By Michael Borella –

Grab a pencil.  Now train a large language model.

Not a big one, mind you.  Let’s go small, something in the range of a few billion parameters, trained on a modest corpus of a trillion or so tokens.  Conservatively, that run requires something on the order of 10²² floating-point operations.[1]  If you can perform one multiply-and-add every ten seconds, a respectable pace for careful long multiplication, and you work without breaks, sleep, or the occasional existential crisis, you will finish in roughly 3 quadrillion years.  The universe is about 14 billion years old, so you may want to start now.

Absurd?  Of course.  And yet, under the dominant reading of § 101 case law, that training run is a “mental process” because nothing in it, step by step, is something a human could not theoretically do.  Each multiplication is trivial.  Each addition is trivial.  Aggregate enough trivial operations and, according to this logic, you have a trivial whole.

This twisted interpretation has real consequences.  Examiners have issued, and courts have affirmed, mental process rejections against claims that no human being could perform within the lifetime of the species.  The ultimate solution is to get rid of the Alice v. CLS Bank Int’l framework altogether and return to the statutory regime, letting §§ 102, 103, and 112 do the work of narrowing overbroad claims.  A claim to an old algorithm fails under § 102.  A claim to an obvious one fails under § 103.  A claim that is unbounded, indefinite, or untethered from its disclosure fails under § 112.  Those sections of 35 U.S.C. were built for exactly this job, and they do it with evidence rather than intuition.

But until the Supreme Court or Congress takes that step, we are stuck with Alice, and we should at least try to make it coherent.  The problem is not just that the mental process doctrine exists, it is also that the doctrine has no ceiling.  In this article, I propose one.

How We Got Here

The mental process exception has questionable roots.  In Gottschalk v. Benson, the Supreme Court observed that the claimed conversion of binary-coded decimal to pure binary could be done mentally, and grouped mental processes with laws of nature as the basic tools of scientific and technological work.

But that observation was not the main driver of the opinion.  The Court’s real concern was preemption, in that the claims covered every use of the algorithm on any general-purpose computer.  The Court went out of its way to say it was not holding computer programs unpatentable, and it suggested that Congress was better placed to settle the question.  However, it proceeded to issue its opinion even though the claims might have been better of remanded to district court for considerations of obviousness.  Moreover, the mental process doctrine that Benson appeared to revive had already been largely dismantled two years earlier by the Court of Customs and Patent Appeals in In re Musgrave.  The Supreme Court ignored the appellate court’s careful reasoning and replaced it with an unworkable and unbounded principle.

The Federal Circuit then gave the concept its modern shape. CyberSource Corp. v. Retail Decisions, Inc. relied on a pen-and-paper test, holding a method of detecting credit card fraud unpatentable because its steps could be performed by a human with pen and paper. Electric Power Group, LLC v. Alstom S.A. extended the idea, treating the collection, analysis, and display of information as essentially mental processes.

The trouble metastasized when these holdings collided with a separate line of cases, typified by Bancorp Services, L.L.C. v. Sun Life Assurance Co. of Canada, holding that performing an otherwise ineligible process faster on a computer does not make it eligible.  Put the two together and you get a syllogism that has haunted software prosecution for a decade.  That is, (1) any algorithm can, in principle, be performed by hand; (2) a computer merely performs it faster; therefore (3) any algorithm is a mental process.  Under this logic, the only limit on the mental process grouping is an imaginary human with infinite time and a lot of paper.  In other words, there is no limit.

The USPTO has pushed back mildly.  Its 2019 Revised Guidance organized abstract ideas under § 101 into three groupings, one of them mental processes.  The 2024 AI Subject Matter Eligibility Update stated the mental process grouping is not without limits.  Most recently, the August 4, 2025 memorandum from Deputy Commissioner Charles Kim told examiners that a claim does not recite a mental process when it contains limitations that cannot practically be performed in the human mind.  Okay, but not especially helpful.

Then came Ex parte Desjardins.  In September 2025, an Appeals Review Panel that included Director John Squires vacated a § 101 rejection of claims to training a machine learning model on sequential tasks without catastrophic forgetting.  The panel agreed the claims recited mathematical calculations, but held that they improved how the model itself operates, and it took pains to say that §§ 102, 103, and 112, not § 101, are the proper tools for policing claim scope.  Tellingly, the panel left the Board’s obviousness rejection intact.  The claims still failed, but for the right reason.  The decision was designated precedential that November, and a December 2025 memorandum wrote it into M.P.E.P. § 2106.04(d).

“Practically” Meaningless

The USPTO’s guidance is a welcome correction, but it hangs on a single adverb.  What does it mean for a step to be practically performable in the human mind?  An hour of work?  A year?  A career?  The guidance does not say, and so the answer depends on the examiner, the art unit, and, frankly, the day of the week.

The courts have not filled the gap.  The Federal Circuit has occasionally recognized that some tasks are beyond human capability.  In SRI International, Inc. v. Cisco Systems, Inc., the Court found that the human mind is not equipped to detect suspicious activity by monitoring network traffic in the claimed manner.  But the Court has never said where the line falls, and it continues to affirm ineligibility for claims whose only alleged advance is speed or scale. Recentive Analytics, Inc. v. Fox Corp. held that applying generic machine learning (a complex series of calculations that cannot be performed mentally in any reasonable amount of time) to a new data environment is ineligible absent an improvement to the machine learning itself.

