
The European Union has set hard deadlines for the EU AI Act, and tech firms must follow them or pay huge fines. European officials passed this broad law to set guardrails on software that recognizes faces, sorts job applications, and builds synthetic text. If you build software or run an online tool, these new requirements will touch how you work.
The short answer is simple. The EU AI Act sorts artificial intelligence tools into risk groups, completely bans a few of them, and demands clear testing and reporting for the rest. Companies that break the rules can face fines of up to 35 million euros or seven percent of their global annual sales, whichever is higher.
Most small teams think rules like this only apply to giants like Google or Microsoft. That is not how the text is written. Anyone offering software inside the European market has to follow along, even if their servers sit in Texas or Tokyo. Let us break down what is happening, what gets banned first, and what you need to track.
What Is the EU AI Act and Who Must Follow It?
The European Union created the AI Act to give buyers and citizens protection from biased or dangerous software systems. Lawmakers worked on the draft for three years before the European Parliament voted to approve it in the spring of 2024. You can read the official timeline on the European Parliament website.
The law covers anyone who sells, hosts, or imports computer programs that use machine learning or algorithmic decision systems in EU countries. If your app has active European users, the location of your business headquarters does not matter. The law looks at where the end user lives. A developer working alone in Toronto faces the same rules as an enterprise team in Berlin if their tool touches European consumers.
Not every computer program counts as artificial intelligence under this framework. Simple scripts that follow fixed if-then rules do not fall under the strict sections. The act targets programs that infer ways to achieve goals, adapt their outputs, and process inputs like text, audio, and visual clips without explicit manual steps.
If you keep up with broader tech shifts on our PromptNova homepage, you know that keeping track of legal updates is just as important as watching new software updates. The rules create four clear risk groups: unacceptable risk, high risk, specific transparency risk, and minimal risk.
Which AI Systems Are Completely Banned?
The first deadline bans tools that the European Union labels an unacceptable risk. These bans take effect six months after the law enters the official journal. Officials decided that some uses cross basic human rights lines, so no company can offer them within member states under any circumstances.
Social scoring systems by governments or private businesses are totally illegal. If a bank or local council wants to track your daily shopping and online habits to give you a public trust score, the law blocks it. China uses social scoring tools, but the European Union wants to stop that model from ever taking root in its cities.
Real-time remote biometric identification in public spaces by police also faces heavy bans. There are very narrow exceptions, such as searching for a kidnap victim or preventing an immediate terror threat. Even then, police officers must get permission from a judge before turning on public facial scanning cameras.
Cognitive behavioral manipulation that targets vulnerable groups is another banned category. An example would be an audio toy that uses voice prompts to urge a young child to perform dangerous acts. Scraping facial images from security footage or the public web to build facial databases, like Clearview AI did, is also forbidden under the text.
Emotion recognition software inside schools and workplaces is banned too. Your employer cannot run webcams to scan your face and guess whether you feel happy, tired, or angry during your shift. The lawmakers decided that these systems are both unreliable and intrusive.
How Do High Risk AI Rules Work?
The largest part of the law focuses on high risk systems. These are tools that do not face total bans, but can still harm a person's health, safety, or basic civil rights if they fail. Companies building high risk software have to complete heavy compliance steps before launching.
What counts as high risk? Here are the main areas named in the legal text:
- Hiring tools that filter resumes or rank job applicants during video interviews.
- Credit scoring tools that decide if a family can get a home loan.
- Educational software that scores standardized tests or decides school admissions.
- Healthcare software used to help doctors triage emergency patients.
- Critical infrastructure controls for water, power, and transport networks.
- Law enforcement tools that assess the risk of a person committing a future crime.
If your tool falls into one of these buckets, you cannot just push code to a server. You must set up a full risk management system that runs throughout the tool's lifecycle. You have to train your models on high-quality data to reduce bias, keep detailed technical logs, and make sure human workers can stop or override the machine at any moment.
Companies must also register their high risk systems in a public European database. This lets watchdogs and consumers see who built the tool, what training data it used, and how the team tests for errors. If you follow modern generative video tools like the OpenAI Sora AI Video Model Features and Release Context, you know that tracking data inputs and system limits is already becoming standard practice for responsible teams.
What Are the Rules for ChatGPT and General Purpose AI?
When the European Union first started writing this law in 2021, large language models were not widespread. The sudden boom of tools from OpenAI, Google, and Anthropic forced regulators to rewrite whole sections to address general purpose AI models. These are base models that can do hundreds of different jobs rather than one single task.
The rules split general purpose models into two tiers. The first tier covers any standard base model. The creators must provide technical documentation, follow European copyright laws when gathering web data, and publish a detailed summary of the content used to train the system. This gives artists and publishers a way to see if their books or articles were copied into the model.
