How a U.S. quasi-law enforcement agency uses TruthScan

Case study cover: Law enforcement use case. How a U.S. quasi-law enforcement agency investigates deepfakes with TruthScan.

At a Glance

Recently, a U.S.-based agency (org name hidden by request) acquired several TruthScan licenses.

The agency investigates suspected violations of law (some involving synthetic or manipulated media). It uses TruthScan to detect artificially generated images, videos, and audio.

As a quasi-law-enforcement agency, it must meet a high evidentiary standard. The agency acquired TruthScan licenses for its team after a month of testing its product.

The Problem

About 56% accuracy when people judge whether media is real or a deepfake. Across the 56 papers in the 2024 meta-analysis, about 44% of human calls are wrong, only 6 points better than a coin flip.
Figure 1. Pooled human deepfake-detection accuracy across 56 studies (2024 meta-analysis).

Now that deepfakes (AI images and videos) look so real, it’s becoming difficult to spot them. As a team dedicated to AI investigations, the agency realized it needed better tools to detect AI-generated content accurately. Its focus is on visual media.

Without accurate detection the agency may have struggled to pursue investigations and draw accurate conclusions. Before TruthScan, its team relied on legacy methods and human detection. A 2024 meta-analysis of 56 papers reported a pooled human deepfake-detection accuracy of approximately 56% across the studies analyzed.

The Solution

New advancements in generative artificial intelligence require new tools and methods specifically tailored to emerging AI threats.

Internal benchmarks put TruthScan’s detection accuracy at 99%. TruthScan detects AI images, audio, videos, documents, and text.

Bar chart titled “Who catches AI-generated media?”: human reviewers about 56% (2024 meta-analysis, 56 papers) and TruthScan 99% (internal benchmarks), with 50% marked as coin-flip chance.
Figure 2. Pooled human deepfake-detection accuracy (2024 meta-analysis of 56 papers) vs. TruthScan’s internal benchmarks.

After testing TruthScan for a month, the agency felt it met their investigative needs and quality standards, and then acquired multiple licenses for their team.

Timeline from first test to team-wide licenses. The problem: deepfakes outpace the eye, and legacy methods and human review fall short. 1-month trial: tested against its standards for investigative needs and quality. The decision: met with TruthScan, then acquired licenses for the team. Today: TruthScan’s findings, paired with human expertise.
Figure 3. From a one-month trial to team-wide licenses.

Now, the agency relies on TruthScan’s technical findings combined with human judgment. Some of the features that are valuable and helpful to its team are:

What the agency’s team relies on: forensic image analysis, detailed heatmaps, explained detection scores, security and compliance (SOC 2 Type II and ISO/IEC 27001:2022), multimodal detection, and dedicated account support with 24/7 coverage by arrangement.
Figure 4. The TruthScan features the agency’s team relies on.

The Features That Matter

Table of features, what each adds, and how it helps investigators: AI image analysis adds technical analysis beyond visual inspection (spot). Detailed heatmaps give more evidence to weigh than a score alone (look closer). Detection scores and explanatory reports support judgment, not a ruling on violations (prioritize). Security and compliance documentation checks controls against sensitive-material rules (evaluate). Multimodal detection covers images and video, plus audio, documents, and text (cover more). Dedicated account support offers a direct contact, with 24/7 support by arrangement (get answers).
Table 1. The features that matter, and their value for investigative work.

How the Agency Uses TruthScan

Its team uses TruthScan to analyze suspicious visual content. Although the organization’s full process is proprietary, it uses TruthScan’s deepfake detection and AI content analysis features.

When something is flagged as AI, it means the agency has to investigate it further.

With TruthScan, the agency can quickly get detection scores across hundreds of content items, which helps them focus their investigations by detection score. When something is flagged as AI, that doesn’t mean the team has made a final ruling; the organization still has to dig deeper and figure out why the content was created and by whom.

Workflow titled “A flag starts the investigation. It doesn’t end it.” TruthScan steps: 1. Analyze, scores across hundreds of items. 2. Prioritize, focus by detection score. Agency steps: 3. Investigate, why it was made and by whom. 4. Conclude, the final ruling stays with investigators. Simplified view; the agency’s full investigative process is proprietary.
Figure 5. Where TruthScan fits in the agency’s investigation workflow.

TruthScan helps investigators review suspicious content and prioritize further examination.

Outcome and Results

When investigating images, investigators usually follow a manual process that can take up to 20 minutes. The typical image investigation process consists of pulling and inspecting metadata, conducting lighting analysis, and comparing suspected AI images side by side.

TruthScan automates that process, cutting initial investigation time down. It does this by analyzing metadata, presence of watermarks, image patterns, and more all at once. The agency recently had 100 cases of suspected AI-generated content to investigate. With TruthScan, they can now investigate roughly four to five times faster.

TruthScan analyzes video content for signs of AI generation. Manual frame-by-frame review can take an investigator up to an hour depending on length; TruthScan analyzes video frames and returns a report and heatmap within minutes.

Initial investigation time. Images: up to 20 minutes manually, 4 to 5 times faster with TruthScan. Video: up to 1 hour manually, minutes with TruthScan. Each pair is scaled to its own manual baseline, and manual video time varies with length.
Figure 6. Initial investigation time for images and video, manual vs. TruthScan.

Looking Forward

TruthScan is committed to fighting AI fraud and providing the most accurate AI detection models. TruthScan deploys new detection model updates and continually trains on new AI generation sources.

As new AI fraud threats emerge, TruthScan’s continual goal is to defend individuals and organizations against them.

Frequently Asked Questions

Because of advances in AI image quality, visual inspection alone is no longer reliable.

Before purchasing licenses, the agency tested TruthScan’s detection software. After a month, they met with TruthScan and acquired licenses.

TruthScan saves investigators time by quickly performing forensic-level image analysis. It adds extra analysis layers that help achieve accurate outcomes while quickly handling previously time-consuming tasks like metadata analysis.

TruthScan offers image, document, video, audio, and text analysis. Investigators select the content type they need to analyze, then upload it or direct it to TruthScan to receive a report.

Detection scores tell you how confident the model is in finding AI content. Heatmaps show where the software found visual patterns or artifacts it flags as suspicious. Reports explain the detector’s results.

Some findings can establish AI generation or editing, such as a verified watermark or authenticated provenance from a trusted source. Other findings (e.g., model scores and unverified metadata) provide supporting indicators for further review. Establishing AI involvement does not, by itself, establish that a factual claim is false or that a law was violated.

Investigators should consider possible false positives and false negatives and assess results alongside other available evidence. TruthScan works by bolstering human expertise. Human judgment remains key in fraud investigations.

TruthScan is SOC 2 Type II and ISO 27001:2022 compliant and supports GDPR compliance. TruthScan also creates custom data handling contracts suited to each business need.

Contact

For any questions regarding this case study, please reach out to Devan Leos at devan@truthscan.com or news@truthscan.com.

For sales-related inquiries, submit a form here.