for 5,000. Business is $333/month for 40,000. A later answer on the same page says 40,000 Business results total $399. Both figures were visible. We did not ask which amount is billed.

TruthScan's own page also says, vendor-reported, that average detection accuracy is 99.3% across 92 image generators and 250,000 real images, with a false-positive rate under 1%. We did not run that study. It is not comparable to the four images above, and it is not our number.

When to pick which

  • Pick Hive when you need an API and a generator name, and you can accept that this recommendation rests on three labeled images.
  • Pick AI or Not when you want a free image allowance plus a large text allowance, and you will check a generator guess instead of trusting it. The screenshot's family was wrong here.
  • Pick Sightengine when you are already sending images to a moderation API and the free tier is enough to try AI image detection. Do not expect AI video detection on that free tier. The family was wrong on two of three labeled images.
  • Pick Is It AI when you want a small no-signup web check and you have read both free-allowance sentences on their plans page. The family was wrong on two of three labeled images.
  • Pick TruthScan when the job is images or PDF pages, a monthly free pool of 25 results fits, and you will label their 99.3% figure as theirs. They did not name a generator on this set.
  • Pick DeepFakeCheck when you want image, video, audio, and text in one no-signup check, a one-time pack instead of a subscription, and a written method. deepfakecheck.io's own result on this set will be added after a re-test on the current production version. The pixel-model AUC of 0.953 is a different test, on 477 images.

Longer notes: Hive, AI or Not, Sightengine, and TruthScan.

What would change this page

A larger labeled set, run through every product on the same day, including confirmed real photos. The pixel-model numbers also need a new run after the 29 October 2026 XMP parsing fix. Until that rerun, the accuracy page is explicit that its figures predate the fix and that no full-system rate exists.

Sources

  • Internal scoreboard, 27 October 2026, with the production-site notes from 28 October 2026. Not a public dataset download.
  • Pixel bake-off and operating points, 28 October 2026, linked from the accuracy page.
  • Hive pricing, accessed 29 October 2026: https://thehive.ai/pricing
  • AI or Not pricing, accessed 29 October 2026: https://www.aiornot.com/pricing
  • Sightengine pricing, accessed 29 October 2026: https://sightengine.com/pricing
  • Is It AI plans, accessed 29 October 2026: https://isitai.com/plans
  • TruthScan pricing, accessed 29 October 2026: https://truthscan.com/pricing