Kind Games, Wicked Law: Centaur Patent Attorneys and Lessons Learned from Deep Blue

In 1997, IBM’s supercomputer Deep Blue defeated reigning world champion Garry Kasparov in a six-game chess match. But the victory was not a simple triumph of silicon over gray matter. Between each of the six games, IBM engineers spent hours tinkering with the code to fine-tune the algorithm and account for Kasparov’s tactical shifts. Together, the IBM engineers and digital Deep Blue formed a human-artificial intelligence (AI) hybrid that became known as an example of a “centaur,” a term based on the Greek mythological creature with the upper body of a human and the lower body of a horse. While today’s large language models (LLMs) dwarf the processing power of Deep Blue [1], this 30-year-old centaur example provides lessons for patent attorneys seeking to embrace current AI tools. The near-term future of patent prosecution belongs to those who can utilize AI as the engine but maintain a human hand on the steering wheel—knowing when to override the machine to win the “wicked” game of patent prosecution.

To understand prosecution through the lens of the centaur model, we should distinguish between “kind” and “wicked” learning environments. In David Epstein’s book “Range: Why Generalists Triumph in a Specialized World,” chess is a kind environment: the rules are fixed, all information is visible and accurate, and feedback is instantaneous. Conversely, the environment of patent prosecution is wicked. Rules change overnight (think cases like Mayo v. Prometheus (2012) or KSR Int’l Co. v. Teleflex Inc. (2007)), information is hidden (the black box of examiner assignments), and feedback delays may last months or years. Epstein argues that, in wicked environments, generalists outperform hyper-specialists by connecting disparate ideas to adapt to novel situations. In this landscape, the machine’s drive to maximize the probability of a win (e.g., allowable subject matter) in the current “move” (e.g., round of prosecution) must be tempered by practical considerations including business objectives, budgetary limits, and the psychological nuances of the human audience.

While AI tools excel at “material gain,” such as maximizing the probability of an allowance or drafting claims that have a high semantic similarity to disclosure documents, they generally lack the contextual information needed to balance multiple strategic objectives simultaneously. Humans access crucial external sources of information: a startup’s need for deliverables before a funding round, or a flagship product’s need for maximum scope regardless of the number of costly Requests for Continued Examination (RCEs). Much like how the IBM team tuned Deep Blue to decline poisoned pawns (and associated material advantage) in favor of freedom of movement and territorial control of the board, access to contextual information enables a seasoned attorney to balance competing objectives. While machines might celebrate allowances, humans recognize the strategic loss of an unenforceable or brittle claim. By overriding the machine’s mathematical evaluation with a long-term strategic valuation, a centaur patent attorney can synthesize these real-world signals to align prosecution with a client’s economic goals.

The centaur patent attorney can combine machine logic with human persuasion. An AI prioritizes semantic consistency, yet patent prosecution is a human game. In game two of his match against Deep Blue, Kasparov resigned in a position that could have resulted in a draw. In that game, he was not defeated by the mathematical certainty of a checkmate, but because Deep Blue’s play felt uncharacteristically human. Meanwhile, a patent prosecutor’s arguments must convince another human (e.g., an examiner, a supervisory patent examiner (SPE), or a judge), not a computer. For example, AI can generate a technically sound motivation to combine argument, but a seasoned attorney understands that the vast majority of examiners give such arguments little weight.[2] A centaur patent attorney recognizes this disconnect and overrides the machine’s logical output, pivoting instead to arguments that resonate with human psychology and are thus more likely to move the needle with the intended audience. As AI is increasingly used throughout patent preparation and prosecution, the centaur patent attorney must use their judgment to adjust for real-world arguments and strategic objectives.

Some might argue that the wickedness of the patent prosecution may be solved by data analytics—that tracking examiner allowance rates and citation histories might turn prosecution into a kind game of predictable probabilities. However, such data can only map the past; it cannot predict new case law or a sudden pivot in a client’s business model. In a wicked environment, the rules are unstable. An AI might calculate a 70% probability of success based on historical data, but a centaur patent attorney understands that a recent Federal Circuit footnote or a client’s new CEO could change everything. A centaur patent attorney does more than merely use data to predict the path; they use their human experiences and judgment to decide if that path is still worth taking.

To manage client-specific context across different platforms, consider creating a custom version of the model, such as a Gem in Google’s Gemini, a Custom GPT in OpenAI’s ChatGPT, or a Claude Project in Anthropic’s Claude. These custom instances of a language model can uniformly consider client background, context, and technical standards. Alternatively, you can maintain a single master conversation thread for one client to ensure total context bleeds across all their inventions. This allows the AI to reference previous work seamlessly, but the downside is that long threads can eventually hit a context limit where the AI begins to lose focus or forget earlier details. As a further alternative, you can use separate chat threads combined with a context fact sheet that you paste in at the start of every session. This enables the cleanest organization and most distinct file history, but the downside is the repetitive manual work of re-introducing the client’s data each time you start a new project and need to update the context manually across all threads.

Ultimately, the “Deep Blue era” of patent law is not about replacing attorneys with AI, rather it is about evolving human attorneys to guide those AIs. As the drudgery of technical drafting is reduced by AI, the value of the centaur patent attorney lies in their ability to act as a high-level curator and strategist. AI tools produce results that may appeal to a naïve audience, but the legal arguments and strategic objective/valuation may not dovetail with client needs. Attorneys must resist automation bias, the temptation to defer to the polished, logical output of the machine, and instead embrace the messy, wicked realities of human-centric patent prosecution. By combining the brute-force processing of AI with strategic data and the psychological intuition of a human generalist, we ensure that the patents we secure for clients are valuable assets in a human-led world.

[1] Indeed, smartphones produced in 2026 have more computational power than Deep Blue.
[2] Remember, the Federal Circuit held that laboratory chemists would look at soda-pop bottle caps to “solve problems with flash chromatography cartridges.” Scientific Plastic Products, Inc. v. BIOTAGE AB, 766 F. 3d 1355, 1360 (Fed. Cir. 2014). Also, the Federal Circuit has blessed thirteen reference §103 rejections. See In re Gorman, 933 F. 2d 982 (Fed. Cir. 1991).