Why Google Never Enforced the Transformer Patent — and Why No AI Tool Can Prove It 'Thinks Strategically'

I read something on LinkedIn the other day that stuck with me. The transformer is the architecture behind nearly all of today's AI. By any measure of value, it should be one of the most lucrative patents in history. Google built it and holds the patent (US 10,452,978, "Attention Is All You Need") — and yet you never hear about Google taking anyone to court over it.

The post made a sharp point about *proof*. A model's internals sit on a competitor's servers. There's almost no way, from the outside, to confirm how a rival model computes attention. To claim patent infringement you have to show the other side is using the elements of your claim — and elements buried somewhere you can't look into are hard to enforce, no matter how brilliant the invention. So the labs mostly don't patent the core method at all; they keep training recipes and model internals as trade secrets instead. When they do write AI patents, the question they ask first isn't "how impressive is the technology?" but "can it be proven from the outside?" The weight of the claim goes where reverse-engineering can reach — not the model's inner workings, but what goes in and what comes out. That line stuck with me all week, because I'd just spent seven months building an AI tool whose whole value lives in exactly that unprovable place.

I was an early adopter, at least by the date on the receipt — I started paying for ChatGPT the day after it launched. But I didn't think it could do the work I'd done for years, especially write the strategy behind a new-business pitch. Through most of last year, generative AI looked to me like an articulate parrot. It could say anything well; not all of it was usable.

Then a model came out that said things worth keeping, and I changed my mind. I figured AI could now write the strategy for a new-business pitch close to what a person writes. Caught up in the prompt-engineering wave, I expected that with the right prompt I could finally escape the pitch writing that had worn me down for more than 25 years. That expectation lasted about as long as a daydream.

Writing the prompt was the first wall. I couldn't get a feel for what to type, what information to feed it, to get an AI to write a pitch strategy. When I got past that, the output was the next problem. The word choices were all over the place. The phrasing swung between polished and clumsy. The overall structure was bolted together with half the screws missing, and the flow didn't connect.

After wrestling with it for more than seven months, I got to the point where the prompt no longer needs babysitting. The way an account lead or creative director builds the story to *sell* the creative and the message to a client — I got an AI to produce a new-business pitch strategy in that mode, as a file. Creative, copy, digital and performance plans, execution programs — there are already plenty of automation tools for those, so I built it to do only the one thing: the strategic story.

Plenty of AI tools say they think strategically, or judge like a consultant. But that "thinking and judgment" is buried deep in the model, where no one outside can confirm it. Synapsis isn't exempt from this. By the patent's own lesson, it's an unprovable area — a part I can't open up for inspection. One thing I can say for certain: with the same input, Synapsis produces what a general-purpose AI can't — because it was designed to.

But my saying so is, in the end, just a claim. By the lesson of that patent I started with, I can't prove to you what happens on the inside either. (How the AI AE works is written up separately.) Whether the same brief really produces a different result — you can check that yourself. Run the same brief through a general-purpose AI and compare.

Read more

• How the engine turns a brief into directions: Can AI write a creative strategy for a pitch?
• The same brief, run two ways: Can ChatGPT write a pitch strategy? We ran one brief
• Why the design, not the model, decides quality: Same AI model, very different results

Frequently asked questions

Does Google hold a patent on the transformer?

Yes — US 10,452,978, "Attention Is All You Need," covers the transformer architecture behind most modern AI. Google has signaled it won't assert the patent aggressively, and there's been no enforcement. The architecture is published and used almost universally.

Why doesn't Google enforce the transformer patent?

Partly because it's published and ubiquitous, and partly because of a proof problem: a competitor's model internals aren't observable from the outside, so proving infringement is nearly impossible. For that reason AI labs tend to protect model internals as trade secrets rather than patents.

How does this apply to AI tools that claim to "think strategically"?

The same proof problem. Any claim about what happens inside a model — strategic reasoning, consultant-like judgment — can't be verified from the outside. So the honest test isn't the claim; it's the output. Run the same brief through two tools and compare what comes back.

Can you prove Synapsis "thinks" strategically?

No, and we don't claim to. What happens inside the engine is as unprovable as any model's internals. What's checkable is the result: with the same brief, Synapsis produces strategic directions a general-purpose AI doesn't — and you can verify that side by side.