The intellectual property deep dive: patents, trademarks, secrets and how to actually use them

What IP is, and the only model you need

Intellectual property is a set of legal rights over creations of the mind: inventions, brand identifiers, creative works, and confidential know-how. Each right is a government-granted (or, for trade secrets and copyright, a government-recognized) monopoly of a particular shape, for a particular time, in a particular place. There are four families you will meet constantly, plus a handful of specialized rights:

• Patents protect how something works — a function, a process, a machine, a composition. Term ~20 years from filing. You must disclose the invention to the public in exchange.
• Trademarks protect brand identifiers — names, logos, sounds, the look of a product. They can last forever if you keep using and renewing them.
• Copyright protects original expression — code, text, design, music, video. It is automatic on creation and lasts a very long time (typically the author’s life plus 70 years).
• Trade secrets protect confidential, commercially valuable information — a formula, an algorithm, a customer list — for as long as you keep it secret. No filing, no expiry, but no protection once it leaks.

Three properties cut across all of them and trip up almost everyone:

IP is territorial. There is no “worldwide patent” or “global trademark.” Rights exist country by country (or region by region). A US patent stops infringement in the United States and nowhere else. To be protected in Germany, China and Japan, you file in Germany, China and Japan (or through a regional or international system that routes you there). This is why a global product needs a deliberate, budgeted filing strategy, not a single application.

Most of the world is first-to-file, and most of the world has no grace period. Whoever files first generally wins, even if someone else invented it earlier. And in Europe, China and most countries, any public disclosure before you file — a conference talk, a demo, a product launch, a press piece — destroys novelty and kills your right to patent it there. The US gives inventors a narrow one-year grace period; relying on it forfeits most foreign rights. The practical rule writes itself: file before you tell anyone who isn’t under an NDA.

Rights are time-limited and use-limited. Patents expire and pass into the public domain (that is the deal: temporary monopoly in exchange for permanent disclosure). Trademarks die if you stop using them or let the name become generic. Copyright is long but not forever. Trade secrets last only as long as the secrecy.

Why a founder should actually care

IP is not an abstraction; it shows up as money and risk at specific moments:

• Defensibility. If your edge is copyable in a weekend, IP is sometimes the only moat the law will give you — especially in hardware, biotech, materials and deep tech.
• Valuation and fundraising. Intangible assets are the majority of enterprise value in modern markets. Investors run IP due diligence: are the inventions assigned to the company, is there a patent position in deep tech, are you free to operate, are you about to be sued? Messy IP can re-price or kill a round.
• M&A. In an acquisition, the buyer’s lawyers will comb your chain of title. A founder or contractor who never assigned their work, an open-source license that “infected” your codebase, or a trademark you don’t actually own can knock millions off the price or blow up the deal.
• Not getting blocked or sued. Freedom-to-operate is the flip side of owning IP: even with your own patents, you can infringe someone else’s. Knowing the landscape before you ship is cheaper than discovering it in a complaint.

Takeaway: treat IP as a product decision, not a legal afterthought. The cheap moves — assignments signed, a provisional filed before you launch, a trademark cleared before you print the logo — cost little and are nearly impossible to fix once missed.

Utility patents: protecting how something works

A utility patent (just “patent” under the European Patent Convention, “invention patent” in China) protects the functional aspects of an invention: a new process, machine, article of manufacture, or composition of matter. In the US these four categories come from 35 U.S.C. §101, and the invention must be “new and useful.” To be granted, an invention must clear three bars everywhere, under slightly different names:

• Novelty — it must be new, not already in the prior art (US: §102; EPC: Art. 54).
• Non-obviousness / inventive step — it must not be an obvious combination to a “person of ordinary skill in the art” (US: §103, framed by Graham v. John Deere (1966) and loosened toward more obviousness findings by KSR v. Teleflex (2007); EPC: Art. 56, the “problem-and-solution” approach).
• Utility / industrial applicability — it must do something useful (EPC Art. 57).

