Deepfakes & AI Fraud
Why fabricated reality is becoming the greatest threat to democracy, business, and society — and how to defend against it.
Deepfakes are AI-generated forgeries of video, audio, and images — and they're growing exponentially: from 500,000 files (2023) to 8M (2025). AI-enabled fraud causes billions in damage ($1.1B in the US alone in 2025), 98% of all deepfake videos are non-consensual pornography, and voice cloning has reached the "indistinguishable" threshold. Detection is possible (Sensity AI: 95–98%), but the attackers' lead is growing. Defense requires a combination of technology (C2PA, SynthID), regulation (EU AI Act, TAKE IT DOWN Act), and media literacy.
The new reality crisis
In January 2024, a finance employee at British engineering firm Arup transferred $25 million to fraudsters — after a video call where the CFO and colleagues were deepfake impersonations. 15 transfers to 5 accounts. No system was breached, no firewall bypassed. Just a human who trusted their own eyes.
This case is not an outlier. It's the new normal. Deepfakes — AI-generated forgeries of video, audio, and images — are growing exponentially, simultaneously threatening democracy, the economy, and individual safety. The technology that cost millions and took days yesterday now produces forgeries in seconds that humans can no longer distinguish from reality.
What research shows
increase in deepfake files in two years: from 500,000 (2023) to 8 million (2025). In Q2 2025, Resemble AI tracked 487 discrete deepfake incidents — a 312% increase year-over-year. Voice cloning has crossed the "indistinguishable" threshold: a few seconds of audio is all it takes for a convincing clone.
Geopolitical weapon: Deepfakes as instruments of disinformation
The most current example: The US-Israeli strikes on Iran starting February 28, 2026 triggered an unprecedented disinformation wave. AI-generated videos of the USS Abraham Lincoln supposedly "sinking" reached 8 million views. Manipulated images showed Khamenei buried under rubble — one still carried a visible "Meta AI" label. BBC Verify identified the three most popular AI-fabricated videos with a combined 100+ million views in the first week.
This is not a new phenomenon. During the 2024 US election, the Treasury sanctioned an IRGC sub-organization and a Moscow-based GRU affiliate — both deployed AI deepfakes for election manipulation. In January 2024, 25,000 New Hampshire voters received an AI-generated Biden robocall: cost $1, creation time 20 minutes. The FCC levied a $6M fine; the perpetrator faces 26 criminal counts. Since 2021, 38 countries have been affected by election deepfakes — impacting 3.8 billion people.
The 2026 Iran war as a blueprint: both sides deploy AI-generated "evidence" of war crimes, fabricated satellite imagery of destroyed military bases, and manipulated news footage. Pro-Iran accounts claimed retaliatory strikes "devastated Tel Aviv" — entirely fabricated. Video game footage (ARMA) was misrepresented as real combat material on TikTok. Google's AI Overviews repeated unverified claims from the conflict. The WEF ranks disinformation as the #1 global risk for the second consecutive year.
What research shows
of people can reliably identify all deepfakes (iProov study, 2,000 participants). The detection rate for high-quality video deepfakes is just 24.5% — fewer than one in four spot the fake. 60% of people believe they can detect deepfakes — a dangerous overconfidence. Meanwhile, AI detection tools lose 45–50% of their effectiveness in real-world conditions. The arms race is in full swing.
Billions in damages: AI fraud in the enterprise
The Arup case was just the beginning. In March 2025, a finance director in Singapore authorized $499,000 during a Zoom call — where no one else on the call was real. CEO fraud via deepfake now hits at least 400 companies daily. Average damage: $500,000 per incident, $680,000 for large enterprises.
What research shows
in deepfake-related fraud losses in the US alone in 2025 — tripling from $360M the prior year. Global losses exceeded $200M in Q1 2025, reaching $347M by Q2. Deloitte forecasts: AI-enabled fraud will grow from $12B (2023) to $40B (2027).
Voice cloning is the most dangerous attack vector: cheap, fast, and convincing. 77% of victims targeted by a voice clone who confirmed financial impact actually lost money. A 3-second audio snippet is enough for a convincing clone. Fortune's December 2025 headline: "2026 will be the year you get fooled by a deepfake."
Social media & youth: The most vulnerable target
Perhaps the most disturbing dimension: 98% of all deepfake videos online are non-consensual pornography — almost exclusively targeting women (99% of victims are female). Production surged 464% in 2023. According to a Thorn study, 1 in 17 young people (6%) have been targeted by deepfake nude creation. In January 2024, sexually explicit AI images of Taylor Swift reached 47 million views on X before the platform blocked "Taylor Swift" as a search term.
