# AI or Not Review (2026): Features, Accuracy, Pricing, API, and Limitations

Explore AI or Not's image, text, video and deepfake detectors, current pricing, independent accuracy tests, API integrations, privacy and limitations.

Canonical URL: https://aiidelist.com/blog/ai-or-not-review

Language: en

Published: 2026-10-11

Updated: 2026-10-11

## Key Takeaways

- **AI or Not is a multimodal AI-content detection platform**, not just an AI image checker. Its services cover images, deepfake faces, text, video, synthetic speech, AI-generated music, and reverse image lookup.
- **Its headline 98.9% accuracy is a company-reported figure**, not a guarantee for every generator or file type. An independent NewsGuard audit published in May 2026 offers a more nuanced picture: AI or Not correctly flagged all 15 substantially AI-edited images but incorrectly flagged one of 15 authentic images.
- **Pricing starts with a free tier**, while Pro is advertised at $5 per month with $10 in monthly detection credits. Business costs $100 per month; enterprise pricing is negotiated.
- **Developers can integrate detection through a REST API or MCP server**, including workflows with Claude, Cursor, and Codex.
- **Detection should inform verification, not replace it.** A positive AI flag does not establish that an event was fabricated, and an authentic-looking score does not prove provenance.

![What Is AI or Not?](https://cdn.aiidelist.com/api/image/pPUOfq6eQzOiadZmQgEH5.webp)

## What Is AI or Not?

[AI or Not](https://www.aiornot.com/) is a web-based and API-accessible service for assessing whether media was generated or modified with artificial intelligence. It launched as an image-oriented detection tool and expanded into audio, video, music, text, and deepfake analysis. The product now positions itself primarily around trust and safety, identity fraud, content moderation, and media authenticity.

AI or Not's operating entity is AIorNot, Inc. Its founder and CEO is Anatoly Kvitnitsky. The product originated from the Optic team and later became an independent company. A founding-team public announcement disclosed a $5 million seed round, involving Foundation Capital, GTMfund, Plug and Play, and other backers. Those details make its corporate background more transparent than that of many anonymous AI-checking websites, although they do not independently validate detection accuracy.

The company displays **500,000+ users** and **7 million+ checks** on its website as of October 2026. These are company-published metrics rather than independently audited traffic or usage figures.

## AI or Not Features: What Can It Detect?

| Feature | What it does | Typical application |
| --- | --- | --- |
| AI image detection | Evaluates image pixels for synthetic-generation patterns and may report likely generator families | Journalism, image marketplaces, insurance claims |
| Deepfake image detection | Screens for AI-manipulated faces and face swaps | Identity checks, fake profiles, impersonation |
| AI text detection | Analyzes text for patterns associated with machine-generated writing | Editorial screening and educational review |
| AI video detection | Examines video frames and associated audio for generative-AI signals | News verification, social platforms, fraud prevention |
| AI voice detection | Checks voice recordings for synthesis or cloning indicators | Impersonation and social-engineering investigations |
| AI music detection | Screens music for AI-generated audio, including content from music-generation tools | Distribution review and rights workflows |
| Reverse image search / tracker | Searches a database of known AI-generated and deepfake images | Source research and repeat-content detection |
| Detection API and MCP | Allows applications and AI assistants to call detection models | Automated moderation, developer tools, internal risk workflows |

![image detection](https://cdn.aiidelist.com/api/image/v4ZjyUW0hqaUn8xjGkx0j.webp)

According to the company's [image detection page](https://www.aiornot.com/ai-image-detector), generator coverage includes Midjourney, DALL-E/GPT Image, Stable Diffusion, Flux, Adobe Firefly, Ideogram, Recraft, Nano Banana, Seedream, Qwen, and others. Coverage is a vendor claim; it should not be interpreted as equal performance across every version, editing workflow, or compression setting.

The [AI Image Tracker](https://www.aiornot.com/tracker) is a distinct free discovery tool: users can browse or search known AI-generated and manipulated imagery. A database match is useful provenance context; failure to find a match is not evidence that an image is real.

