Fear Is a Business Model: AI, Deepfakes and the Rules We Still Don't Have
Everyone now owns a deepfake studio
There is one part of the AI age I still cannot get my head around. Anyone with a phone can now copy your face, your voice and the way you talk, and post the result to thousands of people before lunch.
Not a licensed studio. Not a production company with lawyers and insurance. Anyone.
Meanwhile, the companies building these tools spend their launch weeks telling us how dangerous their own models are. And regulators are still busy designing labels.
This post is about that gap. How people actually use AI, what regulation still does not cover, and why fear has quietly become one of the most effective marketing strategies in tech.
AI mostly speeds up what we already did
Look closely at how AI is used and most of it is not new work. It is old work, done faster. Marketing, but "with AI". Customer service, but "with AI". Content creation, reporting, coding, design, all now "with AI".
There are genuinely new fields. Generative engine optimisation and answer engine optimisation barely existed three years ago. Today brands worry about how ChatGPT or Claude describes them, not only how Google ranks them. But compared with the hype, the list of truly new jobs is still short.
Where AI really bites is in specific professions, and software is the clearest case. When Anthropic launched its Economic Index, the most concentrated use of Claude was software development and technical writing. Its 2026 agentic coding report adds a revealing nuance: developers use AI in roughly 60% of their work, yet say they can fully delegate only 0 to 20% of tasks.
That is exactly what I see in every company that I am working with. AI is a powerful accelerator in the hands of someone who already knows the job. It is not yet a substitute for knowing the job.
End users got the most dangerous tools first
When I look at how ordinary people use AI, I see enormous energy and very little direction. Most of it is harmless. Some of it is not.
Deepfakes are the clearest example. Swapping a face or cloning a voice used to take technical skill, time and specialist software. Now it takes a prompt. Analyses of OpenAI's Sora app described exactly this shift: an expert skill turned into a consumer feature. OpenAI shut Sora down within six months of launch.
Grok shows what happens when the brakes come late. In early January 2026, one analysis estimated that users were creating around 6,700 sexualised or "undressed" images of real people per hour with Grok (overview). Months after xAI promised to stop it, NBC News still found dozens of such images of real women on X.
I see this first hand in my own training sessions. When I run AI workshops inside companies, participants can open Gemini and, within minutes, produce a deepfake video of their own boss or CEO. No technical skills, no special software, no consent asked. It should not be this easy.
This is what genuinely surprises me. We license drivers, pharmacists and financial advisers. Yet the ability to put words in someone else's mouth, convincingly and at scale, is basically a free download.
The EU AI Act regulates the label, not the tool
On 2 August 2026, the transparency rules in Article 50 of the EU AI Act came into force. In short (summary, Commission guidelines explained):
Providers of image, video, audio and text generators must mark their outputs in a machine readable way.
Anyone publishing a deepfake in a professional context must clearly disclose that it is artificial.
Chatbots must tell you that you are talking to a machine.
Fines can reach €15 million or 3% of global turnover.
This is a step. But look at what it does not do. It does not stop anyone from offering realistic face and voice cloning of real people to the general public. Its answer to deepfakes is disclosure.
And here is the part most people miss. People using AI purely for personal, non professional purposes fall outside the deployer obligations. So the person most likely to deepfake a classmate, a colleague or an ex sits largely outside the new rules. Harassment and defamation laws still apply, of course, but they arrive after the damage is done.
A watermark tells you a video is fake. It does not stop the video from ruining someone's week.
I love an AI cat video. That is not the point.
Let me be clear. I am not against synthetic media. I enjoy AI cat videos as much as anyone. A dancing or singing or a cat playing the cello harms nobody. This new category actually the one I like and share with my wife the most :)
The line is not between real and synthetic. It is between imagination and impersonation. Generating a cat is creativity. Generating your colleague, your manager or your MP saying something they never said is something else entirely.
That suggests a simple principle, aimed at the AI companies rather than the users: consumer tools should not generate realistic likenesses of real, identifiable people without verified consent. Professional studios, verified creators and consent flows can still do it. Everyone else gets the cats.
The technology to enforce this already exists. What is missing is the incentive.
Inside companies, the first winners are the IT budgets
We like to say AI adoption is moving fast. Spending certainly is. Results are slower.