The term is 20 years from the filing date (US: §154; EPC: Art. 63), not from grant — so years lost in examination shorten your enforceable life, which patent-term adjustment can partly restore. Scope is defined by the claims, not the drawings or description: independent claims define the broadest protection, dependent claims add limitations, and in the US infringement requires every element of a claim to be present (the “all-elements rule”). This is why claim drafting is where patent-attorney money is well spent: claims too narrow are designed around in an afternoon; claims too broad get invalidated.

Some things are not patentable: abstract ideas, laws of nature and natural phenomena are judicial exceptions in the US. Alice Corp. v. CLS Bank (2014) created a two-step test that invalidated a computer-implemented escrow patent as an “abstract idea on a generic computer,” gutting many software and business-method patents; Mayo v. Prometheus (2012) did the same for a diagnostic correlation; Myriad (2013) held isolated natural DNA unpatentable while lab-made cDNA stayed eligible. In Europe, EPC Art. 52 excludes “computer programs as such,” but a computer-implemented invention with technical character can be patented. The lesson for software and AI inventions: claim a concrete technical improvement, not the idea.

Design rights: protecting how something looks

Design rights protect the appearance of a product — shape, surface, ornamentation — not its function. In the US this is a design patent (15 years from grant, no maintenance fees). In the EU, a registered Community design lasts up to 25 years (renewed in 5-year blocks) and an unregistered Community design gives 3 years of automatic protection against copying — useful for fast-moving fashion and consumer goods. Design rights were the quiet heavyweight in Apple v. Samsung: a chunk of the billion-dollar verdict turned on the ornamental design of the phone, not the software.

Trademarks: protecting your brand

A trademark is anything that identifies the commercial source of goods or services: a name, logo, slogan, sometimes a sound, color or product shape (“trade dress”). Protection runs on a distinctiveness spectrum: fanciful (Kodak, Xerox) and arbitrary (Apple for computers) marks are strongest; suggestive marks are protectable; merely descriptive marks need acquired “secondary meaning”; and generic terms can never be a trademark. Marks can last indefinitely if you keep using and renewing them — but you can also lose one. Genericide killed “aspirin,” “escalator” and “thermos” once the public used them as the product name itself. Goods and services are filed in Nice classification classes (45 of them); you pay per class, and your protection only covers the classes you register. Use the ™ symbol for an unregistered claim and ® only once a mark is actually registered.

Copyright: protecting expression

Copyright protects original works of authorship fixed in a tangible medium — code, text, graphics, music, film. It is automatic: you own it the moment you create the work, with no filing required. It protects the expression, never the underlying idea, system or method (the “idea-expression dichotomy”) — which is why two people can independently write apps that do the same thing. Term is long: in the US and EU, the author’s life plus 70 years (works for hire: 95 years from publication in the US). Registration is optional but powerful: in the US you must register before you can sue for infringement, and timely registration unlocks statutory damages and attorney’s fees. The big limit is fair use (US) / fair dealing (UK/Commonwealth): a four-factor analysis — purpose (is it transformative?), nature of the work, amount used, and market effect — that lets others quote, parody, comment and, increasingly contested, train AI models. Many countries also recognize moral rights (attribution and integrity) that authors keep even after selling the economic rights.

Trade secrets: protecting what you don’t disclose

A trade secret is any information that has commercial value because it is secret and that you take reasonable measures to keep secret — a formula, an algorithm, a manufacturing process, a customer list. There is no application and no expiry: Coca-Cola’s formula has been a trade secret since 1886, outliving any patent many times over. In the US, the Defend Trade Secrets Act of 2016 created a federal civil claim for misappropriation; the EU’s Trade Secrets Directive (2016/943) harmonized protection across the bloc. The catch: protection evaporates the moment the secret is lawfully reverse-engineered, independently discovered, or simply leaks — and then it is gone forever, with no patent to fall back on. “Reasonable measures” is not optional decoration; courts have denied protection to companies that called something secret but emailed it around without NDAs or access controls.

Utility models and the specialized rights

A utility model (“petty patent”) is a faster, cheaper, lower-bar cousin of the patent available in Germany, China, Japan and many other countries — typically no substantive examination, a shorter term (~10 years), and a lower inventive-step requirement. It is a workhorse in China and a useful quick shield for incremental mechanical inventions. Note for US founders: the US has no utility model, and a US “utility patent” is the fully examined 20-year right, not the same thing. Beyond these sit the specialized rights you should at least recognize: geographical indications (Champagne, Parma ham, Roquefort), plant variety rights, the EU’s sui generis database right protecting substantial investment in compiling a database, semiconductor topography rights, and publicity / image rights over a person’s name and likeness.

