The Risks of Deepfake Celebrity Content

The Numbers. Let’s Start There.

Before we talk about anything else — the numbers in the original article are a mess. “$44.5 billion in fraud losses” attributed to Deloitte actually refers to Pindrop’s contact center fraud projection, which is US-specific and covers all AI fraud, not just deepfakes. “94% of deepfakes target celebrities” gets the data backwards: 96–98% of deepfake content volume is non-consensual intimate imagery, most of it targeting ordinary women — not celebrities. Different datasets, different definitions, jumbled together.

So here’s what the verified numbers actually show:

Volume: Surfshark’s April 2025 analysis, combining the Resemble AI and AI Incident Database, found 179 tracked deepfake incidents in Q1 2025 alone — a 19% increase over all of 2024. This tracks named, documented incidents — not estimated total files. The “8 million files” figure is a separate metric covering total synthetic media volume. Celebrities were targeted 47 times in that quarter, an 81% jump over the entire prior year.

Celebrity targeting: Elon Musk leads with 20 documented incidents — 24% of all celebrity-related cases since 2017. Taylor Swift: 11. Tom Hanks: 3. Keepnet Labs reports that 48% of US deepfake incidents in 2025 used celebrity likenesses to boost scam credibility. Fraud is the dominant use case at 38% of celebrity incidents; explicit content follows at 26%.

Business losses: The Arup case — $25.6 million lost in a single deepfake video call, confirmed by the company and CNN — is the clearest data point we have for corporate exposure. Ceartas’s 2026 synthesis of financial sector data puts the average per-incident business cost at nearly $500,000 in 2024, with large enterprises reaching $680,000. These are averages from aggregated industry reports; methodology varies. Treat as directional for planning purposes.

Fraud trajectory: Deloitte projects generative AI fraud hitting $40 billion in the US by 2027, up from $12.3 billion in 2023 — a 32% CAGR. This covers all AI-enabled fraud, not deepfakes specifically. Deepfakes are a subset.

“Celebrities were targeted 47 times in Q1 2025. That’s more than all of 2024 — in three months.”

Surfshark research, April 2025 — Resemble AI + AI Incident Database

How These Scams Actually Work

The mechanics are simpler than most coverage suggests. And that’s the problem.

Celebrity scams follow a pattern. Someone takes publicly available video — interview clips, press appearances, social posts — and runs it through a generation model to produce synthetic footage of the celebrity endorsing a product. Giveaway scams dominate at roughly 30% of fraud cases; crypto investment pitches run close behind. The audience trusts the face. That’s it. That’s the whole attack.

What’s changed since 2023 isn’t the concept — it’s the accessibility. Models that required GPU clusters two years ago now run on consumer hardware. The quality threshold that fools a casual scroll is low. You don’t need to fool a forensics lab; you need to fool someone who’s half-watching TikTok at 11pm.

Corporate impersonation is more sophisticated. The Arup attack — a whole separate beast — combined spear-phishing email, real-time video deepfake of the CFO, and AI-generated versions of four or five colleagues on the same call. The employee who authorized the $25 million transfer had initially been suspicious of the email. The video call erased those doubts. That’s the second-order problem: deepfakes aren’t just tricking people, they’re specifically tricking people who were already being careful.

The Arup Case. Read This Carefully.

What makes this useful beyond the horror-story angle: the employee did several things right. Skeptical of the initial email. Verified via what appeared to be a video call with familiar faces. The attack was specifically designed to neutralize careful behavior. That’s why the lesson here is about process design, not awareness training.

The fix isn’t “train employees to spot deepfakes.” At the quality level used in the Arup attack — publicly available executive video repurposed as real-time deepfake — human detection doesn’t reliably work. The fix is: no transaction above a threshold gets processed without a second verification channel that the initiating call cannot influence. Phone the CFO on a separately-verified number. Send a Signal message. Something out-of-band.

Detection Tools: Honest Assessment

The market is real, the tools vary a lot, and the accuracy claims deserve scrutiny.