## How Does AI or Not Work?

For images, AI or Not says its detector evaluates pixel-level content rather than depending solely on file metadata, visible watermarks, or C2PA provenance credentials. That distinction matters because screenshots, crops, re-encodes, and social-media uploads can remove or weaken embedded provenance signals while leaving some statistical image features available for analysis.

For video, the company describes a multimodal approach that considers frames and accompanying audio. Its text checker compares writing characteristics against patterns associated with human and model-generated language. These are probabilistic assessments of content characteristics, not direct records of who created a file.

A reported **AI likelihood score should not be read as a calibrated probability that an image is fraudulent**. A photograph can be genuine yet lightly edited with AI; an entirely synthetic scene can appear convincing; and a generator attribution can be incorrect. The scope of the question matters: *Was any AI involved? Was the full scene synthesized? Was a person's face swapped? Is the claimed event real?* Those are separate verification tasks.

## How Accurate Is AI or Not? Company Claims vs. Independent Testing

AI or Not advertises **98.9% accuracy**. The company's description attributes this figure to its testing on public academic image datasets. The precise dataset composition, thresholds, and test conditions should be examined before applying that percentage to a different content type or real-world use case.

### What the May 2026 NewsGuard audit actually found

In an [independent audit published on May 8, 2026](https://www.newsguardtech.com/special-reports/leading-ai-image-detection-tools-mislead-online-users-often-declaring-authentic-content-fake/), NewsGuard evaluated five image detectors using 15 authentic news-related images, 15 versions with light AI-assisted edits, and 15 versions substantially altered to change their meaning. Under its 50% AI-likelihood classification threshold, the findings for AI or Not were:

- **Authentic images: 14 of 15** correctly classified as real; one false positive.
- **Substantially altered images: 15 of 15** classified as AI-generated.
- **Lightly AI-edited images: 13 of 15 (approximately 87%)** classified as AI-generated.

The same audit illustrates why comparisons must identify the target behavior:

| Detector | Authentic images correctly classified | Substantially AI-altered images flagged |
| --- | ---: | ---: |
| AI or Not | 14/15 | 15/15 |
| Hive | 15/15 | 9/15 |
| Sightengine | 15/15 | 5/15 |
| ZeroGPT | 12/15 | 14/15 |
| ScamAI | 9/15 | 12/15 |

These figures are **results from one narrow 45-image audit**, not general market-wide accuracy rankings. The trade-off is important: a model more sensitive to AI-assisted edits can catch more heavily altered evidence but can also label cosmetically enhanced authentic images as AI. NewsGuard also reported disagreements among the five tools on 35 of the 45 images.

### Additional company-reported tests

In July 2026, [AI or Not published a benchmark](https://www.prnewswire.com/news-releases/ai-or-not-detects-100-of-meta-ai-images-and-holds-98-accuracy-after-those-images-are-cropped-or-tampered-302829352.html) covering 205 Meta AI images. It reported identifying 103/103 untouched images and 100/102 cropped or tampered images. This was a **company-conducted test**; it should not be presented as an independent evaluation of AI or Not. A separate Reuters analysis discussed in that release examined Meta's native labeling under cropping, not AI or Not's comparative accuracy.

A sensible reading of the available evidence is that AI or Not is a potentially useful screening layer, especially for synthetic imagery and meaningful manipulation, but it is not a forensic proof system. Text, music, voice, and video accuracy should be evaluated separately instead of inheriting the image benchmark number.

