Are Software Patents Harder to Get Since Alice?
The definition decides the answer. Among 1,308,614 US utility applications filed 2020 to 2022, software defined narrowly was granted 75.8% of the time against 76.4% for everything else, which is the same rate. Define software one class wider and it was granted 79.3% against 75.5%, which is better than everything else. Business methods are the real exception. That group average also hides a split: applications whose first Office action raised section 101 were granted far less often than those that never met one, and software’s ordinary overall rate rests on the two thirds that never did.
The belief being tested here is a common one, and it bundles several claims together: that software patents are harder to get and harder to enforce since Alice, that claims now have to be tied to a concrete technical improvement, and that a patent is a strong position rather than an automatic win. Those are different claims, and the patent office record can answer some of them and not others.
| The claim | Verdict | What the record shows |
|---|---|---|
| Alice raised eligibility rejections | True, per the USPTO | The USPTO's own study found the chance of a first-action 101 rejection rose 31% in 33 affected technology areas in the 18 months after the decision. |
| Harder to get, for software overall | Depends on the definition | Narrowly defined, 75.8% granted against 76.4%. Adding networks and image processing, 79.3% against 75.5%. |
| Harder to get, for business methods | True | Business-method applications (G06Q) were granted 57.2%, and 38.3% ended without a patent. |
| A harder road to the grant | True | 70.9% of non-final actions on AI applications carried a 101 rejection in July 2025, against 16.0% across all fields. |
| Claims must show a concrete technical improvement | Not tested | A statement of the legal standard, which this record cannot test. Examiners do write in the Guidance's terms: "practical application" appears in 81.2% of 101 rejections in 2025. |
| Harder to get, once a 101 rejection is raised | True | Of 2020 filings in narrow software, those whose first action raised 101 were granted 60.0% of the time against 85.5% for those that did not. The same gap appears in 2021 filings. |
| Harder to enforce | Not tested | This record covers the patent office only. Nothing here measures litigation or enforcement. |
The definition decides the answer
There is no official list of software patents. Any answer depends on which CPC classes you count, and the two defensible choices point in opposite directions.
| Definition of software | Applications | Granted | Everything else |
|---|---|---|---|
| Narrow: G06F, G06N, G06Q | 170,853 | 75.8% | 76.4% |
| Broad: adds H04L, G06T, G06V | 262,810 | 79.3% | 75.5% |
Neither number is wrong. The narrow definition keeps business methods in and leaves networking out, and the broad one does the reverse. Which way that comparison points rests entirely on which of those two lists the speaker has in mind, and most people stating the claim have not picked one. The road to the grant is a separate question, and the answer there does not depend on the definition.
What the classes actually look like
Pooling them hides the spread. The classes usually called software do not behave alike, and the gap between the best and the worst is wider than the gap between software and everything else.
G06N is the one to read carefully rather than quickly. Its 64.0% is not a measured failure rate: 15.7% of those applications were still pending on the data date, the highest share of any class here, because machine-learning filings are recent and many have not finished. Among the G06N applications that have resolved, 75.9% were granted. The class share can only hold at 64.0% or climb as the rest finish, since a pending application can still go either way, and the ones still pending are not a random sample of the class, so 75.9% is not a forecast either.
Classed by the first CPC symbol only, so an application with claims spanning two classes is counted once, in one of them. H04L, G06T and G06V mix software with hardware, which is exactly why the definition above changes the answer.
Business methods are the exception
One group does behave the way the belief describes, and it is the group Alice was about. Following the business-method art units themselves rather than a CPC class, the fall and the partial recovery are both visible.
| Filing year | Business-method art units | Rest of center 3600 | Applications |
|---|---|---|---|
| 2004 | 47.2% | 67.1% | 8,782 |
| 2005 | 44.5% | 62.6% | 8,443 |
| 2006 | 40.8% | 61.4% | 8,019 |
| 2007 | 37.0% | 63.3% | 9,256 |
| 2008 | 38.5% | 66.8% | 9,036 |
| 2009 | 39.8% | 69.6% | 8,020 |
| 2010 | 43.4% | 70.7% | 9,125 |
| 2011 | 37.6% | 71.3% | 9,472 |
| 2012 | 34.5% | 73.9% | 11,375 |
| 2013 | 28.8% | 75.7% | 13,344 |
| 2014 | 31.0% | 76.7% | 13,226 |
| 2015 | 38.0% | 77.9% | 11,126 |
| 2016 | 42.0% | 79.9% | 10,158 |
| 2017 | 46.2% | 81.0% | 10,457 |
| 2018 | 48.9% | 80.0% | 10,789 |
| 2019 | 53.3% | 80.2% | 12,328 |
Applications filed in 2013 were mostly still pending when Alice was decided in June 2014, and fewer than three in ten were ever granted. By 2019 filings the rate had recovered to 53.3%, still well below the 80.2% of their neighbours in the same technology center.