I want to be careful here. Recentive was decided at step two, not on mental process grounds.  The question of whether a claim recites a mental process is analytically separate from whether a claim that recites one is nonetheless integrated into a practical application or amounts to significantly more.  My concern is narrower.  The first question should have a definite answer for the majority of claims, and right now it does not.

The Bancorp “mere speed” principle is sound when applied to its facts.  If a claim recites a process that bank clerks performed by hand before computers came along, a computer doing it faster adds nothing.  But that principle quietly assumes the process was humanly performable to begin with.  When the process could never have been performed by any human, at any speed, there is nothing for the computer to accelerate.  The computer is not doing a human task faster.  It is doing a task that only a machine can do.

The Proposal: A Human-Feasibility Threshold

Failing to render Alice a misguided and sad footnote in the history of innovation, here is the rule I would adopt:

A claim recites a mental process only if a single person, using pencil and paper, could perform the smallest embodiment the claim requires (1) within one working year of total effort, and (2) fast enough to meet any deadline the claimed process actually depends on.

Think of it as two clocks.  The first measures how much work the claim demands in total.  The second measures how quickly the claim demands an answer.  A claim that fails either clock is not something a human mind can practically perform, and it should not be treated as one.

To make the first clock administrable, we need a unit of measure.  I propose elementary computing operations such as one arithmetic step, one comparison, one memory transfer.  A person working carefully with a pencil, writing down intermediate results and checking the work, needs about ten seconds for each.  A working year is about 2,000 hours, or around 7,200,000 seconds, which comes to roughly 720,000 operations.  Rounding to a clean power of ten, the threshold is 1,000,000 elementary computing operations.

I would pair that ceiling with a rebuttable middle zone as shown in the table below.

Minimum required operations	Human equivalent	Treatment
Below 100,000	Less than about one working month.	May recite a mental process; ordinary analysis applies.
100,000 to 1,000,000	One month to one year.	Presumed not a mental process but rebuttable by identifying a human-feasible embodiment within the claim’s scope.
Above 1,000,000	More than a working year.	Not a mental process as a matter of law.

Total effort is only half the story.  Plenty of computations are small enough for a person to grind through eventually but useless unless they finish within milliseconds.  An anti-lock braking controller might need only a hundred or so operations to decide whether to release brake pressure, but it needs that answer every ten milliseconds, while the car is skidding.  A person with a pencil would finish the first cycle about seventeen minutes after the crash.

The second clock captures this.  Using the same unit of measure, that a person needs about ten seconds per elementary operation, the human time to produce one result is roughly ten seconds multiplied by the number of operations that result requires.  Even a single trivial response has a floor in that simple human reaction time is around 200 milliseconds, and no amount of practice gets below it.  If the claimed process requires each result faster than that human time, the claim is not a mental process, no matter how few operations it involves.

The obvious objection is that this lets applicants escape the grouping by tacking “within one millisecond” onto a list of operations.  So the deadline must be functional, one the claimed process actually depends on to work, not a speed requirement added for its own sake.

The test is would be simple.  Does a slower result defeat the purpose of the claimed invention?  A braking decision that arrives after the wheel locks is worthless.  A packet filter that decides after the packet has been forwarded has filtered nothing.  A control signal that arrives after the robotic arm has moved past the target is just a log entry.  By contrast, a fraud score that arrives in fifty milliseconds rather than fifty minutes is merely faster, and faster is not enough.

Functional deadlines can be derived from two places.  Some are explicit, such as a claim requiring a response within a stated interval tied to a physical or protocol event.  Others are inherent, arising because the claimed process controls a physical system, acts on data that exists only transiently, or interacts with another machine that imposes its own timing.  Bare recitations of “real time” get no special treatment.  They are construed in light of the specification like any other term, and if the specification does not tie them to a deadline the process depends on, they do not start the second clock.

The two clocks also reinforce each other.  A claim requiring continuous operation accumulates effort quickly.  That braking controller, running a hundred cycles per second at a hundred operations per cycle, blows through the one-year effort threshold in under two minutes of driving

Guardrails

Three further features keep this from becoming a mere counting exercise.

First, both clocks operate with respect to the claim and specification as written.  The relevant burden comes from explicit limitations (model size, data volume, iteration counts, response deadlines) or from limitations that the specification makes inherent under the broadest reasonable interpretation.  If a claim reads on a ten-parameter toy model with no deadline, the ten-parameter toy model with no deadline is what we evaluate.  Applicants who want the benefit of scale or speed must make it clear.

Second, a capacity consideration catches what the clocks miss.  A claim requiring simultaneous handling of data that a human cannot hold or perceive, such as raw sensor waveforms or state spanning millions of records, is outside the grouping even if the operation count is modest and no deadline applies.

