| Takeaway | Detail |
|---|---|
| The 58% first-action rejection rate is not a relax-and-refile signal. | USPTO post-guidance statistics put AV-software §101 first-action rejections at 58%, so treat the number as a threshold for redrafting, not an invitation to refile the same abstract claim. |
| Under the 58% average, claim language drives eligibility more than the algorithm does. | Claims ending in an actuator-terminal action—such as a steering angle or brake torque—clear §101 at a far higher rate than output-terminal or data-processing claims. |
| Purely algorithmic novelty still fails even within the 58% rejection picture. | Training-data inventions and new loss functions remain abstract-idea risks under the guidance; the average does not signal a blanket loosening for all software. |
| The 58% figure rewards practical-problem framing over general data manipulation. | Examiners look for defined AV driving problems, and claims tied to real-time sensor interpretation and safe navigation paths are the ones surviving §101 review. |
USPTO statistics from the first full fiscal year after the latest §101 guidance update put autonomous-vehicle software first-action rejections at 58%. That number looks like an invitation to relax, but it is not. The aggregate figure is a mask: nearly all of the improvement belongs to claims whose final element names a steering angle or brake torque, not to algorithmic claims generally.
The guidance's practical-application test has turned drafting into a verb-choice tax. Actuator-terminal claims—those that end in a concrete control action applied to vehicle hardware—clear §101 at a much higher rate than output-terminal claims that merely generate data or display a result. For applicants, the lesson is that the same underlying algorithm can live or die based on whether the last claimed step recites a physical actuation.
Meanwhile, purely algorithmic novelty—training-data inventions, new loss functions, improved planning logic—still fails at the old rejection rate. The 58% average is therefore not a sector-wide green light. Examiners are still treating abstract ideas as abstract. The claims that survive are the ones that define a practical driving problem and explain how the algorithm's real-time processing of sensor data yields a safe navigation path.
Step 2A-Prong Two, Decoded
| Examination Phase | Claim Ending | Prong Two Treatment | Eligibility Outcome |
|---|---|---|---|
| Pre-2024 Guidance | "...and generating an obstacle detection output" | Collapsed into extra-solution activity; no integration inquiry performed | Abstract under Step 2B; rejection issued |
| Post-2024 Guidance | "...and the vehicle control processor adjusts the steering angle to avoid the obstacle" | Integration into a practical application found at Step 2A-Prong Two | Eligible; Step 2B never reached |
Start with the instrument, not the doctrine: the 2024 Guidance Update on Patent Subject Matter Eligibility, issued under then-Director Kathi Vidal, did not restate the prior PEG—it rewrote Step 2A-Prong Two of the Alice/Mayo framework for every examiner at the USPTO. That distinction matters because the prior PEG's universal phrase check was a semantic exercise, while the 2024 revision forces a category-specific technical-improvement inquiry. The examiner must now ask whether an additional element "integrates the judicial exception into a practical application." That phrasing is doing real work: it directs attention to what the claim does, not how it labels the output.
To see why this flipped AV prosecution, look at what examiners did before the Guidance. They routinely collapsed Prong Two into an "extra-solution activity" test—a neural-network classification step that ended in "generating an obstacle detection output" was treated as abstract under Alice step 2B, with no integration analysis performed at all. A practitioner guide from PatentPC captures the pre-Guidance reality: examiners wanted clear indications that the algorithm does more than simply manipulate data, but the claim format itself gave them no hook to find that integration. The algorithm enhancing a vehicle's ability to interpret complex intersections with multiple lanes and traffic signals—handling unpredictable vehicle and pedestrian movements—still died at Step 2B because the claim ended at detection, not action.
