The Ghost in the Patent: Why AI Can Draft Documents but Cannot Make Judgement Calls
The Category Error: Fluent Intelligence and the Illusion of Understanding
Generative AI can now draft patent documents with striking fluency. Descriptions, embodiments, even claims can be produced in seconds. This surface-level linguistic competence has encouraged a dangerous inference: that patent practice is fundamentally a drafting task, and therefore largely automatable.
That inference is a category error. It confuses the what of patent practice (the production of text) with its why (the strategic intent that gives the text meaning). The true risk to the profession is not the emergence of a superior artificial intelligence, but the gradual displacement of human judgment. When practitioners use AI to replace thinking rather than to support it, they begin to treat strategy – the core professional skill – as a cognitive burden to be outsourced. Drafting is automated; judgment quietly follows.
AI fails at the core of patent strategy in two fundamental ways:
First, AI is not capable of situated decision-making. Patent strategy is never a purely abstract optimization problem. It is embedded in a concrete situation: a specific inventor and applicant, a specific competitive landscape, an examiner culture, a litigation horizon, a client’s risk tolerance, budget constraints, and long-term commercial intent. Human practitioners continuously integrate weak signals, tacit knowledge, and contextual cues that are not fully articulable and often not even consciously formalized. AI, by contrast, operates on representations detached from lived context. It can recombine patterns and propose plausible options, but it cannot inhabit the situation in which a strategic choice must be made. Its “judgment” is statistical, not situated.
Imagine it is 10:00 on a Monday morning. You receive an International Search Report (ISR) on a client’s flagship PCT application. The examiner has rejected claim 1 for lack of novelty but indicates that dependent claim 9 is allowable. An AI views this as a logic gate. Optimizing for “Grant,” it calculates the statistical success of various responses and offers a menu of binary options: fight the rejection with arguments, or accept claim 9 for a quick win.
The human practitioner sees a different landscape. They recall the difficult phone call with the client last week: the company’s Series B funding is contingent on a positive signal from the patent office immediately. They analyze the examiner’s objection to claim 1 and see that while it is based on a flawed interpretation that could be rebutted, this specific examiner is notoriously difficult to turn around in the international phase. The decision, therefore, is not a calculation but a maneuver. The attorney recommends limiting the international phase to claim 9 to secure a positive International Preliminary Report on Patentability (IPRP) to satisfy the investors, while quietly planning to fight for the broader claim 1 later, maybe via divisional applications in only the most critical target markets.
This is situated decision-making. The AI sees a probability of success; the human sees a conflict of values.
Second, AI cannot take responsibility for a decision. Strategy is not merely the generation of options; it is the commitment to one path over others, with full awareness that alternatives are being foreclosed. Every patent filing embodies such a commitment: claims not pursued, jurisdictions not entered, disclosures not made. These are normative decisions with legal and commercial consequences. AI cannot bear those consequences. It cannot answer for a lost enforcement opportunity, a fatal admission, or a misaligned filing strategy. Responsibility always attaches to a human agent – and that fact is not a temporary technological gap, but a structural feature of professional and legal practice.
When the human practitioner is reduced to validating machine output, they surrender what German philosophy aptly calls Geist (mind, spirit): the situated understanding, purposive intention, and willingness to stand behind a decision. This Geist is the true “ghost” in the patent. Like a mountain guide who turns back despite favorable forecasts because the snow “feels wrong,” the patent attorney’s irreplaceable role lies not in producing information, but in exercising judgment under uncertainty, and accepting professional accountability for the chosen course.
Functional Intelligence and the Trap of Satisficing
Patent practice is often taught and practiced as a form of expert analysis in a legal-technical “data space”. The patent attorney is trained to identify relevant facts, survey the legal terrain, enumerate risks, and present structured options. This was the model in which the author himself was trained almost two decades ago. The professional posture is analytical and technically rigorous, but deliberately non-directive: “Here are the examiner’s objections and three possible claim amendments. How would you like to proceed?”
This mode of practice reduces professional value to the ability to synthesize information and enumerate possibilities. Strategy becomes a menu. The attorney functions as an intermediary between complex systems – law, technology, procedure – organizing them into intelligible choice sets.
The problem is that this is precisely the domain in which AI excels. As Markus Gabriel observes in The Meaning of Thought (Polity, 2020), artificial systems possess functional intelligence: the ability to process data, detect patterns, and generate coherent outputs, without possessing Geist – situated understanding, intention, or meaning. In this sense, AI is not a deficient analyst; it is the ultimate data-space analyst.
When attorneys rely on AI in this purely analytical mode, they risk falling into what Advait Sarkar calls “outsourced reason” in Artificial Intelligence as a Tool for Thought (Microsoft Research, 2025): the gradual substitution of human judgment with machine-generated plausibility. The cognitive danger is not merely efficiency-seeking, but the phenomenon Sarkar identifies as satisficing. Because AI outputs meet a threshold of apparent correctness, they dampen critical engagement. The draft looks like a patent claim; it reads like a patent claim; therefore, it is treated as good enough. The traditional “blank page problem” is replaced by a more insidious “filled page problem”, in which the attorney shifts from active strategic construction to passive evaluation; reviewing, trimming, and approving a statistically probable draft rather than deliberately engineering a legally and commercially optimal one.