Third, an anti-dissection rule.  The challenger may not first strip away limitations as “generic computer components” and then evaluate only what remains.  Both clocks apply to the entire claim, not just aspects cherry-picked out of it.  This rule is consistent with the Kim memorandum’s instruction to analyze claims as a whole.

Taking it For a Spin

A few hypothetical claims show how the test plays out.  For each, the table below gives a rough, order-of-magnitude estimate of the operations required by the smallest embodiment the claim covers, along with any functional deadline.  The results line up about where intuition says they should.

Hypothetical claim	Operations per result	Functional deadline	Result
Flag a transaction if its IP address matches one used by a prior fraudulent transaction (CyberSource-style).	100 to 1,000	None	Mental process
Same fraud check, required “within 50 milliseconds” with no reason tied to how the process works.	100 to 1,000	None (speed for its own sake)	Mental process
Classify one 28 × 28 image with a single-layer linear classifier.	About 10,000	None	Mental process
Same classifier, rejecting defective parts as they pass a sensor on a production line at ten parts per second.	About 10,000	About 100 milliseconds	Not a mental process (second clock)
Decide whether to release brake pressure from wheel-speed readings, every 10 milliseconds.	About 100	About 10 milliseconds	Not a mental process (second clock)
Inspect each packet on a network link and drop it before forwarding if it matches an attack signature.	About 1,000	Microseconds	Not a mental process (second clock)
Generate one output token from a model recited as having at least one billion parameters.	About 2 billion	None	Not a mental process (first clock)
Pretrain a transformer language model on at least one billion tokens.	1 quadrillion and up, even for a tiny model	None	Not a mental process (first clock)

Three results deserve comment.

The simple image classifier lands in the mental process zone, and that is the right answer.  About eight thousand multiply-adds at ten seconds each is roughly 22 hours of work: a tedious three days, no doubt.  Put the identical computation on a production line, though, and the answer flips.  A person would classify a part that is gone in a tenth of a second.  Same math, different invention.  The second clock is what lets the doctrine honor the real-word difference.

The fraud rows show the other side of the line.  Adding a fifty-millisecond deadline to a CyberSource-style check changes nothing, because nothing about the check depends on that speed.  That is Bancorp doing its proper job.

Finally, under my test, Desjardins-style claims escape the mental process grouping only if they recite, explicitly or inherently, enough scale or a functional deadline to start one of the clocks.  A claim to training “a machine learning model” with no size limitation reads on a trivially small model.  That is not a bug. Desjardins itself was decided on practical application grounds, which is the right place for a claim whose contribution is a better training technique rather than sheer computational heft.  My proposal does not displace that analysis.  It simply makes it harder to mischaracterize what the claim actually protects under prong one of Alice.

Anticipating the Objections

“Applicants will just add scale limitations.”  Yes, and good.  A claim reciting a billion-parameter model is narrower than one that does not.  We already accept this dynamic with specific data structures and concrete technical parameters.  Narrowing a claim to the embodiments that actually require a machine is the kind of drafting behavior we should encourage, not punish.

“The second clock is just speed in disguise.”  It would be, without the functional deadline requirement. Bancorp says that doing a human task faster on a computer adds nothing.  But a functional deadline does not make a human task faster.  It helps define tasks no human can perform at all, because the answers are worthless after the clock expires.  A person who computes a braking decision too late has not performed the claimed process slowly.  They have simply not performed it.

“Everything can be framed as time-sensitive.”  Applicants will try, and examiners and judges should push back.  The question is whether a slower result defeats the claimed purpose, and the specification has to show why it would.  A deadline that appears only in the claims, with no explanation of what breaks if it is missed, is speed for its own sake and gets no credit.

“The numbers are arbitrary.”  Somewhat.  Any bright line is.  But ten seconds per operation and 1,000,000 operations are not pulled from thin air.  Together they describe a person’s working year expressed in arithmetic steps.  Reasonable people could argue for a working month or a full career instead.  What matters more is that there is a number, so that examiners, the PTAB, and courts are applying the same test rather than their intuitions about what feels like thinking.

“Escaping the mental process grouping does not make a claim eligible.”  Correct, and I am not suggesting otherwise (for now).  The mathematical concepts methods of organizing human activity groupings remain.  We can save discussion of their silliness for another day.

“Why a single person and not a team?”  Because the doctrine speaks of the human mind, not a committee of them.  Allowing teams invites the absurd hypothetical of ten million clerks working in parallel, which is just a computer made of people and anything but practical.  It would not help on the second clock anyway.  Ten million clerks still cannot pass a note faster than one of them can react.

Conclusion

The mental process grouping exists to keep people from patenting thought.  But a doctrine that classifies training a language model or stopping a skidding car as thought has lost track of what thought is.  The fix proposed here is to provide limits.

So put down the pencil.  You were never going to finish that training run anyway, and the car has already hit the wall.

---

[1] Based on calculations set forth in Jared Kaplan et al., Scaling Laws for Neural Language Models (arXiv:2001.08361, Jan. 23, 2020), https://doi.org/10.48550/arXiv.2001.08361 (estimating total training compute as approximately 6ND, where N is the number of model parameters and D is the number of training tokens).