AI rules at work, explained

The EU AI Act is the first broad law anywhere that regulates AI by how risky its use is, not by what the technology is. It sorts AI systems into tiers. A few uses are simply banned: things like social scoring and certain kinds of manipulation. “High-risk” uses (AI that screens job applicants, scores creditworthiness, or makes decisions in healthcare) carry the heaviest obligations. Most everyday tools sit in lighter tiers, where the main rule is transparency: you should be told when you’re talking to AI, not a person.

It rolls out in stages, and the stages keep moving. The bans took effect on 2 February 2025. Rules for general-purpose AI models (the large systems that power chatbots) began on 2 August 2025. The heaviest rules, for high-risk systems, were originally set for 2 August 2026 (European Commission AI Act timeline; DLA Piper). Then, in May 2026, the EU agreed to push that high-risk deadline back (to 2 December 2027 for standalone systems like hiring and credit tools) under a package called the Digital Omnibus (Gibson Dunn; Hogan Lovells). The delay isn’t fully final until it’s published in the EU’s Official Journal, expected in the coming weeks. The lesson worth keeping: this is a moving target, and the date you read last year may not be the date that’s true today.

“But I’m not in Europe.”

This is the part people miss. The Act is written to reach outside the EU on purpose, modelled on the way GDPR did. Under Article 2, it can apply when an AI system’s output is used inside the EU, regardless of where your company, your servers, or your staff sit (William Fry). If you work for a Canadian or US company with European customers, clients, or employees, your organization may already be in scope.

What this means for you, practically: you almost certainly don’t need to read the statute. But if your team is building or deploying an AI system that touches hiring, lending, insurance, or anyone in Europe, that’s the moment to ask your legal or compliance team a direct question rather than assume someone else has it covered. Knowing the law exists, and roughly what it cares about, is enough to know when to raise your hand.

This is the rule most likely to affect you this week, so I’ll be blunt about it. When you type into a free, consumer chatbot, treat it like speaking in a public square. You don’t fully control where the words go.

Here’s the concrete mechanism, not a vibe. OpenAI uses conversations from consumer ChatGPT accounts (Free, Plus, and Pro) to train future models by default, unless you turn that off in settings or use a business tier like Business or Enterprise (OpenAI Help Center). So the client list, the unreleased numbers, the employee’s health note you pasted in to “help write a sensitive email”: that content can become part of how the system gets trained. You can’t un-paste it.

This is Nicolle’s Golden Rule #1, and it has two halves: don’t outsource your judgement, and don’t share sensitive data.

The simple test before you paste

Ask one question: would I be comfortable if this appeared outside the company? If the answer is no, don’t put the real thing into a public tool.

What counts as sensitive

• →Client or customer names, accounts, and personal details
• →Unreleased financials, deal terms, or strategy
• →Anything covered by a contract or NDA
• →Health, HR, or other personal data about real people
• →Source code or proprietary documents your company owns

The workaround that keeps the productivity

You usually don’t need the real data to get the help. Feed the AI a fictional, what-if version instead. Swap the real client for “a mid-sized manufacturing client.” Change the real $4.2M figure to a made-up one. Strip the names. The model helps you with structure, tone, and logic just as well. And nothing confidential leaves the building.

When you genuinely need to work on the real thing, two safer routes exist:

• →Use your employer’s sanctioned, enterprise-tier tool, where data is generally excluded from training by default. But confirm that’s actually how yours is configured.
• →Run a private or local AI model on your own machine, where the data never leaves your device. I walk through that option on the local AI page →.

Say you ask an image tool for a logo, or a chatbot for ad copy, and you ship it. Whose is it? The honest answer: less yours than you’d think.

In the United States, the Copyright Office settled the core question in its January 2025 report. Material generated purely by AI from a prompt is not protected by copyright. A prompt (even a long, clever, much-revised one) does not by itself make you the author, because you’re not controlling the specific expressive output the way an author does (U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2).

What can be protected is the human contribution: your own creative edits, your arrangement and selection, the parts where a person clearly shaped the work. AI used as a tool inside a human creative process doesn’t poison the whole thing. Pure AI output, with a human just pressing “generate,” does.

Why this matters at work

• →A purely AI-made logo or jingle may be one nobody can stop a competitor from copying: there’s nothing to enforce.
• →The reverse risk exists too: AI is trained on existing work, and it can produce something close to material someone else does own. “The AI made it” is not a defence.
• →Your output is only as defensible as the human judgement layered on top. Which is, conveniently, the same reason the work is any good in the first place.

This is the practical version of a line I keep coming back to: AI is the intern, not the author. It drafts. You decide, you shape, you own the result, legally and otherwise.

Most companies now have an AI policy. Most employees have not read it past the title. Given that about half of workers admit to breaking these rules (many because they don’t know what they say), fifteen minutes here is the cheapest insurance you’ll buy all year (KPMG).

You don’t need to memorize it. You need answers to five questions. Find these, and you’ve read the policy.

• →Which tools am I actually allowed to use? Approved list, or anything goes? Is the free version of a tool treated differently from the paid enterprise one?
• →What data am I forbidden from entering? Almost every policy bans client, customer, and confidential data in public tools. Know the exact words yours uses.
• →Do I have to disclose when I used AI? On client work, in code, in published content. Some employers require a flag, some don’t.
• →Who is accountable for the output? The answer is always you. The policy just makes it official. AI can’t be held responsible for a mistake, so a qualified human has to own it.
• →Who do I ask when I’m unsure? There’s usually a named team or inbox. Using it is a sign of competence, not weakness.

Watch for these gaps

A policy can exist and still leave you exposed. Be alert when:

• →The policy bans a tool but offers no approved alternative. That’s the setup that drives people to use it in secret. (In the KPMG study, the breaches were highest at companies with outright bans: 67% of those employees admitted to feeding sensitive data into public tools anyway.)
• →“AI” is treated as one thing, with no distinction between a public chatbot and a private enterprise tool. The risk is wildly different.
• →There’s a mandate to use AI but no guidance on how. That’s how you get usage for the sake of usage, instead of usage for the sake of an outcome.

One reframe worth holding onto: a good AI policy exists to protect you and the people in your data, not to slow you down. The point is to keep your judgement (and the company’s trust) intact, not to win a usage contest.