This is the quiet trap of functional intelligence. When professional value is defined as mapping the option space rather than choosing a path through it, the human practitioner becomes replaceable. The attorney has been acting as a cartographer – skilled, informed, and precise, yet ultimately detached from the responsibility of navigation itself.
The Mountain Guide: Identity and Accountability
The cartographer is obsessed with the document. Like a technician using a high-precision GPS, they use AI to generate technically accurate text and valid claims. When the GPS (AI) suggests a route – “Narrow claim 1 to overcome D1” – the Cartographer accepts it because the map is clean and the file is legally tidy.
A mountain guide knows that the map is not the territory. They do not merely display the map; they recommend the route. A hiker has little use for a guide who pauses at every junction to ask, “Which way would you like to go?” Guidance, by definition, entails direction. Patent strategy is no different. It is not a single calculation or optimization problem, but an act of navigation across what Markus Gabriel calls distinct fields of sense (Sinnfelder); irreducible domains of meaning.
To the mountain-guide patent attorney, these fields of sense present themselves as four winds that often blow in opposite directions:
- Legal Validity (The Rules)
- Technical Accuracy (The Facts)
- Commercial Utility (The Money)
- Psychology/Ethics (The Human Element)
Strategic decisions arise precisely at the intersections of these fields, where no dataset yields a uniquely correct answer. A broader claim may increase commercial leverage while heightening validity risk; a narrower claim may be easier to grant but commercially of limited value. Additional disclosure in the patent description may strengthen enablement but may map out a workaround path. These trade-offs cannot be resolved by factual analysis alone. They are judgment calls made under uncertainty, where gains in one dimension imply losses in another.
The mountain guide’s task is to synthesize these competing considerations and move from the neutral data space into what might be called the consequence space: the domain in which choices have irreversible legal and commercial effects. This is where AI cannot follow. While AI can model possibilities and rank alternatives, it cannot inhabit the future in which those alternatives will be tested – by examiners, competitors, courts, or markets.
This role is grounded in the concept of skin in the game articulated by Nassim Nicholas Taleb in Skin in the Game (Random House, 2018). Taleb argues that bureaucracy is a construction by which a person is separated from the consequences of their actions. AI is the ultimate bureaucrat. It produces text but bears no risk. The patent attorney, unlike an AI system, is a legally accountable and professionally insured actor who bears the consequences when a strategy succeeds or fails. Responsibility is not an administrative detail to be minimized; it is the source of professional legitimacy.
Recent empirical research supports this view. In Product Manager Practices for Delegating Work to Generative AI (2025), Mara Ulloa et al. introduce a “Selective Delegation Framework,” showing that professionals systematically refuse to delegate tasks where failure would impose personal or professional cost. Their conclusion is explicit: accountability must not be delegated to non-human actors. This finding aligns with Cristina DiGiacomo’s first commandment of human–AI coexistence: humans must own the outcomes of AI-assisted decisions. AI may generate content, but responsibility for its impact remains irreducibly human.
This distinction exposes a further limitation of AI in patent practice. Drafting competence is not strategic judgment. An AI does not experience a claim as a legal boundary that will be attacked, defended, or enforced. It does not experience uncertainty as risk. The mountain guide does. Like a guide who senses a subtle change in weather despite favorable forecasts, the patent attorney relies on situated judgment to decide when to proceed, and when to turn back.
This demands a shift in professional identity. Historically, the identity of the patent attorney was tied to the craft of drafting. As AI commoditizes that craft, professional value must migrate elsewhere – to the one element that cannot be automated or delegated: accountability. The future of the profession lies not in competing with machines at producing text, but in embracing the non-delegable burden of judgment and responsibility that machines, by their nature, cannot assume.
Four Guardrails for the Mountain Guide
Before proposing concrete guardrails, it is worth observing how professional authorities have begun to respond to generative AI in patent practice. The European Patent Institute (epi) has published Guidelines for the Use of Generative AI in the Work of Patent Attorneys (epi Information 4/2024), emphasizing essential duties: maintain client confidentiality, understand model limitations, verify outputs, and ensure that AI use never excuses errors or omissions. These guidelines rightly situate responsibility with the practitioner and warn against professional misconduct and data leakage.
However, their focus is largely preventive. They excel at identifying risks to professional conduct – confidentiality, quality assurance, client consent – but do not offer substantive guidance on how AI can be deployed to improve strategic outcomes for clients. They tell us what not to do, but not much about what good looks like when AI is used to elevate strategy rather than displace it.
The four guardrails below are offered to fill that void. They are not mere constraints on risk; they are disciplines that enable the attorney to engage AI as a tool for thought – preserving purposive human intention (Geist) and strengthening strategic value for clients.