Which right fits what

A quick mapping, because founders often reach for the wrong tool:

• A novel hardware mechanism, drug, chemical or process → patent (and consider a utility model abroad for speed).
• The way your product looks → design right / design patent.
• Your name, logo, app icon → trademark (clear it before you print it).
• Your source code, content, UI copy, designs → copyright (automatic) plus trade secret for the parts you keep closed.
• A back-end algorithm or recipe that nobody can see or reverse-engineer → trade secret.

Takeaway: the four families overlap. One phone embodies utility patents (the radio), design rights (the shape), trademarks (the logo), copyright (the software) and trade secrets (the supply chain). Map your product the same way and protect each layer with the right tool.

The United States

The USPTO runs a first-inventor-to-file system since the America Invents Act took effect on 16 March 2013 — the date the US abandoned its historic “first-to-invent” rule. Two features matter to founders. First, the provisional application: a cheap, informal filing that locks in a priority date and buys you 12 months to file the full non-provisional (or go international) while you refine the invention or raise money. Second, the PTAB and inter partes review (IPR): a faster, cheaper administrative route to challenge a patent’s validity than full litigation, created by the AIA and now central to patent disputes. Separately, the US International Trade Commission can issue Section 337 exclusion orders that block infringing imports at the border — a fast, powerful weapon against overseas manufacturers.

Europe: the EPO and the new Unified Patent Court

The European Patent Office examines one application and grants a European patent — but historically that was a “bundle” of national patents you then had to validate (and translate, and pay annuities on) country by country, enforced in each national court. The landmark change: on 1 June 2023 the Unitary Patent and the Unified Patent Court (UPC) went live. You can now get a single unitary patent covering most EU states at once, litigated in one specialized court whose decisions take effect across all participating countries. That cuts cost and gives you one injunction for a whole market — but it cuts both ways: a single UPC revocation action can knock your patent out everywhere at once, so many holders are still weighing whether to “opt out” existing patents. Note the EPO is not an EU body: it has 39 member states, broader than the EU.

The EU single market: EUIPO

For brands and designs, the EUIPO is a genuine one-stop shop: a single EU Trade Mark (EUTM) or Registered Community Design covers all 27 EU member states with one filing, one fee, one renewal. It is one of the best-value moves in IP — one application, a market of ~450 million people. The trade-off is all-or-nothing: an EUTM can be refused or cancelled based on a conflict in a single member state.

China

China’s CNIPA is now the largest patent and trademark office in the world by volume, and treating Chinese IP as a backwater is a years-out-of-date mistake. Three things every founder should know. First, China is strictly first-to-file, including for trademarks, and trademark squatting is rampant: bad actors register foreign brands before the real owner arrives, then demand payment — so register your marks in China early, in Chinese characters too. Second, utility models are cheap, fast and widely used. Third, enforcement has improved dramatically: China created specialized IP courts (Beijing, Shanghai, Guangzhou and now a national appeals tribunal at the Supreme People’s Court), and the 2021 Patent Law amendment introduced punitive damages (up to five times actual damages for willful infringement) and patent-term adjustment/extension. Damages awards have climbed sharply.

Japan, Korea, India and emerging Asia

Japan (JPO) and South Korea (KIPO) are sophisticated, examiner-driven systems with strong enforcement. India is the one to study before you assume Western rules apply: its Patents Act Section 3(d) bars patents on new forms of known substances without enhanced efficacy — the provision used to deny Novartis a patent on the cancer drug Glivec (2013) — and India has used compulsory licensing (the 2012 Natco/Bayer order on the cancer drug Nexavar) to force licensing of patented medicines on public-interest grounds. Pharma and biotech founders especially need country-specific advice in these markets.

