THE DARK SIDE OF AI IN NUMBERS: 2019–2025 TRENDS AND THE 2030 FORECAST

Every capability that makes AI useful — speed, scale, fluency, mimicry — becomes a weapon in the wrong hands. This report puts numbers to AI’s dark side: where it’s growing, how fast, and where the forecasts point by 2030. The through-line inverts our usual one — harm concentrates exactly where humans over-trust the output, or weaponize the tool on purpose.

Thank you for reading this post, don't forget to subscribe!

The trend: a dark side growing year over year

The clearest signal isn’t any single number — it’s the slope. Across independent trackers, AI harm is compounding at double- and triple-digit rates.

Regionally, the 2022–2023 deepfake surge was led by North America (+1,740%) and Asia-Pacific (+1,530%), with Europe up 780% (Sumsub). Deepfake fraud attempts climbed from roughly 0.1% to 6.5% of all fraud in three years, and now account for about 40% of biometric fraud attempts.

Deepfake fraud: the $25-million heist era

The defining case came in 2024, when a finance employee at engineering giant Arup wired about US$25 million after a video conference in which every other “colleague,” including the CFO, was an AI deepfake. It’s the extreme end of a now-common problem: Regula’s 2024 survey found 49% of businesses had faced an audio or video deepfake fraud attempt, with average losses near $450,000 — and over $600,000 in financial services. Seeing and hearing are no longer proof.

AI as a weapon: cybercrime gets a skill-floor cut

In mid-2023, jailbroken “dark LLMs” appeared for sale: WormGPT (June 2023) and FraudGPT (July 2023) marketed phishing, malware, and business-email-compromise tooling for €60–100/month — no skill required. Variants built on Grok and Mixtral resurfaced in 2024–25. The effect shows up in the data: an estimated 82.6% of phishing emails are now AI-written, voice phishing surged 442% across the second half of 2024, and over 80% of cyberattacks now involve AI in some form. Models themselves remain soft targets — prompt injection is OWASP’s #1 LLM risk, with research showing jailbreak success rates above 90% against unprotected systems.

Synthetic abuse: the human cost behind the numbers

The fastest-growing harm is also the gravest. The Internet Watch Foundation found AI-generated child-abuse videos jumped from 13 in 2024 to 3,440 in 2025, two-thirds rated in its most severe category. More broadly, roughly 96% of all deepfakes are nonconsensual intimate imagery, 99% of it targeting women (Sensity). Consumer scams scale the same way: voice clones now need as little as three seconds of audio, and one in four adults reports hitting an AI voice scam (McAfee).

“Harm spikes where humans over-trust or weaponize AI, and recedes where verification is put back in.”

The 2030 forecast

The counter-story: the defenses are scaling too

The same forecasts that show harm rising show the response hardening. Regulation is arriving fast: the EU AI Act’s transparency rules (Article 50) require AI content to be labeled and machine-detectable from August 2, 2026, with fines up to €35M or 7% of global turnover for the worst violations. In the US, the TAKE IT DOWN Act (signed May 2025) made nonconsensual intimate deepfakes a federal crime, and roughly 45 states now have deepfake laws on the books. Technically, content provenance (C2PA), detection AI, and — most reliably — human verification protocols like payment callbacks are the practical line of defense. The pattern holds: harm spikes where humans over-trust or weaponize AI, and recedes where verification is put back in.