The gap to the rest of center 3600 widened by 15.8 points between 2008 to 2010 filings (28.4 points behind) and 2012 to 2014 filings (44.2 points behind). Reading the series year by year, which was done after that window comparison rather than planned, it was already widening before then: 19.9 points in 2004, 33.7 by 2011. A step that size by 2011 fits Bilski v. Kappos (June 2010) at least as well as anything later, so the series shows timing and isolates no single decision.
Filing year is not examination year. An application filed in 2012 was usually examined in 2014 to 2016, so a filing-year series lags the decision it is being read against. Nothing here shows that Alice caused the change: the timing is consistent with it, with Bilski before it, and with the 2019 eligibility guidance after it.
Where the belief is right: the road, not the destination
The grant rates above describe where applications end up. They say nothing about what it takes to get there, and that is where the belief holds up best. Eligibility rejections are heavily concentrated in software and AI.
The same measure, month by month
One month is a snapshot. The series is what shows the arc, and it is the reason this page treats the road and the destination as separate questions.
| Month | All fields | Center 2100 | Center 3600 | AI (G06N) | AI actions |
|---|---|---|---|---|---|
| 2024-01 | 11.9% | 30.8% | 28.0% | 53.7% | 149 |
| 2024-02 | 12.6% | 31.2% | 29.7% | 57.9% | 195 |
| 2024-03 | 12.9% | 32.8% | 30.5% | 53.5% | 217 |
| 2024-04 | 12.9% | 33.7% | 28.7% | 53.5% | 155 |
| 2024-05 | 12.8% | 32.4% | 30.1% | 52.5% | 240 |
| 2024-06 | 13.8% | 33.6% | 31.8% | 57.4% | 223 |
| 2024-07 | 13.4% | 34.6% | 30.4% | 62.5% | 296 |
| 2024-08 | 13.8% | 35.1% | 32.4% | 59.2% | 434 |
| 2024-09 | 14.5% | 36.1% | 33.5% | 60.2% | 532 |
| 2024-10 | 14.0% | 35.8% | 33.3% | 61.8% | 191 |
| 2024-11 | 14.3% | 36.3% | 32.4% | 58.8% | 177 |
| 2024-12 | 15.0% | 37.1% | 34.9% | 59.1% | 362 |
| 2025-01 | 13.9% | 39.1% | 32.4% | 61.8% | 254 |
| 2025-02 | 14.2% | 41.4% | 32.1% | 62.0% | 234 |
| 2025-03 | 14.8% | 41.8% | 34.5% | 63.9% | 631 |
| 2025-04 | 15.0% | 42.9% | 34.0% | 64.9% | 541 |
| 2025-05 | 16.0% | 46.2% | 35.1% | 65.3% | 738 |
| 2025-06 | 16.0% | 46.3% | 35.6% | 66.9% | 706 |
| 2025-07 | 16.0% | 48.2% | 34.5% | 70.9% | 746 |
| 2025-08 | 15.4% | 44.7% | 34.3% | 63.4% | 636 |
| 2025-09 | 14.7% | 39.1% | 32.6% | 61.0% | 613 |
| 2025-10 | 13.7% | 35.5% | 31.4% | 54.1% | 665 |
| 2025-11 | 14.7% | 41.7% | 32.3% | 60.2% | 548 |
| 2025-12 | 14.4% | 37.8% | 31.5% | 56.8% | 639 |
Monthly AI counts run from 149 to 746 actions, so a single month moving a few points is noise. The arc is not: the AI share ran 53.7% in 2024-01, peaked at 70.9% in 2025-07, and was 56.8% by 2025-12.
Once raised, eligibility also tends to stay. Of first non-final rejections citing section 101, the ground came back in the next action 56.2% of the time overall, and 76.9% of the time in transportation and e-commerce, the center that contains the business-method art units.
A rejection is not an outcome. These figures count how often eligibility is raised and how often it returns. They do not follow what became of the applications that received one. The class-level grant shares higher up pool applications with and without a 101 rejection, so a field can carry both the highest rejection rate and an ordinary grant share without this record showing how those two facts fit together. The outcome of the rejected applications themselves is measured further down, under After a 101 rejection, and it is not the same answer.
After a 101 rejection
Everything above counts how often eligibility is raised. It does not follow the applications that met one, and until now this page said so and left it there. That question now has an answer, and it changes the shape of the page's own conclusion.