Guardrail 1: Defining the Route (Metacognition)
Advait Sarkar warns in Artificial Intelligence as a Tool for Thought (Microsoft Research, 2025) against skipping metacognitive reasoning: high-level planning and mental decomposition that precedes execution. In patent practice, this means that the attorney must define the route before the machine writes the first word.
A common cartographer’s error occurs when strategic intent is silently delegated to the machine – for example, asking an AI to “write the claims” and then evaluating the result without first determining the optimal hierarchy of distinguishing features and fallback positions. In such cases, the attorney is no longer guiding the process; they are auditing it after the fact.
The mountain guide does the opposite. They perform the metacognitive work themselves: constructing the invention narrative, deciding which technical features carry commercial value, and determining how scope should be allocated across the main claims and dependent claims. Only then is AI engaged – to execute a human-defined plan, not to hallucinate one of its own. AI may assist in exploring alternatives, but it must not be allowed to decide what strategy is.
Guardrail 2: Feeling the Terrain (Material Engagement)
Fully outsourcing drafting risks the loss of what Sarkar calls material engagement: the hands-on cognitive interaction with the “clay” of one’s work. Without this engagement, the attorney becomes a “middle manager of their own thinking” – reviewing text they did not author and cannot deeply defend.
A cartographer may allow AI to generate dependent claims without ever inhabiting the reasoning that underlies them, only to find themselves unable – perhaps during oral proceedings – to explain why a particular fallback exists or what strategic function it serves. The words are present; ownership is not.
The mountain guide, by contrast, insists on touching the clay. They personally shape the protection hierarchy and scope boundaries before linguistic drafting begins. In the author’s own patent drafting workflow, this principle is enforced through a mandatory “Claim Strategy” step: a functional, high-level outline where the human manually refines the protection logic before any claim language is generated by AI. Text follows thought, not the other way around.
Guardrail 3: Forging the Non-Obvious Path (Combating Convergence)
Large language models are subject to what Sarkar calls mechanized convergence: a gravitational pull toward statistically average, widely represented solutions. This is not a defect; it is a structural feature of probabilistic systems.
In patent practice, however, this tendency is dangerous. In the author’s experience, state-of-the-art LLMs do not meaningfully distinguish between granted patents and published applications. A dominant signal in their training data therefore seems to be the surviving, narrowed version of an invention in a granted patent – the compromise that passed examination. This may help explain why claims generated by AI tools tend to drift toward a “safe” center: easily grantable, but commercially thin and readily designed around.
The cartographer accepts this comfort as prudence. The mountain guide recognizes it as a strategic failure. Valuable positions are often non-obvious, uncomfortable, and defensible only through skill. To counter convergence, the attorney must actively push the AI away from the statistical average, directing it toward maximum-scope positions that require human judgment to justify and human accountability to defend. To achieve this, the mountain-guide patent attorney can use the protocol of extremes. Do not ask for “the best claim.” Ask for the boundaries:
- Prompt A: “Draft a claim that is extremely broad and aggressive. Ignore validity risks.”
- Prompt B: “Draft a claim that is extremely narrow and virtually guaranteed to be granted.”
By forcing the AI to generate the boundaries, the attorney can navigate the space between them.
Guardrail 4: Sensing the Warning Signs (Productive Resistance)
Sarkar argues that AI should not function as an obedient servant, but as a provocateur—a source of productive resistance that prevents the user from moving uncritically through a task. Agreement is easy; challenge is valuable.
The cartographer treats a confident AI draft as a favorable weather forecast and proceeds accordingly. The mountain guide, by contrast, uses AI to surface warnings rather than reassurance. This can be implemented through a series of mandatory claim checks that act as critiques, not corrections. These provocations do not “fix” the draft; they force the attorney to confront weaknesses, trade-offs, and unresolved strategic questions.
In doing so, AI becomes a mirror for judgment rather than a substitute for it. The attorney is compelled to re-enter the consequence space – to engage in the legal reasoning, risk assessment, and responsibility-taking that no machine can assume.
Conclusion: Only the Ghost Can Decide the “Why”
The rise of generative AI exposes a fault line within the patent profession. Those who reduce their role to purely neutral facilitation – mapping options, summarizing objections, and asking clients “which way?” – are increasingly replaceable. These are the cartographers. By contrast, the mountain guides – those who synthesize technical, legal, commercial, and psychological realities into a recommended course of action – operate in a domain AI cannot enter.
Machines excel at producing the what: the text, the variants, the enumerated possibilities. They can even generate the what else: alternative claims, fallback positions, and plausible amendments. What they cannot do is decide the why. They cannot commit to a strategic boundary, inhabit the consequences of that commitment, or bear responsibility when the strategy is tested by examiners, competitors, or courts.
The “ghost” in the patent – the Geist of situated understanding, purposive intention, and accountability – remains irreducibly human. AI can assist judgment, challenge assumptions, and expand the space of possibilities. But it cannot assume authorship of strategy, because strategy is inseparable from responsibility.
AI will not replace the patent attorney who uses machines to think better. It will replace the one who uses machines to think less.
