Resemble AI's DETECT-World Looks Past Generator Fingerprints to Spot Deepfakes

October 14, 2026Source: Resemble AI

Resemble AI says its new DETECT-World model detects deepfakes by checking whether content makes physical sense, not only by matching artifacts of known generators. The company also launched Investigator, a set of forensic tools, and Agent Detection, a product that tells humans, bots and AI agents apart.

Key points

Resemble AI's DETECT-World checks whether content is physically coherent, in addition to looking for generator artifacts.
The company reports video accuracy of 98.2% (up from 80.96%) and image accuracy of 95.8% (up from 86.70%).
DETECT-World ranked first of 18 systems on the Podonos audio benchmark as of August 2026, per Resemble AI.
Investigator offers six forensic workflows through an API, and Agent Detection classifies visitors as human, AI agent or needing review.
Image and video figures come from internal benchmarks that the company says are pending.

A model that checks whether content makes sense

In a Q3 recap published on its blog, Resemble AI described a shift in how it approaches detection. Its previous model, DETECT-3B Omni, ranked first on DFBench and covered audio, image and video across 51 languages. Like most detection models, it relied heavily on learning artifacts left by known generators. According to the company, that becomes harder as the number of generators grows, because a detector that has never seen a new model's outputs can lose performance until it is retrained.

DETECT-World still looks for those artifacts, but it also asks whether the content is coherent. The company lists questions such as whether lighting is consistent across a face, whether shadows fall correctly relative to the light source, whether background movement makes sense from frame to frame, and whether lips match the audio. Resemble AI says these signals are not tied to a single generator, and that the physics do not change when a new generator launches.

The reported results

Resemble AI reports the following figures for DETECT-World:

Video deepfake detection accuracy rose from 80.96% to 98.2%.
Image accuracy rose from 86.70% to 95.8%.
Audio accuracy reached 99.47%, with a false positive rate of 0.66%.
Latency dropped 66% to 399 milliseconds.

The company notes that the image and video figures come from internal benchmarks, which it describes as pending. On the independent Podonos audio deepfake detection benchmark, DETECT-World ranked first out of 18 evaluated systems as of August 2026, according to Resemble AI. It was the only system on the board with both false positive and false negative rates below 1%. DETECT-3B Omni ranked second.

Resemble AI also tested the model against Haotian AI's real-time face-swap tool, which the model had not seen before. The company says it identified the generated content with approximately 95% accuracy without prior exposure to the generator, retraining or fine-tuning.

Investigator and Agent Detection

Beyond the model, Resemble AI launched two products. Investigator consists of six managed forensic investigators available through its API, each built around a specific verification workflow: social media moderation, insurance claim materials, breaking news footage, documents and receipts, forensic evidence, and identity documents and presented individuals. Each investigation combines multiple forensic analysis layers and returns evidence and explanations with the result. The company says enterprise customers repeatedly asked for more than an "is this AI or not" answer, for example a trust and safety analyst who sees a 94% manipulation probability but still needs to know what kind of manipulation occurred and whether to remove the content.

Agent Detection is described as a new product built on the company's System One model. Site owners add a single line of code and, according to Resemble AI, can see their traffic within minutes. Rather than relying on user-agent strings, which agents can spoof, it analyzes behavioral signals such as cursor movement, typing rhythm and interaction patterns. Each verdict includes reasoning for why a visitor was classified as a human, an AI agent, or something needing further review.

Operators could then reroute agents to APIs or MCP servers, restrict access to areas such as billing or credential management, or block activities such as account deletion. Head of Product Obaid Ahmed wrote that he does not think the answer is to block every AI agent, since some do useful work, and that organizations will need to set policy for routing, allowance and restriction.

Context: awards, provenance and a shift in focus

Resemble AI said it officially sunsetted voice generation as a paid product and put the full company behind detection. DETECT-World won Best Technological Solution at the Global Anti-Scam Summit America, and the company won Best Technology at Visual 1st. It also said it signed the EU Code of Practice on Transparency of AI Generated Content alongside Google, Anthropic, Meta and others, as a technology provider of marking and detection solutions. The company frames watermarking as a way to establish provenance, while detection identifies signs of manipulation.

At the GMI Summer Signal conference in San Francisco, CTO Tedi Papajorgji opened with a guessing game using a press conference video, a voice clip and a photo of a dog on a couch. All three were synthetic. Ahmed wrote that fewer than 10% of attendees guessed all three correctly, even though they knew they were judging real versus fake.

What it means for users

For organizations facing voice-clone calls or AI agents on their sites, the company's argument is that detection must hold up as new generators appear, and that results need context to be actionable. If the reported figures hold up outside Resemble AI's own testing, a model less dependent on known generators could reduce the need for constant retraining. The image and video numbers remain tied to internal benchmarks, so independent verification is still pending.