So the page's own answer needs splitting in two. "A harder road, an ordinary destination" holds for software as a group, and does not hold for the applications that actually meet an eligibility rejection. Both are true at once, because the applications that never meet one are granted better than average.
| Class | First actions with 101 | Met a 101 | Did not | Difference | 2021 |
|---|---|---|---|---|---|
| Narrow software | 33.0% | 60.0% | 85.5% | -25.5 | -26.0 |
| G06Q | 61.9% | 46.1% | 79.0% | -32.9 | -33.4 |
| G06N | 42.7% | 64.5% | 84.7% | -20.2 | -22.6 |
| G06F | 21.5% | 72.2% | 86.6% | -14.4 | -13.0 |
| G06V | 17.0% | 73.8% | 84.1% | -10.3 | -9.1 |
| H04L | 12.8% | 82.5% | 88.7% | -6.2 | -6.9 |
| G06T | 16.2% | 85.0% | 87.1% | -2.1 | -5.0 |
| All other classes | 7.5% | 66.4% | 79.1% | -12.7 | -13.7 |
| All utility | 10.9% | 65.3% | 80.1% | -14.8 | -15.9 |
Class mix explains some of the all-utility gap and not most of it. Reweighting the applications that met a 101 rejection to the class mix of those that did not moves their rate from 65.3% to 71.5%, against 80.1%. About six of the fifteen points are the kinds of application that attract eligibility rejections; the rest sits inside the classes.
Whether it is dropped matters about as much as whether it was raised
An eligibility rejection in a first action is not a verdict. The next action is where it either goes away or does not.
A kept rejection does arrive with more rounds of prosecution, which is the obvious objection to the chart above, and it has now been tested rather than left standing. Comparing applications that ended with the same total number of rejections, two, three, four, or five or more, the kept group was granted less often in every one of them, by 14.6 to 21.7 points. Reweighting the kept group to the dropped group’s mix of rounds leaves 19.9 points of the raw 21.2, so holding the number of rounds fixed accounts for about 1.3 of them and not for the gap. The objection is a real one to raise, and this pair of figures was reported after the results to show that: 52.8% of applications whose 101 was kept went on to three or more rejections, against 41.1% of those where it was dropped.
| Rejections in all | Kept | Applications | Dropped | Applications | Difference |
|---|---|---|---|---|---|
| 2 | 54.8% | 1,828 | 74.9% | 1,796 | -20.2 |
| 3 | 60.7% | 725 | 80.7% | 558 | -20.0 |
| 4 | 39.0% | 730 | 60.7% | 403 | -21.7 |
| 5 or more | 42.0% | 592 | 56.6% | 294 | -14.6 |
The pattern repeats outside this cohort. For 2021 filings the reweighted difference is 21.4 points against a raw 22.7, and for narrow software it is 21.2 in 2020 and 27.7 in 2021. For software only the two-rejection group has 200 or more applications in both arms, so there the verdict rests on that group with the three smaller ones pointing the same way.
Two things that does not settle. The number of rounds comes after the rejection was kept or dropped, so this holds fixed something a kept rejection may itself cause. And it does not touch what leads an examiner to keep one, which is the claims, the invention and the reply, and which may be what ends the application. Every application here received a second rejection, so they are already the harder cases. This still goes together rather than causes.
Art units that raise it more often grant somewhat less
Grouping 387 art units by how often their first actions raise eligibility, the pattern is not a steady slope. It is the top group.
| Technology centre | Art units | Correlation | 95% interval | Interval |
|---|---|---|---|---|
| 2100 Computer architecture | 13 | -0.775 | -0.929 to -0.39 | excludes zero |
| 3600 Transportation and e-commerce | 59 | -0.541 | -0.7 to -0.331 | excludes zero |
| 2400 Networks | 45 | -0.412 | -0.629 to -0.135 | excludes zero |
| 3700 Mechanical | 43 | -0.226 | -0.493 to 0.08 | includes zero |
| 2800 Semiconductors | 60 | -0.153 | -0.392 to 0.105 | includes zero |
| 2600 Communications | 59 | -0.118 | -0.363 to 0.142 | includes zero |
| 1600 Biotechnology and organic chemistry | 43 | 0.04 | -0.263 to 0.336 | includes zero |
| 1700 Chemical and materials | 64 | 0.122 | -0.128 to 0.357 | includes zero |
Two numbers in that caption are worth comparing. Pooled across every centre the association is -0.049, an interval straddling zero and effectively nothing. Ranked inside each centre it is -0.197. The difference is not a finding about patents, it is a property of the arithmetic: centres differ in both how often they raise eligibility and how often they grant, so pooling them lets the differences between centres stand in for a relationship within them. The within-centre figure is the one to read.