The second tier covers models that carry systemic risk. The law uses a specific computing metric here: any model trained using more than 10 to the power of 25 floating point operations, or FLOPs, hits this tier. The biggest models from major tech firms pass this line today. You can read technical breakdowns of compute thresholds on research pages like the Wikipedia entry on the Artificial Intelligence Act.
Makers of these massive models must run model evaluations, perform red-teaming tests to find security holes, track and report serious incidents to European regulators, and ensure strong cybersecurity protections. The European AI Office, a new regulatory body set up in Brussels, directly oversees these model makers.
Open source models receive some exemptions, but only if they stay below the systemic risk threshold and do not fall into high risk categories. If an open source model releases all its weights and code publicly, it skips some paperwork, but it still has to respect copyright laws.
Watermarks and Deepfakes: What Must Be Labeled?
Have you ever seen a photo or video online and wondered if it was real? The EU AI Act addresses this directly by requiring clear watermarks and labels on synthetic media. Any system that generates synthetic audio, photos, or video clips must mark those files in a machine-readable format.
If you publish a deepfake that shows real people doing or saying things they never did, you must disclose that the content was artificially created. The disclosure must be clear and visible to anyone viewing or listening to the material. An exception exists for obvious parody or artistic work, but creators must still display the label so users are not misled.
Text generators face disclosure rules too. If an online service uses an AI chatbot to talk to customers, it must inform people right away that they are speaking with a machine, unless that fact is obvious from the context. News articles and public interest texts generated by software without human editing must carry similar notices.
This is a big change for media outlets, marketing agencies, and social platforms. Watermarking standards are still being refined by groups like the Coalition for Content Provenance and Authenticity, but tech teams must get their technical tagging ready now to comply with the enforcement dates.
The Official Deadlines You Need to Track
The European Union did not turn on all rules in one day. Instead, they staggered the enforcement over three years so businesses have time to adapt. Here is the exact schedule you should mark on your calendar:
- 6 Months: Bans on unacceptable risk systems take effect. Emotion tracking at work, social scoring, and untargeted web scraping for faces must stop.
- 12 Months: Rules for general purpose AI models begin. Model creators must follow copyright rules and share model training summaries. Governance authorities take their seats.
- 24 Months: Most high risk AI rules become fully enforceable. Systems used in schools, jobs, loans, and public services must meet all data and testing standards.
- 36 Months: High risk systems that are built into products already covered by older EU safety laws, such as medical devices, cars, and airplanes, must comply.
Missing these dates carries severe costs. Small businesses and startups get slight discounts on penalties, but fines can still reach 15 million euros or three percent of global revenue for general violations. Supplying false or misleading information to regulators carries fines of up to 7.5 million euros or 1.5 percent of sales.
The European Commission provides guidance documents and sandbox environments where developers can test their tools safely before launching. These sandboxes help small teams check if their software meets standards without fearing immediate fines while testing.
How Tech Teams Should Prepare Right Now
You do not need to hire an expensive law firm tomorrow, but you do need an accurate inventory of the software tools your team uses. Start by listing every algorithm and third-party API in your product stack. Determine whether your tool handles sensitive decisions like hiring, identity verification, or user profiling.
Talk to your vendors. If your app relies on outside models, ask those providers for their compliance roadmaps. Under the law, if you take a base model from another company and modify it for a high risk purpose, you might inherit the legal duties of a provider. That means you become responsible for technical files, logging, and safety testing.
Set up clean records for your training data. Keep logs showing where you got your text, images, or code. Document how you cleaned the data and what steps you took to prevent bias against protected groups. Having clean paperwork ready will save your team months of frantic work when European regulators start asking questions.
Watch for updates on the official European Commission digital policies page. They publish ongoing codes of practice that explain how to meet the rules in day-to-day coding work.
What This Means for the Future of Technology
Some critics claim these strict rules will push tech innovators out of Europe. They point to delays in model releases across the continent as evidence that red tape slows down development. On the other side, consumer groups argue that clear guardrails build trust, making ordinary people more comfortable using automated tools in daily life.
We saw a similar story play out with GDPR privacy rules in 2018. Many companies complained at first, but the rules eventually became a global standard that shaped privacy laws in California, Brazil, and across Asia. The EU AI Act will likely have the same global reach. If you want to sell software anywhere in the developed world, building to European standards will soon be the easiest path.
Whether you run an independent app, manage an enterprise engineering team, or read tech news to stay informed, watching these rules matters. The era of unregulated automated systems is closing. Building clean, honest, and reliable software is no longer just good engineering practice; it is now the law.
Keep following our guides and updates on the PromptNova main hub to stay ahead of upcoming regulations, model releases, and technical shifts. The best time to check your code is before the deadlines arrive.
Comments
Post a Comment
Thank you for your comment