AI Law: What Every Creator & Founder Needs to Know

In the US and Canada, copyright requires human authorship — so content created entirely by an AI tool, with no meaningful human creative input, gets no copyright protection. The US Copyright Office confirmed this position in its January 2025 report: prompting an AI alone does not make you the author of the output. That said, if you made substantial creative choices — selecting, arranging, or modifying AI-generated material in ways that reflect your own expression — those human-authored elements can qualify for protection. The unprotected AI portions and the protected human-authored portions exist side by side in the same work.

The AI platforms generally assign output ownership to you — OpenAI, Anthropic, and Midjourney all state in their terms that you own what you generate. The critical catch: that contractual ownership doesn't automatically mean you have copyright. If the output lacks sufficient human creative expression, copyright law won't protect it regardless of what the platform's terms say. You also grant the platform a broad license to your prompts and outputs under most standard terms, and enterprise-tier agreements often differ significantly from free-tier terms. Read the plan you're actually on, not the headline marketing.

Yes — unlike copyright, trademark law has no human authorship requirement. The USPTO cares whether your mark is distinctive, used in commerce, and doesn't conflict with existing marks, not how it was created. An AI-generated logo that you've adopted as your brand identifier and used in commerce is eligible for trademark registration. However, you won't be able to rely on copyright to stop someone from copying the logo's visual design unless you layered in sufficient human-authored creative choices. A trademark clearance search is essential before investing in any logo, AI-generated or not, because registration doesn't protect you from a prior user with superior common-law rights.

This is the most actively litigated question in AI law right now, and the honest answer is: not yet settled. More than 50 US federal lawsuits are pending against AI developers including OpenAI, Stability AI, Anthropic, and Google, with artists, publishers, and record labels all claiming that scraping copyrighted works for training data is infringement. One 2025 federal court ruling found Anthropic's use of books to train Claude was fair use; others have allowed infringement claims to survive motions to dismiss. The US Copyright Office acknowledged in May 2025 that some training uses will qualify as fair use and some will not — courts will decide case by case. In Canada, the same legal question is open: no binding ruling exists, and the federal AI legislation (AIDA) that might have addressed it was killed when Parliament prorogued in January 2025.

In most US states, no — and the law is moving fast to close remaining gaps. At the federal level, the TAKE IT DOWN ACT (signed May 2025) prohibits non-consensual intimate deepfakes and requires platforms to remove them. Tennessee's ELVIS Act (effective July 2024) explicitly covers AI voice cloning without consent. The proposed NO FAKES Act, if passed, would create a federal IP right in every person's voice and likeness. In California, existing right-of-publicity law and proposed legislation give creators strong grounds to act against unauthorized AI replicas used for commercial purposes. Canada has no equivalent federal deepfake statute yet, but provincial privacy laws and common-law personality rights provide a partial remedy. If someone is profiting from an AI clone of your voice or image without your consent, you likely have actionable claims today — but the specific remedies vary by state and province.

More than most people realize. While platforms broadly let you keep ownership of outputs, you're typically granting them a royalty-free, worldwide license to use your prompts and content for model improvement, product development, and in some cases marketing. Anthropic's consumer-tier terms (as of late 2025) include training on your inputs unless you opt out. Midjourney's terms grant them the right to reproduce and sublicense your prompts and generated images. For paid or enterprise tiers, these terms are usually narrower and more protective — a meaningful reason to use business accounts for commercial work. Before you put a client's brief, a screenplay, or proprietary business strategy into an AI tool, check whether your inputs are being used to train models and whether confidentiality is protected.

Three things require careful attention. First, IP ownership and warranties: if you're delivering AI-assisted work to a client, who owns it, and are you warranting that it's original and free of third-party claims? Standard 'all work is original and owned by me' warranties may be impossible to make honestly about AI-generated content given its uncertain copyright status. Second, indemnification exposure: some clients are now demanding indemnification for any AI copyright claims arising from deliverables — that's a significant risk to accept without caps. Third, disclosure obligations: an increasing number of contracts, platform policies, and industry standards require disclosure when AI was used in production.

If you're collecting personal data from users in California, you're subject to the CCPA/CPRA — which requires a privacy policy, opt-out rights for data sales and sharing, and deletion rights. If you handle data of Quebec residents, Quebec's Law 25 (fully in force since September 2024) is among the strictest privacy laws in North America: it requires explicit, specific consent for AI processing of personal data, mandatory privacy impact assessments for any AI system that collects personal information, and strict rules on transferring data outside Quebec. Penalties run up to 4% of worldwide revenue. Federal Canadian law (PIPEDA) applies across the rest of Canada, with breach notification obligations. The practical reality for most AI startups: build your privacy architecture around the strictest rule that applies to your user base.

The EU AI Act is the world's first comprehensive AI regulation, formally effective August 2024 with most substantive provisions enforced from August 2026. It takes a risk-tiered approach: a small category of AI uses are outright prohibited; high-risk AI systems face strict compliance requirements including conformity assessments, transparency obligations, and human oversight mandates; and general-purpose AI models have their own transparency and copyright compliance requirements. It applies to you if you offer AI products or services to EU users, regardless of where your company is incorporated. In Canada, the federal AIDA legislation died with the prorogation of Parliament in January 2025, leaving no equivalent national law — though Quebec Law 25 fills significant gaps for that province.

This is still being worked out in courts, but the current framework is: you (the user) carry meaningful liability for how you deploy AI outputs, the AI developer may face liability if its system was negligently built or its outputs foreseeably infringe rights, and platforms hosting AI-generated content retain DMCA safe harbor protection only if they respond to takedown notices. If you publish AI-generated content that defames someone, infringes copyright, or is used to defraud — you face direct liability on the same legal theories that would apply to human-created content. The fact that an AI produced it is not a defense. Contracts between users and AI companies typically disclaim the company's liability for outputs, meaning users bear more risk than they often assume.

Platforms can remove your AI-generated content under their own terms of service for virtually any reason — DMCA is not a shield for content the platform finds objectionable. DMCA's counter-notice process applies specifically when your content was taken down in response to a copyright takedown claim: if you believe the takedown was wrong, you can file a counter-notice, and the content should be restored within 10–14 business days unless the claimant files suit. However, if your AI-generated content closely mimics a copyrighted work — style, character, or elements that cross into protected expression — the takedown may be legally justified. AI-generated content is not automatically exempt from infringement claims just because a machine produced it.

You need legal counsel when the stakes of getting it wrong outweigh the cost of professional advice — which happens earlier than most people expect. Concrete triggers: you're building a product or startup that uses AI to process user data (privacy law compliance is mandatory from day one, not at Series A); you're signing a contract that includes IP warranties, indemnification for AI outputs, or AI-specific disclosure obligations; you've received a cease-and-desist or copyright claim involving AI-generated content; you're a creator and someone has cloned your voice, image, or style with AI without consent; you're a founder raising money and your AI-generated assets are being treated as owned IP. AI law is moving fast enough that general-practice attorneys often lack the specialized knowledge to advise accurately; look for counsel with a demonstrated track record in IP and technology law specifically.