Everyone Wants to Trademark AI. Most Won't Succeed.

In 2025, Amazon Technologies filed a US trademark application for a single word: AI. Two letters of plain vocabulary, claimed the way you might try to trademark "good" or "the". Amazon was not alone. 3,747 trademark applications filed that year put AI in the mark itself, and a long line of them, "Dynamo AI", "Pando.ai", "ThreatLens AI", "Bits AI", were companies racing to plant a flag on the same two letters.

Trademark filings are a slow, deliberate signal. A company pays a lawyer, picks a name, and commits to it months before a product ships. So a flood of AI names is a genuine bet on where business thinks the next few years are going. But there is a second story in the data, and it is the more interesting one: a large share of this gold rush is unlikely to ever register. You cannot easily trademark a description, and AI describes the product.

From 2016 to 2022, trademarks that put AI in the mark held remarkably steady, around 500 to 650 a year, roughly 1 in every 1,000 US filings. Then ChatGPT launched at the end of November 2022, and the line broke.

The easy assumption was that 2023 would be a one-off spike, the noise of a hype cycle. It was not. AI filings rose again in 2024, and jumped much harder in 2025. The 2025 total, 3,747, is more than double 2023 and just over seven times the pre-ChatGPT level. Because the register itself barely grew (total US filings drifted from 551,000 in 2022 to 600,000 in 2025), this is not a rising tide lifting everything. It is real concentration. AI went from 1 filing in 1,050 to 1 in 160.

The monthly view pins the timing. December 2022, the month after ChatGPT launched, was still flat at 44 filings, dead on that year's average. The climb starts in January 2023 and never properly stops. It steadied through 2024, then a second wave through 2025 pushed it past 380 a month. March 2026 set a record at 444, the busiest month for AI-named trademarks ever recorded.

Filing a trademark is not the same as owning one. Of the AI-named marks filed in 2022 and 2023, enough time has now passed to see how they landed: just 40% reached registration. For US trademarks overall in the same two years, the figure is 62%. AI-named marks fail at roughly 1.6 times the normal rate.

The reason is a basic principle of trademark law. A mark has to be distinctive, not merely descriptive of the goods. An examiner will not let a company fence off ordinary words that competitors need to describe their own products. Call an analytics tool "Smart AI Analytics" and the AI part is describing exactly what it does, so it is weak or unregistrable on its own. The more literally a name leans on AI, the harder it is to protect. Amazon's application for the bare word AI is close to the extreme case, the kind of mark examiners are most sceptical of.

One sub-trend shows the lesson being learned in real time. Marks using GPT as a word spiked to 41 in 2023, the year after GPT-4. OpenAI registered GPT-4 itself, examiners pushed back on the rest, and filings using GPT fell to around 20 a year. The GPT name-grab burned out fast, even as the broader AI label kept climbing. A good share of the companies filing today are paying lawyers to chase words they will not be allowed to keep.

The filings cluster heavily in one place. Nice Class 42, scientific and technology services, absorbed 5,485 of the AI marks filed since 2023, more than double the next class (Class 9, electronics and software). AI branding is, for now, overwhelmingly a software-and-services phenomenon rather than a consumer-goods one. Whether that holds as AI features spread into physical products is one of the things the next few years of this register will show.