The international systems: how you actually go global

You never file 150 separate applications. A web of treaties, administered largely by WIPO, lets one filing reach into many countries:

• Paris Convention (1883): gives you a 12-month priority right — file in one member country, and you have 12 months to file in others while keeping the original date for novelty.
• PCT (Patent Cooperation Treaty): one “international” patent application preserves your priority date in 150+ countries and delays the expensive, country-by-country “national phase” decisions to ~30 months from priority — buying ~18 extra months to decide where it’s worth paying. The PCT does not grant a patent; it defers and centralizes the early steps.
• Madrid System: one international application to register a trademark in many countries.
• Hague System: the same idea for industrial designs.
• Berne Convention: makes copyright automatic and mutually recognized across ~180 countries — no registration needed to have the right.
• TRIPS (the WTO agreement): sets the minimum IP standards every WTO member must provide, which is why the basics rhyme worldwide.

Takeaway: the standard global play for a startup is — file a home application (often a US provisional) to bank a priority date, then within 12 months file a PCT (patents) and Madrid (trademarks) to keep the whole world open, then within ~30 months spend real money only in the countries that matter. And remember the territoriality trap: outside the US there is usually no grace period, so file before you disclose.

The filing process, step by step

For a patent, the pipeline is roughly: (1) prior-art search — check what already exists, both to gauge your odds and to draft around it; (2) draft the application — a technical description and, crucially, the claims; (3) file a provisional (US) to lock the date cheaply; (4) within 12 months, file the non-provisional and/or PCT; (5) prosecution — an examiner reviews it and sends “office actions” rejecting claims; your attorney argues and amends, often over two to four years; (6) grant, then (7) maintenance fees to keep it alive. Trademarks are simpler and faster: clear the mark (search for conflicts), file in the right classes, respond to any examiner refusal, and register — typically months, not years. Copyright needs no filing to exist, only to sue (in the US).

What it really costs — and how long

Order-of-magnitude numbers, because founders consistently underestimate the lifecycle (not just the filing):

• US utility patent: roughly $10,000–$20,000+ in attorney and USPTO fees to draft, file and prosecute through grant (more for complex tech), then maintenance fees at 3.5 / 7.5 / 11.5 years — on the 2025 USPTO schedule, $2,150 / $4,040 / $8,280 for large entities (halved for small entities, quartered for micro). Grant typically takes 2–4 years.
• Europe: an EPO patent commonly runs €5,000–€30,000+ once you add examination, validation and translation across countries, plus annual renewal fees that rise each year.
• PCT: a few thousand dollars in fees to file the international application — then the real money lands at national phase, multiplied by every country you enter.
• Trademark: roughly $1,000–$2,000 per class with an attorney in the US; far less if you self-file, far more across many countries.

The pattern: filing is the cheap part; going broad and keeping rights alive is what costs. A worldwide patent family with annuities can run into six figures over its life. Budget accordingly, and file only where you can realistically enforce or sell.

Patent or trade secret? The decision that defines your strategy

These two are alternatives, and choosing wrong is costly — a patent publishes your invention to the world. Lean patent when: the invention is reverse-engineerable from your shipped product anyway (so secrecy is impossible), you want to license it, or you need a hard asset for investors. Lean trade secret when: the thing is genuinely hard to reverse-engineer (a back-end algorithm, a manufacturing process, a ranking model), it could outlive a 20-year patent, or detecting infringement would be impossible anyway. Google’s search ranking and Coca-Cola’s formula are trade secrets for exactly these reasons. A third option is defensive publication: deliberately publish an idea so that nobody can patent it, keeping it free for you to use without paying to own it. For most software, the honest answer is a blend: trade-secret the core that’s hidden, copyright the code, patent the few genuinely novel and detectable techniques, and trademark the brand.

The startup IP strategy — cheap moves that compound

You don’t need a hundred patents; you need the right few and a clean house. The high-leverage moves:

• Provisional-first. Bank priority dates cheaply with provisionals as you invent, before any launch or pitch, and convert only the ones that prove out within the 12-month window.
• Get every assignment signed. Every founder, employee and contractor must assign their IP to the company in writing, ideally on day one. This is the number-one thing diligence catches.
• Run a freedom-to-operate check before you commit to a design in a crowded field — owning a patent doesn’t mean you’re free to ship.
• Clear your trademark before you spend on a brand, and register it in your real markets (including China) early.
• Mind open-source licenses in your codebase from the start (see the partnerships chapter).