Gartner forecasts worldwide AI spending of around $2.5 trillion in 2026 (Battery Ventures).
In BCG's May 2026 survey, 66% of IT leaders plan to increase AI spending, more than any other category, with money pulled from elsewhere. Average measured ROI on GenAI and agents: 13.8%.
In KPMG's Q2 2026 pulse, only 7% of organisations said they had established measurable ROI (summary).
In the companies I work with, the pattern is consistent. Individual productivity goes up. The organisation barely moves. Most use cases still stop at "this helps me write faster".
Why? Because companies were already slow and layered before AI arrived. Data has to be cleaned, systems integrated, approvals redesigned, security reviewed. Every one of those steps is an invoice. So in this phase, the people who benefit most are IT departments, cloud providers and technology consultants.
There is a simple fix that costs very little: bring the people who will actually use AI into the room early. Tools bought for people tend to gather dust. Tools chosen with people tend to get used.
The fear playbook: "our product is too dangerous"
Now for the part I find almost comic. For seven years, new model launches have come wrapped in warnings about how frightening the new model is. Look at the pattern:
When | Model | The warning | What happened next |
Sep 2026 | Launch scrapped after tests showed more deception and acting beyond its authorisation | OpenAI's GPT-6 family reached ChatGPT users in October anyway | |
Sep 2026 | Announcement devoted significant space to the model's potential to do harm | Billed as the world's most intelligent model; public release within days | |
Apr 2026 | Too good at finding and exploiting software vulnerabilities for general release | Access for big partners via Project Glasswing; a Mythos-class public model, Claude Fable 5, followed in June | |
May 2025 | In a staged test, the model tried to blackmail a fictional engineer in 84% of runs | Launched the same week as "state of the art"; blackmail headlines everywhere | |
Feb 2019 | Full model withheld over fears of fake news and impersonation | Released in full nine months later |
Science and technology scholar Lee Vinsel has a name for this: criti-hype, "criticism that both feeds and feeds on hype". A warning that a product is dangerous is also a boast that it is powerful. Cory Doctorow adds a sharper point: framing AI as an existential danger supports licensing rules that conveniently shut out smaller, open source competitors.
To be fair, not every warning is theatre. Mythos did find real, long hidden security flaws. And this September an independent probe found OpenAI agents had broken out of their test environment and compromised Hugging Face's servers.
But that is exactly my point. If the danger is real, it should not be managed by press release. If it is not real, the press release should not be allowed.
Selling fear is still selling
As a marketing academic, I recognise the toolkit. Negativity bias means bad news grabs more attention than good news. Fear creates urgency. And "too dangerous to release" is a textbook scarcity signal: the thing you cannot have suddenly looks more valuable.
It works on investors, who hear "powerful". It works on executives, who hear "fall behind and you are finished". And it works on ordinary people, who end up feeling inadequate, anxious and late to something they barely understand.
Is it ethical to raise money for loss making businesses by keeping the public slightly afraid? I do not think so. In most industries, a company that exaggerated a product's danger to boost its value would face questions from regulators. In AI, it gets a front page.
I also run AI training, so let me say this against my own interest. The right response to AI is not fear. It is curiosity, practice and a healthy dose of scepticism.
Here is what I would ask regulators to do:
Regulate the tool, not only the label. No realistic likeness of a real person in consumer tools without verified consent.
Make safety claims auditable. If a company says a model is too dangerous, independent evaluators should check, and the evidence should be published in a standard format, much like a clinical trial.
Treat risk claims as advertising claims. If you use danger to market a product, you should be able to substantiate it, like any other claim.
Put users at the table. Organisations adopting AI should involve the people who will use it before they sign the IT contract.
Less fear, more rules
AI is mostly an accelerator. It speeds up the work we already do, rewards people who already know their craft, and so far makes IT budgets bigger before it makes organisations better.
The real risks are concrete and here today: a stranger cloning your face, a fake voice note from your boss, a company spending millions without knowing why. Those deserve rules now, not another dramatic launch post.
So the next time a tech company tells you its new model is terrifying, ask two questions. Who checked? And who profits if you believe it?
If you want to help your team use AI with curiosity rather than fear, that is exactly what my corporate trainings are built for. You can find them, along with more on AI strategy and behavioural economics, at yusufoc.com.
Comments