The numbers for children are alarming: AI-generated child sexual abuse material reported to NCMEC surged from 4,700 (2023) to 440,000 in H1 2025 alone. A Kentucky teenager took their own life after being blackmailed with an AI-generated nude image. UNICEF declared: "Deepfake abuse is abuse." 13% of US school principals reported deepfake bullying incidents (22% at high schools). The Lancet Psychiatry classifies deepfake victimization as a new category of "digital trauma."
What research shows
of US teenagers use generative AI tools (as of 2025). 34% use AI image generators, 22% use video generators. Europol estimates: by 2026, 90% of online content may be synthetically generated. Youth exposure to AI-generated misinformation and manipulated content is no longer avoidable — it's everyday reality.
Detection: The state of the art
The good news: deepfake detection is a rapidly growing market — from $5.5B (2023) to a projected $15.7B (2026), with 42% annual growth. The bad news: it's an arms race. According to Gartner (September 2025, 302 cybersecurity leaders), 62% of organizations experienced at least one deepfake attack in the past 12 months — 43% via audio calls, 37% via video calls. 80% of companies have no established deepfake response protocol.
Sensity AI
Forensic deepfake detection with 95–98% accuracy. Monitors 9,000+ sources for malicious deepfake activity. All-in-one platform for image, video, and audio analysis. Used by government agencies, financial institutions, and cybersecurity firms.
Reality Defender
Enterprise platform for real-time deepfake detection. Covers audio, video, image, and text-based deepfakes. Direct integration into content management systems. Batch analysis and real-time detection pipelines.
Google SynthID
Invisible watermarking from DeepMind, embedded in AI-generated content. Integrated into Google's models (Gemini, Imagen). Resistant to previous watermark removal techniques. Built on the C2PA standard.
DeepSafe (Open Source)
Modular, containerized platform for deepfake detection. Aggregates state-of-the-art models in an ensemble approach. PyTorch + EfficientNet-based. Enterprise-grade accuracy without licensing costs.
Regulation: The legal framework
Legislation is trying to keep pace with the technology — with limited success. The key developments:
EU AI Act: Enforcement August 2026. Article 50 mandates machine-readable disclosure of AI-generated content. Penalties: up to €35M or 7% of global annual turnover. | TAKE IT DOWN Act (US, May 2025): First US federal law — criminalizes non-consensual deepfakes, platforms must remove within 48 hours. | DEFIANCE Act (US, Jan. 2026): Passed Senate unanimously — victims can sue for up to $250,000 in damages. | NO FAKES Act (US, April 2025): Federal right to one's own voice and likeness, not yet enacted. | China: Mandatory AI labeling since September 2025, algorithm registration, platform monitoring. 46 US states now have deepfake laws (as of Feb. 2026).
What research shows
Content Credentials is the most promising technical approach: an open standard for digital content provenance, backed by Adobe, Google, Microsoft, Intel, BBC, and hundreds more. C2PA embeds cryptographically verifiable metadata into media — from camera to publication. The EU references C2PA as the standard for AI Act Article 50. The challenge: voluntary adoption, and only effective if widely implemented.
Strategies: What organizations and individuals must do now
For enterprises
1. Multi-factor verification for all financial transactions — no transfer based on a single communication channel. 2. Deepfake awareness training for finance and HR teams. 3. Integration of deepfake detection tools into communication systems. 4. Code words or out-of-band verification for critical decisions. 5. Regular penetration tests with social engineering components.
For individuals & families
1. Minimize publicly available audio and video recordings (voice cloning needs only 3 seconds). 2. Establish family passwords for phone verification. 3. Actively train media literacy — especially with children and teenagers. 4. For suspicious calls or videos: hang up, call back through a known channel. 5. Report and document non-consensual deepfake content immediately.
For society
The long-term solution is not purely technical. It requires: comprehensive media literacy in schools from elementary level, mandatory labeling of all AI-generated content (the EU AI Act is a start), criminal consequences for malicious deepfake use, and a culture of healthy skepticism — without descending into paranoia.
What research shows
is the projected global market for deepfake detection solutions by 2026 — growing at 42% annually. The market is responding. But technology alone isn't enough. The most effective defense combines detection tools, regulation, and media literacy. Organizations investing in all three layers today will be the least vulnerable tomorrow.
Our approach at Radical Innovators
Deepfakes are not an abstract future scenario — they are an operational risk in the here and now. At Radical Innovators, we help organizations defend on three levels: technology (integration of detection tools and content authentication), processes (redesign of approval and communication workflows), and people (awareness programs that go beyond standard phishing training). Our network includes specialists in AI security, forensics, and crisis management.
The greatest vulnerability isn't the technology — it's the blind trust in our senses. In a world where any video, any voice, any image can be fabricated, critical thinking becomes the most essential core competency. Organizations that don't prepare their teams today will pay the price tomorrow.
— Martin Kocijaz, CEO Radical Innovators