## AI or Not Pricing in October 2026

The official [pricing page](https://www.aiornot.com/pricing) uses a flexible credit system. A subscription includes credits that can be allocated across eligible detection models; the examples for words, images, and video are **alternative ways to use the same credit balance**, not cumulative allowances.

| Plan | Listed price | What is included |
| --- | --- | --- |
| Free | $0 | $5 introductory credits, 1 million words of free AI text detection per month, 20 AI image checks, and an API key |
| Pro | $5/month | $10 in monthly credits, pay-as-you-go top-ups, and monthly credit rollover |
| Business | $100/month | $100 in monthly credits, 25% lower image/audio/video rates than Pro, and limited reseller rights |
| Enterprise | Custom quote | Custom models, private instances, on-premises deployment, model calibration, dedicated support, and full reseller rights |

### Published standard per-use rates

| Detection operation | Pro rate |
| --- | ---: |
| AI-generated image | $0.02 per image |
| Deepfake image | $0.02 per image |
| Text | $5.00 per million words |
| AI voice | $0.005 per second |
| AI music | $0.005 per second |
| AI video | $0.01 per second |
| Deepfake video | $0.01 per second |
| Reverse image search | $0.02 per image |

**Important billing caveat:** A request can invoke more than one analysis. The company's example image API request notes that AI-generation and deepfake detection are separate billable operations; NSFW and quality checks may also appear in a response. Use selective `only` or `excluding` parameters where supported and verify the actual invoice or dashboard usage rather than multiplying the image count by one listed fee blindly.

The Business plan's reseller rights are limited to **$5,000 of customer-facing consumed credits per calendar month** under the published terms; higher-volume or differently structured reselling requires a separate agreement. Subscription terms provide for automatic renewal, with no routine prorated refund for the unused portion of the current billing term.

## How to Use AI or Not for Free

1. Open the [official website](https://www.aiornot.com/) or the dedicated [free text detector](https://www.aiornot.com/ai-detection).
2. Select the appropriate content type and upload a supported file or enter text using the available interface.
3. Review the AI-generation verdict, related confidence scores, and any deepfake or generator breakdown available for that detector.
4. If the result matters, inspect the original file, seek earlier publications, check contextual details, and compare with other evidence before making a decision.
5. Create an account when an API key, tracked usage, or larger detection allowance is required.

The basic web interface lists PNG, JPG, WEBP, GIF, MP3, and MP4 among supported media formats. The text-detection page also references DOC, DOCX, and TXT. API-specific format and size limits may differ, so check the endpoint's current documentation rather than assuming all uploads share the same requirements.

## AI or Not API: Integration for Developers

AI or Not provides an [API documentation portal](https://docs.aiornot.com/) and a [Python client on GitHub](https://github.com/aiornotinc/aiornot-python). The image detector page documents the synchronous endpoint `POST https://api.aiornot.com/v2/image/sync` with Bearer-token authentication and multipart image upload.

A minimal request looks like this:

```bash
curl --request POST \
  --url https://api.aiornot.com/v2/image/sync \
  --header "Authorization: Bearer $AIORNOT_API_KEY" \
  --form "image=@example.jpg"
```

Production systems should keep tokens server-side, validate response errors, cap file size and request concurrency, store only the risk information that is necessary, and require human review for high-impact decisions. They should also track charges per enabled model and monitor for generator drift.

### MCP support for AI assistants

The [official MCP integration](https://www.aiornot.com/ai-detection-mcp) allows supported clients including Claude, Cursor, Codex, VS Code, and Windsurf to call image, text, audio, video, and batch-detection tools through an API key. This is particularly relevant for developer-facing moderation pipelines and agent-assisted investigations. MCP support does not make results more accurate; it makes the same detection service easier to invoke within a workflow.

## Data Privacy and Content Retention

The company's product pages state that submitted detection content is deleted immediately after checking. Its broader [privacy policy](https://www.aiornot.com/privacy-policy), however, separately describes collecting and retaining account, contact, IP/device, analytics, and other personal data as needed for service operations and legal requirements.

These statements are not necessarily contradictory: **short-lived uploaded media and persistent account records are different data categories**. Organizations processing confidential IDs, medical information, private recordings, or customer content should obtain contractual details about retention, regional processing, subprocessors, and audit rights rather than relying solely on a marketing phrase such as 'zero retention.' Enterprise on-premises and private deployment options may be more appropriate for highly regulated workloads.