What investors and acquirers look for in IP diligence: clean chain of title (everything assigned to the company), a defensible position in deep tech (real patents or hard-to-copy secrets), freedom to operate, and no looming infringement or open-source contamination.

IP as an asset: valuation and deals

IP is property; it can be valued, sold, licensed and borrowed against. Valuation uses three lenses — cost (what it took to create/replace), market (what comparable IP sold for), and income (the future cash flows it enables, e.g. discounted royalties). Patents change hands as strategic assets: in 2011 a consortium (Apple, Microsoft and others) paid $4.5 billion for Nortel’s ~6,000 patents, and Google bought Motorola Mobility for $12.5 billion — largely for its patent portfolio — the same year. Most startups won’t see those numbers, but the principle holds: a credible IP position is value an acquirer pays for, and a missing one is a discount.

Takeaway: spend on IP like a portfolio — cheap, broad options early (provisionals, PCT, assignments), real money later only where the market and enforceability justify it.

US patent litigation: slow, expensive, jury-driven

US patent litigation is the most expensive in the world — commonly $3–5 million or more to take a case through trial — and it is fought before juries. Where you can sue narrowed after TC Heartland (2017) tightened venue rules; today a large share of cases concentrates in Delaware, the Eastern District of Texas, and the Western District of Texas (Waco), the last shaped by a single patent-heavy judge. Because trial is so costly, the cheaper administrative challenge — an IPR at the PTAB — is now a standard defensive move: instead of betting millions on a jury, you ask the patent office to cancel the patent on prior art.

NPEs: the patent-troll problem

A non-practicing entity (NPE) or patent-assertion entity (PAE) — colloquially a “patent troll” — owns patents not to build products but to extract settlements by asserting them, often against startups that can’t afford to fight. The economics are asymmetric: defending costs millions, so a demand for a few hundred thousand “to make it go away” is rational even when the patent is weak. Defenses have organized around this: RPX and Unified Patents aggregate defense and buy or invalidate troublesome patents; the LOT Network is a cross-license that immunizes members if a patent ever falls into troll hands; and an IPR can kill a weak patent for a fraction of litigation cost. For a young startup, the practical posture is: don’t panic at a demand letter, get counsel, and check whether the patent is already being challenged.

Remedies: damages and injunctions

If infringement is proven, the two main remedies are money and an order to stop. Damages are either a reasonable royalty (what a willing licensee would have paid, analyzed through the 15 Georgia-Pacific factors) or lost profits if the patentee can prove the sales it lost. Injunctions — the order to stop selling — changed profoundly with eBay v. MercExchange (2006): the Supreme Court ended automatic injunctions and required a four-factor equitable test, which made it much harder for NPEs (who lose no sales) to shut down a product, while practicing competitors can still get them. Where infringement is willful, §284 allows enhanced damages up to three times the award, and §285 plus Octane Fitness (2014) lets courts shift attorney’s fees in “exceptional” cases. This is why a written legal opinion of non-infringement, and not forwarding patents around internally, matters once you’re on notice.

Enforcement beyond the US

Outside US courts, the toolkit differs. The UPC now offers a single injunction across most of the EU — fast and, for a defendant, frightening. The EPO has a post-grant opposition window (nine months after grant) in which anyone can attack a European patent centrally and cheaply. The ITC’s Section 337 can exclude infringing goods at the US border within ~16–18 months — faster than district court and devastating for import-dependent defendants. And customs recordation lets trademark and copyright owners have border authorities seize counterfeits on sight, the front line of anti-counterfeiting.

The landmark cases worth knowing

• Apple v. Samsung (2011–2018): the defining smartphone war; over a billion dollars in verdicts, much of it on design patents, and a Supreme Court trip over how to calculate design-patent damages.
• Oracle v. Google (2010–2021): whether reimplementing Java APIs in Android infringed copyright. After a decade, the Supreme Court held Google’s use was fair use — a foundational result for software interoperability.
• Alice v. CLS Bank (2014): made “abstract idea on a generic computer” a fatal objection, invalidating swaths of software patents.
• Myriad (2013) and Mayo (2012): drew the line on patenting nature and diagnostics.
• eBay v. MercExchange (2006): ended automatic injunctions — arguably the single most important case for the troll economy.

