The EU AI Act used to be talked about as a looming deadline, something that would eventually change how AI companies operate, at some point down the line.
That point has already passed.
Since 2 August 2026, regulators have had the power to actually fine AI companies, not just tell them off.
It is not the whole law switching on all at once.
It is one specific, meaningful part, and it has now been active for over a month.
Here is what that means in practice, and how much of it a normal person browsing ChatGPT or Claude will actually notice.
What actually happened
From 2 August 2026, the European Commission's AI Office can fine providers of general-purpose AI models, the big foundation models behind tools like ChatGPT, Claude, and Gemini, for breaking the rules set out in the AI Act.
Before this date, the law existed on paper.
Now it has an enforcement arm that can hand out real financial penalties.
The numbers are large by design.
Under Article 101 of the Act, the Commission can fine a general-purpose AI provider up to EUR15,000,000 or 3% of its annual worldwide turnover, whichever figure is bigger.
That applies when a provider is found to have broken the rules intentionally or through negligence, refused to hand over documents the Commission asked for, given false or incomplete information, or blocked the Commission from evaluating a model.
Article 99 sets a similar scale of fines, up to EUR15 million or 3% of turnover, for breaches of several other parts of the Act, including the transparency rules under Article 50.
Practices the Act bans outright carry the steepest penalty of all: up to EUR35 million or 7% of turnover.
For context, turnover-based fines like this are the same mechanism the EU already uses under GDPR, its data protection law.
The logic is the same too: a flat fine that a huge company could shrug off does not change behaviour, but a fine tied to a percentage of global revenue does.
The AI Office is not limited to writing cheques for wrongdoing after the fact, either.
It can request technical documentation from providers, run its own evaluations of a model, and demand corrective action.
In other words, it can go looking, not just wait for a complaint.
Alongside the fines, new transparency rules also took effect on the same date.
Article 50 now requires AI chatbots to make it clear that someone is talking to an automated system, not a person.
It also requires AI-generated or manipulated content, including deepfakes, to carry a machine-readable mark so software can detect it automatically.
Providers of general-purpose AI models now have to document what their models were trained on for regulators, publish a copyright policy, and put out a reasonably detailed summary of the training content itself.
According to the Act's own text (Articles 99 and 101), those are the fine thresholds and the conditions that trigger them.
Wilson Sonsini's legal analysis of the Act, along with reporting from CNBC and Help Net Security, all name OpenAI, Anthropic, and Google among the companies whose models now fall under these powers, since the rules apply to anyone making a model available in the EU, regardless of where the company itself is based.
There is one grace period worth knowing about.
Models that were already on the EU market before 2 August get extra time on the watermarking requirement specifically: that obligation does not bite until 2 December 2026.
Why it actually matters
Honestly, most people using AI day to day will not notice anything different.
Nobody's ChatGPT window is about to change because of a date on a regulation.
That is worth saying plainly rather than dressing this up as more dramatic than it is.
What has changed is quieter but real: AI companies now face genuine financial consequences for how they train models, what they disclose about that training, and how they label AI-generated content.
Up to now, a lot of the AI Act's requirements existed as commitments companies were working toward.
Now there is a regulator that can request evidence, run its own checks, and fine a company that falls short.
Over time, that is the kind of pressure that shapes what tools actually look like.
The transparency rules are the part most likely to eventually show up somewhere a reader can see it: chatbots stating plainly that they are automated, and AI-generated images, video, or audio carrying some kind of detectable marker.
Whether that becomes something people actually notice in the products they use, rather than just a line in a policy document, is the part still worth watching.