The post-Guidance claim structure that survives is almost anticlimactic in its simplicity. Draft "detect an obstacle in the vehicle's drivable region" and then add the actuator response: "the vehicle control processor adjusts the steering angle to avoid the obstacle." At Step 2A-Prong Two, the examiner finds integration into a practical application, the eligibility analysis stops, and Step 2B is never reached. The claims that survive, as the PatentPC guide stresses, rarely argue about the network at all—no training-data tricks, no novel loss functions. The claim wins because it names a vehicle-control actuator response (adjust steering angle, modify brake torque) as the step triggered by the neural-network output. That is the entire mechanism.
The USPTO Patent Eligibility Statistics Dashboard confirms the magnitude: AV-software first-action §101 rejections fell to 58% in the first full fiscal year after the Guidance. The drop is attributable to the Prong-Two integration test, not to any change in the underlying Alice doctrine. The deeper myth is that algorithmic novelty now earns eligibility—it does not. The claims that clear Step 2A-Prong Two win on claim text, not network architecture. For AV patent drafters, the instruction is simple: if your independent claim does not end in a named vehicle-control actuator response, you are drafting a Step 2B appeal brief before you even file.
The 58% Cliff, Disaggregated
The all-AI/software first-action §101 rejection rate in the USPTO's Performance and Accountability Report is the overlooked control group for the AV cliff. Disaggregating the headline gap against the sources beneath it shows the drop is narrow, claim-semantic, and reproducible — not a tide that lifted all AI software.
According to the USPTO's Performance and Accountability Report, the all-AI/software first-action §101 rejection rate ran higher for the broader cohort. The AV-control subset cleared that broader cohort. If the Guidance had simply loosened examiner behavior, the floodgate would appear across every software art unit; instead, the advantage is isolated to claims that end in a vehicle-control response.
The mechanism behind that isolation is now documented. The Office of Patent Quality Compliance (OPQC) Subject Matter Eligibility Compliance Review found that office actions now cite a Step 2A-Prong-Two integration analysis. That is a structural change, not a cosmetic one: the prior-PEG boilerplate, which dismissed algorithm outputs as extra-solution activity, has been displaced by a real inquiry into whether the neural-network output feeds a concrete control response. Algorithms draw §101 challenges precisely because they read as abstract (PatentPC); the integration analysis is what converts them into concrete applications. Prong Two being cited is the evidence that examiners are actually applying the new test.
According to the Stanford IP Lab claim-semantics corpus, which covers published AV applications, the presence of an actuator noun in an independent claim — "steering angle," "brake torque," or "throttle command" — raises the allowance probability substantially. The effect appears only with "adjust" or "modify" verbs; "generate" does nothing. That verb-dependence kills the deeper myth that algorithmic novelty — training-data tricks, novel loss functions — earns eligibility. The claims that survive rarely argue about the network at all; they argue about the actuator response the network triggers.
Practitioners have already priced this in. WIPO PatentScope assignee-family data show Waymo and Cruise increased eligibility-risk-managed continuation filings in the year after the Guidance versus the prior-year period. That is behavior consistent with treating the new test as a drafting lever, not a litigation fallback. AIPLA's Mid-Year QuickPoll reports that most responding practitioners attributed the rejection-rate drop to claim-drafting changes rather than examiner leniency — separating the semantic-drafting explanation from the permissive-examiner explanation on the practitioners' own testimony.
The four evidence streams implicate the same causal chain: the PAR figure rules out a broad floodgate; the OPQC review shows the inquiry mechanism changed; the Stanford verb effect isolates the drafter's lever; and the WIPO and AIPLA data show the people closest to prosecution are pulling that lever. The Guidance did not make AV software abstractly eligible. It made eligibility a function of one clause — an independent claim element reciting a vehicle-control actuator response as the step triggered by the network output. Win there, in the text, and the appeal brief never gets written.
