You don't need the maths to defend yourself#
Generative AI — text models like GPT, image models like Midjourney, voice models like ElevenLabs, video models like Sora — has become a fraud tool. To defend yourself, you don't need to understand the architecture. You need a working mental model of what the technology can and cannot do.
Here's that model in three claims.
Claim 1 — These systems are pattern reproducers, not understanders#
Generative models are trained on enormous datasets and learn statistical patterns of how text, images, voice, or video go together. At runtime, they sample new instances that resemble the training distribution.
This matters because:
- They can produce highly plausible imitation without understanding what they're producing.
- They are very good at imitating styles you've seen many examples of (English writing, a celebrity's voice from a podcast, a specific generic photo style).
- They are noticeably worse at things rare in their training data — niche dialects, internal company jargon, unusual lighting in a video.
Claim 2 — The cost of personalization collapsed to near zero#
Five years ago, a custom phishing email took a human attacker minutes to write. A voice clone took weeks of skilled work. A convincing fake video required a Hollywood team.
Today, each of those takes seconds to generate from a stock model. Personalization at scale is now cheaper than spam used to be.
Defensive consequence: assumptions that 'this email is too well-written to be a scam' or 'I would recognize my mother's voice' are no longer reliable.
Claim 3 — Generative systems still leak in detectable ways — but the leaks shrink#
In 2023 you could spot a deepfake video by its weird hands, blurred ear edges, or inconsistent earrings. In 2024 most of those tells were patched. In 2025 the few remaining tells are subtle and inconsistent.
Defensive consequence: do not pin your defence to today's visual artifacts. They will be gone in 12 months. Pin your defence to process: verification through channels the attacker can't control.
What changed for defenders#
The net effect of generative AI on the threat landscape is:
- Volume up. Mass-produced personalized attacks against everyone.
- Quality up. No more 'badly translated' or 'awkward grammar' tells.
- Specific defences down. Visual-artifact training is a losing battle.
- Process defences unchanged. Verification-through-separate-channel works against AI exactly as well as it worked before AI.
The whole rest of this module is about building the process defences that survive the next iteration of the technology.