Takeaway: litigation is a weapon of last resort and enormous cost. Most disputes settle. The cheap insurance is upstream: clean ownership, freedom-to-operate searches, opinions of counsel when on notice, and membership in defensive networks if you operate in a patent-dense field.

Co-design and joint development: the ownership trap

When two companies build something together — a co-designed product, a joint research project, a hardware/software integration — the most dangerous default is joint ownership, and not because sharing is bad but because the law’s default rules are surprising. Under US law, absent an agreement, each joint owner of a patent can independently license or exploit it without the other’s consent and without accounting for profits — so your “partner” can license your shared invention to your competitor. Under UK and much of European law, the opposite default applies: neither co-owner can license without the other’s consent, which can deadlock commercialization. Either way, the fix is the same: never rely on the default. A collaboration agreement should spell out, before work starts, the split between background IP (what each side brings in — stays theirs) and foreground IP (what’s created jointly — assign it deliberately to one party with a license back, or define exactly how it’s shared), plus who can file patents, who pays, and who controls enforcement.

Licensing: renting IP instead of selling it

A license is permission to use IP on agreed terms — the engine of most IP monetization. The dimensions that matter:

• Exclusivity: exclusive (only the licensee, even excluding the owner), sole (licensee + owner), or non-exclusive (many licensees). Exclusivity commands a premium and changes who can sue infringers.
• Scope: field of use, territory, and term — you can license a patent for medical use in Europe only, for five years.
• Money: running royalties (a % of sales), milestone payments, and/or an up-front lump sum — often combined.
• Plumbing: sublicensing rights, audit rights (to verify royalty reports), most-favored-nation clauses, improvements, and what happens on termination.

An assignment, by contrast, is an outright sale — the buyer becomes the owner. Get the distinction right in every term sheet: founders routinely “license” when they mean “assign” and vice versa.

Standard-essential patents and FRAND

Some patents are unavoidable because they cover an industry standard — you cannot build a 5G phone, a Wi-Fi chip or an H.265 video decoder without infringing them. These standard-essential patents (SEPs) come with a deal: to get their tech into the standard, holders commit to license on FRAND terms — fair, reasonable and non-discriminatory. The fights are over what “fair” means, and they are titanic: the Qualcomm, Nokia and InterDigital licensing disputes, the Apple–Qualcomm settlement, and ongoing SEP battles in the UPC and Chinese courts. Patent pools (MPEG LA, Via LA, Avanci for automotive) bundle many owners’ SEPs into one license to make this tractable. If your product implements any standard, budget for SEP royalties — they are a real, recurring cost of doing business.

Assignment, employees and the chain of title

Owning your own IP is not automatic, and the gaps are where deals die. Key rules: an employee’s inventions usually need a written assignment to vest cleanly in the company (in some places they vest by statute, in others not — don’t guess); the US copyright “work made for hire” doctrine only auto-assigns certain categories and only for true employees, so independent contractors must sign an explicit assignment or they keep the copyright in what you paid them to build. For inventions out of universities, the US Bayh-Dole Act (1980) lets universities own and license federally-funded inventions — the foundation of academic spin-outs and tech-transfer offices. And a founder who created core technology before incorporating must formally assign it into the company, or it isn’t the company’s. The recurring lesson: papered assignments from everyone, early.

Open source: the license is the contract

Open-source software is governed by IP licenses, and ignoring them is how startups get a nasty surprise in diligence. The split that matters: copyleft licenses (GPL, and especially AGPL for network services) require that if you distribute software built on them, your derivative work must also be released under the same open license — the “viral” effect that can force you to open-source your own code. Permissive licenses (MIT, BSD, Apache-2.0) let you use the code in closed products with minimal obligations — and Apache-2.0 adds an explicit patent grant, which is why many companies prefer it. Practical hygiene: keep an inventory of your dependencies and their licenses, avoid pulling AGPL/GPL code into a proprietary product unless you intend the consequences, require a Contributor License Agreement if you run an open project, and consider dual licensing (open core + commercial license) if open source is your distribution. License non-compliance is a real liability an acquirer will price — or walk away from.