3 of the 8 intervals exclude zero. Testing 8 centres at a 95% interval would be expected to put about 0.4 of them there by chance alone, and 3 or more happens about 0.6% of the time, so this is not an artefact of having run 8 tests.
| Fifth | Share of first actions raising 101 | Art units | Applications | Mean granted of resolved | 95% interval |
|---|---|---|---|---|---|
| Lowest | 0.0% to 1.6% | 77 | 32,134 | 78.6% | 76.3 to 80.9 |
| Second | 1.6% to 3.6% | 77 | 38,463 | 80.3% | 78.4 to 82.2 |
| Middle | 3.6% to 7.1% | 77 | 30,647 | 79.0% | 76.0 to 82.1 |
| Fourth | 7.1% to 14.2% | 77 | 26,994 | 82.1% | 79.8 to 84.4 |
| Highest | 14.5% to 88.3% | 79 | 28,674 | 73.7% | 70.7 to 76.7 |
Art units are offices rather than people, and none is named here. An art unit's mix of applications is not random, so a unit that raises eligibility more often is also examining different subject matter.
Coverage: a first action counts here only when it could be matched to its own text, which was possible for 50.3% of 2020 applications (177,530 of 352,623), because the action text is incomplete for 2021 to 2024. The matched and unmatched halves were granted within about two points of each other, which limits the bias without excluding it. Eligibility rejections are found by matching the action text; checked against a reading of 300 actions, the pattern found 98.1% of their 101 rejections and raised no false ones.
What happens on appeal
An applicant who cannot move an examiner can appeal to the Patent Trial and Appeal Board. The record there is the one place in this data where an eligibility dispute is settled by someone other than the examiner who raised it, and the pattern has been steady for most of a decade.
| Decision year | Decisions | Affirmed outright | Affirmed in part | Reversed |
|---|---|---|---|---|
| 2016 | 804 | 56.3% | 17.2% | 23.9% |
| 2017 | 1,703 | 75.7% | 7.0% | 13.8% |
| 2018 | 2,554 | 82.5% | 3.4% | 13.0% |
| 2019 | 2,507 | 73.8% | 4.3% | 19.4% |
| 2020 | 2,071 | 77.0% | 4.6% | 15.5% |
| 2021 | 1,043 | 74.1% | 5.3% | 17.2% |
| 2022 | 666 | 78.7% | 4.7% | 12.8% |
| 2023 | 682 | 78.4% | 3.8% | 14.4% |
| 2024 | 805 | 82.6% | 2.4% | 12.8% |
| 2025 | 1,095 | 80.0% | 4.1% | 12.6% |
| 2026 | 758 | 78.2% | 3.6% | 14.2% |
These are appeals, which are a selected group: an applicant appeals when the disagreement is worth the cost and the delay, so this is not the outcome of eligibility disputes generally. 2026 is a partial year.
The one causal claim, and it is not ours
Nothing above establishes cause. A filing-year series cannot separate Alice from everything else that changed. The USPTO's own economists ran the comparison that can, measuring Alice-affected technologies against others around each event, and their findings are quoted here rather than recomputed.
Read together with the rest of this page, that is the shape of the answer. Alice changed the road measurably. In the cohorts measured here, it did not show up as a lower grant share for software as a whole. Business methods are where it shows in both.
What this record cannot say
- Enforcement is outside the data. Everything here comes from the patent office: applications, Office actions and appeal decisions. None of it covers litigation, licensing or what a granted patent is worth, so the claim that software patents are harder to enforce is neither supported nor contradicted.
- Only published or granted applications appear. An application abandoned before publication, usually one filed with a request not to publish, never enters the record, so every grant share here may run slightly high. About 3.6% of 2020s serial numbers are missing, which could lower the 2021 share from 76.3% to about 73.9% at most.
- Provisional applications are not counted at all, since they are never examined or published.
- Filing year is not examination year, and nothing here shows that Alice caused any change. The timing is consistent with it, and with the 2019 guidance that followed.
All public USPTO data, read in the ObviouslyNot continuation study (studies S4b, S5, S14, S19b, S21, S22, S24, S27 and S28). Outcome by class is the Open Data Portal record of 26 October 2026; the filing-year series is the PEDS final snapshot to 7 March 2025; rejection grounds are every non-final Office action mailed 2024 to 2025, matched on the action text.
Describes the record; it is not legal advice and says nothing about any one application.
The figures on this page are available as a CSV: /data/are-software-patents-harder-since-alice.csv. It is generated from the tables and charts above, so it carries the same numbers and cannot fall out of step with them.