Patent law only pays for outliers. Novelty and inventive step exist to screen out the expected, so a patent portfolio is, by design, a collection of exceptions.

Hold that thought against new research from Harvard, Wharton, Northwestern and Columbia, published in Harvard Business Review on 14 August 2026, on what generative AI actually does inside the innovation pipeline. The finding is uncomfortable: used naively, AI deepens the human bottlenecks it is meant to remove. In ideation, a model gravitates to the statistically typical answer – and once a person reads that answer, they fixate on it and produce fewer unusual ideas than they would have unprompted.

Because most models draw on similar training data, independent teams at different companies converge on the same handful of design directions: individual productivity rises while collective diversity falls. Screening compounds the problem, because committees mistake AI fluency for quality and fund the polished pitch over the original one. For anyone responsible for intellectual property, the translation is blunt. The statistically typical idea is the unpatentable idea.

An innovation pipeline that quietly optimises for the expected is manufacturing prior art, not property. And convergence carries a second cost: the same “invention” is being generated at the same time inside your competitors’ businesses, and in a first-to-file world, whoever turns it into a patent application first owns it.

What AI convergence means for invention harvesting and patent strategy

The durable lessons are about process design, not tool selection.

First, audit the output. Map the past twelve months of invention disclosures and patent filings against the landscape: if your applications cluster where everyone else’s do, AI has homogenised your R&D without anyone deciding it should – and using the tools was never the moat, as I argued in AI Isn’t Your Advantage—Your IP Strategy Is.

Second, redesign the funnel for outliers. The research shows that prompting a model deliberately toward bolder, more distinct territory works, while telling a fixated human to “think more broadly” does not – so build divergence into the machine step, score invention disclosures blind to their polish, and keep a human genuinely directing the inventive work. That last discipline is also what inventorship law demands of a patent that has to survive challenge.

Third, guard the loop itself. The researchers describe an automation trap in which ideation, screening and testing are all mediated by models trained on each other’s output, and the whole system drifts from reality – the same failure of human ownership I examined in AI Transformation Is Not a Strategy Problem — It’s an Ownership Problem. And when a capability is available to everyone, value moves to what cannot be downloaded – the pattern from When Anyone Can Build It in an Afternoon, What’s Left to Own?.

The closing test is simple. If your innovation pipeline would generate the same ideas as your competitor’s, you no longer have an R&D advantage – you have a subscription.

Read the article: Research: The Innovation Problems AI Can’t Solve, Julian De Freitas, Ayelet Israeli, Gideon Nave, Artem Timoshenko and Olivier Toubia, Harvard Business Review, 14 August 2026.