## Who Should Use AI or Not?

**Newsrooms and fact-checkers** can use it to triage suspicious visuals, but should combine model output with original-source verification, reverse searches, publication history, and context. **Marketplace and social-app operators** can use it to prioritize review of synthetic profiles or suspicious uploads. **Insurers and KYC providers** may use it as an additional risk signal for claim photos and identity images, not as a substitute for document authentication or liveness validation. **Developers** can integrate its API with an existing moderation or anti-fraud workflow. **Students and educators** should avoid treating a text-detector score as proof of academic misconduct.

### Strengths

- Broad media coverage across images, text, video, audio, and deepfake categories.
- Public API, Python integration, and an MCP interface for developer and agent workflows.
- Clear introductory pricing with inexpensive text and image checks.
- Independent 2026 testing that showed strong performance on a small set of substantially manipulated images.
- Enterprise and reseller options for businesses that need customized deployment or product integration.

### Limitations

- The 98.9% headline does not establish accuracy on every modality, file type, or generator release.
- Real images may be falsely flagged, particularly after quality degradation or AI-assisted edits.
- A generator-family score is an inference rather than a definitive creation record.
- Independent comparisons are sensitive to sample selection, threshold, and definition of 'AI-generated.'
- Privacy assurances for uploaded content should be evaluated separately from broader personal-data policies.
- Multimodel checks and video/audio duration can make billing more complex than the entry-level subscription price suggests.

## AI or Not vs. Alternatives

For image-checking teams, [NewsGuard's comparison](https://www.newsguardtech.com/special-reports/leading-ai-image-detection-tools-mislead-online-users-often-declaring-authentic-content-fake/) provides a useful warning against choosing a detector purely by a single accuracy percentage. Hive and Sightengine made no false-positive errors on the 15 authentic samples tested, but flagged fewer heavily AI-edited samples than AI or Not. ZeroGPT and ScamAI showed different trade-offs. This comparison **does not establish their overall detection quality, full feature parity, pricing, or suitability for specific industries**.

Detection models should also be combined with provenance mechanisms such as valid publisher credentials and C2PA content records where available. A detection score asks whether content statistically resembles AI output; provenance verification asks whether a verifiable origin or edit history can be established. They answer different questions and work best as complementary signals.

## Frequently Asked Questions

### Is AI or Not free?

Yes, a limited free offering is advertised, including an introductory credit balance, monthly free text detection, some image checks, and an API key. Intensive or broader multimodal use may require paid credits or a subscription.

### Can AI or Not detect ChatGPT and Midjourney images?

The vendor lists GPT Image/DALL-E and Midjourney among its supported generator families. Detection depends on the actual output and processing history, and should not be regarded as infallible.

### Can AI or Not identify AI-written text?

It offers a text detector for writing associated with major large language models. Its output is a risk indicator and should not be used as sole evidence of authorship, cheating, or policy violations.

### Does AI or Not work on deepfake video and cloned voices?

The platform advertises video, deepfake-video, and synthetic-voice analysis. Results and supported formats should be checked for the exact API or web workflow in use.

### Is AI or Not accurate enough for a legal or financial decision?

Not on its own. Use it to flag items for review, preserve original files and audit trails, and seek stronger independent evidence before rejecting a claim, accusing someone, or taking adverse action.

### Is AI or Not a legitimate company?

It has an identified legal entity, public company leadership, published terms and privacy documents, an active product, and publicly disclosed seed financing. Legitimacy of the business should not be confused with perfection of its model predictions.

## Conclusion

**AI or Not is best understood as a practical, multi-format AI-content screening platform with unusually accessible developer integrations and a low starting price.** Its independent NewsGuard results support a useful role in detecting meaningful image manipulation, while also showing why false positives and lightly edited originals require caution. For individual users, the free tools are a reasonable starting point; for developers, the API and MCP integrations are worth evaluating with a representative, labeled dataset and clear cost controls.

To explore it, visit [AI or Not](https://www.aiornot.com/), review the [current pricing](https://www.aiornot.com/pricing), and compare its output against original-source evidence before relying on any high-stakes verdict.