Here’s the actual state of detection in 2026: AI-generated content detection works well on known generation patterns. When the tool was trained on content from similar models, accuracy rates of 95%+ are achievable. When attackers use newer or adversarially crafted generation methods — specifically designed to evade known detectors — accuracy drops significantly. This isn’t hypothetical; it’s documented in the literature. The generation-detection arms race is real.

Human detection is worse. Studies consistently find humans correctly identify deepfakes at rates around 24–50%, with accuracy declining as generation quality improves. Don’t rely on “it looked off” as a defense strategy.

For developers wanting to prototype something before committing to a vendor, here’s a working starting point using OpenCV for blink inconsistency detection. This is a basic heuristic — not a production detector:

import cv2import numpy as np# Requires: pip install opencv-python dlib# dlib's face predictor handles eye aspect ratio (EAR)def eye_aspect_ratio(eye_points):    # EAR = (|p2-p6| + |p3-p5|) / (2 * |p1-p4|)    A = np.linalg.norm(eye_points[1] - eye_points[5])    B = np.linalg.norm(eye_points[2] - eye_points[4])    C = np.linalg.norm(eye_points[0] - eye_points[3])    return (A + B) / (2.0 * C)def flag_blink_anomalies(ear_sequence, threshold=0.25, min_blink_frames=2):    # Very few blinks OR unnaturally regular blinking = flag    blink_events = []    in_blink = False    count = 0    for ear in ear_sequence:        if ear < threshold:            if not in_blink:                in_blink = True            count += 1        else:            if in_blink and count >= min_blink_frames:                blink_events.append(count)            in_blink = False            count = 0    blink_rate = len(blink_events) / (len(ear_sequence) / 30)  # per second    # Normal blink rate: 12-20 per minute (0.2-0.33 per second)    return blink_rate < 0.1 or blink_rate > 0.6, blink_events

This is a heuristic, not a classifier. Newer deepfake models have specifically addressed blink artifacts. Use it as a signal layer in a multi-signal approach — not a standalone detector.

Who This Actually Hits

The Part That Complicates Everything

Here’s the thing nobody says clearly: most deepfake content — by volume — isn’t scams targeting businesses or celebrities. It’s non-consensual intimate imagery targeting ordinary women. Keepnet, Ceartas, programs.com — multiple sources converge on 96–98% of deepfake content being NCII The entire policy and commercial detection ecosystem is mostly oriented around the high-visibility cases: celebrity fraud, executive impersonation, political disinformation. The much larger and more personally harmful problem gets less attention because it’s harder to monetize as enterprise software.

This matters because the detection and reporting infrastructure for NCII deepfakes is significantly weaker than for fraud. If you’re a creator who finds synthetic imagery of yourself circulating without consent, your options are: report to individual platforms (slow, inconsistent), seek legal remedies (expensive, jurisdictionally fragmented), or use takedown services (works sometimes, not systematically).

A fuller treatment of that problem is beyond the scope of this piece, but any honest discussion of “the deepfake celebrity problem” that doesn’t mention it is leaving out the majority of the actual harm.

Questions People Actually Ask

Sources

• Surfshark — Deepfake Statistics 2025 (Resemble AI + AI Incident Database), Apr. 2025
• Ceartas — Global Deepfake Statistics & Impact: 2025 Report (updated Feb. 2026)
• Keepnet Labs — Deepfake Statistics & Trends 2026
• Bright Defense — 150+ Deepfake Statistics (updated Apr. 2026)
• Programs.com — Latest Deepfake Facts & Statistics (2026)
• Fortune — Arup deepfake CFO fraud, $25M, May 2024
• CNN — Arup confirms deepfake scam, May 2024
• World Economic Forum — Lessons from the Arup $25M deepfake attack (Rob Greig, CIO), Feb. 2025
• Facia AI — Top deepfake incidents 2024–2025, Oct. 2025
• SQ Magazine — Deepfake Statistics 2026
• Global Security Mag — Deepfake epidemic for the famous, Apr. 2025