Takeaway: partnerships and licensing are where IP turns into revenue — and where it quietly leaks away. The discipline is unglamorous and decisive: define background vs foreground IP before you collaborate, paper every assignment, read your open-source licenses, and never accept the law’s default ownership rule by accident.

What AI is breaking, and where it stands now

AI is stress-testing every assumption above, and the law is unsettled and moving. Three live fronts, as of 2026:

Can AI output be copyrighted? The current US position is that copyright requires human authorship: the Copyright Office and the courts (Thaler v. Perlmutter; the “Zarya of the Dawn” Midjourney comic) have refused protection for works generated autonomously by AI, while allowing copyright in the human-authored selection, arrangement and meaningful editing around AI elements. Practically: the more a human shapes and edits, the more protectable the result — purely “press generate” output is in a gray-to-unprotected zone.

Is training on copyrighted data infringement? This is the multi-billion-dollar question, fought in cases like The New York Times v. OpenAI/Microsoft, Getty Images v. Stability AI, and the artists’ suits against Stability and Midjourney. The core defense is fair use (training as a transformative use); plaintiffs argue it’s mass copying that competes with their market. Outcomes are still landing, and a parallel market of licensing deals (publishers and platforms selling training data) is forming alongside the litigation. Builders should track this closely: it shapes both your data supply and your liability.

Can an AI be an inventor? No, in nearly every major jurisdiction. The DABUS test cases — an attempt to name an AI as the inventor on patents — were rejected in the US, UK and at the EPO (which all require a human inventor); South Africa was a lone outlier that granted one. Inventions made with AI assistance are patentable, but a human must be named as the inventor.

Add to this the surge of work on likeness and voice — deepfakes are driving publicity-rights claims and proposed legislation (e.g. US “NO FAKES” proposals) to protect a person’s digital replica.

Takeaway: if AI is core to your product, IP is a strategy question, not a footnote: keep humans meaningfully in the loop to preserve copyright, get clear rights or licenses to your training and input data, name human inventors on patents, and watch the case law — the ground is still moving.

What to actually do, by stage

The synthesis, stripped to moves you can act on. None of this is legal advice; it is the checklist that makes the eventual lawyer cheaper.

Day zero (incorporation):

• Get IP assignment agreements signed by every founder, and assign in any technology created before the company existed.
• Put assignment clauses in every employee and contractor agreement from the first hire — contractors especially.
• Clear and register your name and logo as a trademark in your real markets (and China), in the right classes.

Before you launch or pitch publicly:

• File provisional patents on anything genuinely novel and detectable — before the public talk or launch, because most countries have no grace period.
• Decide patent vs trade secret for each core technique; lock down secrecy (NDAs, access controls) for what you keep closed.
• Run a freedom-to-operate check in crowded fields.

As you grow:

• Within 12 months of your first filing, decide on PCT (patents) and Madrid (trademarks) to keep the world open; spend on national phase only where you can enforce or sell.
• Keep an open-source dependency inventory; avoid accidental copyleft in proprietary code.
• In any collaboration, paper background vs foreground IP before work starts; never accept default joint ownership by accident.
• If you implement a standard, budget for SEP/FRAND royalties.

When raising or selling:

• Expect IP due diligence: clean chain of title, a defensible position, freedom to operate, no open-source contamination, no live infringement risk. Fix gaps before diligence, not during.

When a demand letter arrives:

• Don’t respond off the cuff and don’t forward the patent around internally. Get counsel, assess validity (is it already in IPR?), and consider defensive options. Most disputes settle; panic is the expensive choice.

Intellectual property rewards the boring, early, cheap moves and punishes the missed ones. Sign the assignments, file before you talk, read the licenses, clear the brand. Do those four things and you’ve avoided the great majority of the ways IP destroys value — and kept the option to build the kind of position that creates it. This is general information, dated to 2026 and not legal advice; consult a qualified attorney in your jurisdiction for your specific situation.