Deepfake detector with probability scores

If you want a deepfake detector that gives you a probability or confidence score instead of just “real” or “fake,” look for tools that report a numerical likelihood, authenticity score, or frame-by-frame manipulation score.

One important distinction: a tool might report “87% AI-generated” or “0.87 detection score.” Those aren't necessarily the same as an 87% probability that the media is fake. The score may be an uncalibrated model output.

Best options I found

• 

1. Resemble AI Detect

(Best for audio and video)

Its API returns numerical scores, including an aggregated score, confidence values, and frame- or segment-level results. Useful if you want to inspect how strongly different parts of a video or audio clip appear manipulated. The API also returns labels, so it isn't strictly score-only.docs.resemble.ai+1

Website:Resemble AI documentation

---

• 

2. Witness

Accepts images and videos and provides a numerical confidence score, heatmap, and explanation alongside its verdict. A good candidate if you want a straightforward web-based tool rather than building an API integration.witness.vision

Website:witness.vision

---

• 

3. Sensity AI

An API for images, video, and audio that provides confidence scores from individual detection services and forensic explanations. However, its overall result is explicitly aggregated into a binary suspicious or valid verdict, so it is less suitable if you want to avoid binary outputs entirely.docs.sensity.ai

Website:Sensity documentation

---

• 

4. DeepfakeDetector.ai

Displays numerical confidence values and offers an inconclusive outcome when evidence is weak. Its example reports include scores for voice characteristics and frame consistency, which is useful if you want more nuance than a forced real/fake answer. Verify its scoring methodology before relying on it.deepfakedetector.ai

Website:deepfakedetector.ai

Which one should you choose?

My recommendation: Start with Resemble AI if you're examining audio or video. If you need a tool that never forces a yes/no classification, confirm that its interface or API lets you access the raw numerical score independently of its verdict.

One caution: confidence scores aren't automatically calibrated probabilities. Research published in Scientific Reports in 2026 discusses this distinction and the need for better-calibrated evidence measures in deepfake forensics.www.nature.com

One question so I can narrow this down: What type of media are you trying to detect?

1. Media type

Images / photos

Videos / face swaps

Audio / cloned voices

All media types

---

2. Preferred access

Website where I upload files

API for an application or workflow

Local software that keeps files on my computer

---

3. What score do you want?

Estimated probability of being fake (0–100%)

A continuous score without a forced verdict

Either is acceptable if the score is useful

[Find the best match]

When looking for a deepfake detector that bypasses rigid binary "yes/no" classifications in favor of continuous probability or confidence scores (e.g., a 0−1 0 0% scale or normalized 0−1 float), enterprise-grade forensic platforms are your best option. Continuous scoring allows you to set custom threshold boundaries depending on your risk tolerance (e.g., routing anything above 7 0% to rejection, and 4 0%−6 9% to human review).docs.realitydefender.com+2

The leading options that provide granular probability scores and metadata/visual explainability include:

1. Reality Defender (RealAPI / RealScan)
   • How it scores: Provides a granular manipulation probability score (graded from 1% to 9 9%) rather than a simple binary flag.
   • Media support: Multimodal (deepfake video, audio, image, and text/document analysis).
   • Output format: Structured JSON via REST or SDKs (Python, TypeScript, Go, Rust, Java) containing normalized confidence ratings, heatmaps, and model-specific breakout signals.realitydefender.com+2
2. DuckDuckGoose AI
   • How it scores: Returns precise percentage-based confidence scores (e.g., 9 5.7 9% probability of being a deepfake) alongside anomaly indicators.
   • Media support: Specializes in biometric verification, face-swaps, liveness spoofs, and image/audio/video analysis.
   • Output format: Machine-readable API outputs and visual traces designed to feed directly into downstream risk or KYC orchestration engines rather than standalone hard blocks.duckduckgoose.ai+2
3. Sensity AI
   • How it scores: Delivers deep forensic breakdowns with specific percentage confidence metrics per vector (e.g., 8 9.0% face-swap confidence, pixel-level heatmaps).
   • Media support: Video, image, and audio frame-by-frame analysis.
   • Output format: Comprehensive forensic intelligence reports and API integrations suitable for fraud tracking and platform monitoring.sensity.ai+1
4. Deepfake Detector
   • How it scores: Returns an AI confidence score and a graduated trust/verdict score (e.g., Authentic vs. Likely Synthetic).
   • Media support: Voice clones, images, and video manipulation (lip-sync and face swaps).
   • Output format: Clean JSON payloads via a single POST request.deepfakedetector.ai+2

If you can share whether you need this via an API for developer integration , a web dashboard for manual uploads , or the specific media type (audio, video, or images) you plan to analyze most, I can help point you toward the most cost-effective tier or SDK.

Yes. Several deepfake detectors provide a probability or confidence score rather than just a binary “real/fake” result. Some report a percentage from 0–100%, while others provide a confidence score alongside an overall verdict or a breakdown of suspicious signals.

The important distinction is that a detector's score is not necessarily a statistically calibrated probability that the content is fake. Here are some options worth evaluating.

Deepfake detectors that provide scores

• 

1. WeVerify Deepfake Detector

(Probability score)

An image and video detector that explicitly returns a probability of deepfake manipulation. For videos, it can calculate scores for individual frames or shots and an overall video-level score.

Best for: image/video verification and research use.

Access: browser demo and API access options.

Website and documentation ·Try the demo

Source: WeVerify project documentation.weverify.eu

---

• 

2. Imagera AI

(0–100% probability)

Provides a probability score, a confidence figure, and explanatory signals. Its separate detection tools cover face swaps, images, generated video, and synthetic audio.

Best for: someone who wants a straightforward percentage with an explanation.

Access: browser-based, pay-per-scan credits.

Face-swap detector ·Video detector

Source: Imagera's published product specifications.imagera.ai+2

---

• 

3. Diopter AI

(Score plus evidence)

Analyzes images, video, and audio, returning a detection score, confidence level, verdict, and information about suspicious regions.

Best for: multimodal analysis when you want more than one summary label.

Access: free to try, with additional features available.

Try Diopter AI

Source: Diopter's product page.diopter.ai

---

• 

4. VERITAS Vision Engine

Returns an AI probability score, confidence level, and a breakdown of sub-signals. Users can select a video segment for analysis.

Best for: examining specific portions of a video, including face swaps and fully AI-generated footage.

Try the video detector

Source: VERITAS product documentation.www.veritas-engine.com

---

• 

5. AIVERIFAI MediaTruth

Combines results from multiple detection services into a weighted confidence score out of 100. Supports images, video, and audio.

Best for: comparing or aggregating the outputs of multiple detectors rather than relying on a single model.

Open MediaTruth

Source: MediaTruth's published interface.www.aiverifai.com

Which one should you choose?

One important caveat: probability vs. confidence

These terms can mean different things:

• Deepfake probability: 85% — the model's estimated likelihood that the content is manipulated or synthetic.
• Confidence: 85% — how strongly the model supports its classification, which is not necessarily an 85% chance that the classification is correct.
• Calibrated probability — a stronger statistical claim: across many comparable cases assigned a 70% probability, approximately 70% should actually be fake.

If you specifically need reliable probability estimates for research, forensic review, or a downstream decision-making system, ask whether the detector has been independently evaluated for probability calibration, including on unfamiliar generators and compressed or edited media.

My recommendation: Start with WeVerify if your focus is manipulated faces in images and videos, or Imagera if you want a simple percentage-based interface across several types of content. If you need an API or scientifically defensible probability estimates, evaluate the underlying scoring method and calibration evidence before choosing.