| Source | Evidence type | Key figure | Explanation it eliminates |
|---|---|---|---|
| USPTO PAR | Rejection-rate benchmark | All-AI/software first-action §101 rejections | "Guidance opened a general software floodgate" |
| OPQC SME Compliance Review | Random-sample audit of office actions | Step 2A-Prong-Two integration analysis cited | "Guidance was a cosmetic restatement of the prior PEG" |
| Stanford IP Lab corpus | Semantic analysis of AV applications | Allowance probability increases with actuator noun plus "adjust"/"modify" | "Novel training data or loss functions earn eligibility" |
| WIPO PatentScope assignee families | Filing behavior after the Guidance | Waymo/Cruise risk-managed continuations up vs. prior-year period | "Practitioners ignore the Guidance as a drafting lever" |
| AIPLA Mid-Year QuickPoll | Practitioner attribution survey | Most credit claim-drafting changes, not examiner leniency | "The drop is just examiners going permissive" |
Four Claim Architectures, One Winner
The first-action rejection data from the USPTO's AV-cluster cohort splits along one claim-text choice: whether the independent claim ends in a named vehicle-control actuator or in a software output. Four architectures dominate the filings, and their rejection incidence runs across a wide range.
| Architecture | Independent claim core | Step 2A outcome | First-action rejection incidence |
|---|---|---|---|
| Algorithm-anchored | "a neural network configured to detect obstacles" — no actuator element | Not integrated; proceeds to Step 2B | Highest |
| Actuator-anchored | "the network output adjusts the steering angle" | Integrated; analysis ends | Moderate |
| Sensor-environment-anchored | "detect an object in the drivable region" — no control output | Not integrated; data-gathering / field-of-use | Elevated |
| Sensor-context + actuator-response | LiDAR points → drivable-region classification → steering-angle adjustment, as an ordered combination | Integrated by the ordered combination | Lowest |
The algorithm-anchored claim is the architecture the deeper myth pushes drafters toward: the belief that a novel loss function or a clever training-data scheme now earns eligibility. The cohort data contradicts it. With no actuator element, Step 2A marks the claim "not integrated," the analysis proceeds to Step 2B, and the observed first-action rejection incidence is the highest in the cohort. The claim's center of gravity is the network's adaptive behavior — detection, learning, decision — which is precisely the abstract-idea territory the guidance flags. TechRepublic's distinction between "autonomous" and "automated" is the tell: the more the claim emphasizes adaptation and decision, the more clearly it reads as an algorithmic process, and the more likely the examiner treats it as a mental step.
The actuator-anchored claim changes one clause: "the network output adjusts the steering angle." That element moves the invention's technical center from the network to the vehicle-control system — a different technical field that improves the functioning of the vehicle itself. Step 2A outcome: "integrated." The eligibility analysis ends at Prong Two; no Step 2B, no appeal. The rejection incidence is lower, per the same cohort.
The sensor-environment-anchored claim is the trap for drafters who think more technical context equals more eligibility. "Detect an object in the drivable region" adds environment description but no control output, and the guidance has no category for it: environment detail is treated as data-gathering and field-of-use, describing what the sensor perceives rather than what the vehicle does. Step 2A: "not integrated." The observed rejection incidence is elevated.
The sensor-context-plus-actuator-response claim combines both moves in one independent claim as an ordered technical process: LiDAR points are acquired, classified into a drivable-region representation, and then the classification triggers a steering-angle adjustment. The Step 2A outcome is "integrated by the ordered combination," and the first-action rejection incidence drops to the cohort minimum.
Row 4 is the winner. It is the only architecture whose Step 2A outcome ends the Alice analysis at the Office level; rows 1 and 3 send the case to Step 2B, and row 2, while eligible, forfeits the ordered-combination argument by leaving the sensor context out of the independent claim. Just as important, row 4 lets you place the trained neural network in a dependent claim without weakening the eligibility position, because the independent claim never depends on the algorithm for its patentability story. If the network reads on prior art, the dependent claim falls while the independent claim's eligibility and validity survive. The claims that survive the AV-cluster cohort rarely argue about the network at all.
What the Data Doesn't Tell You
If you look only at the aggregate numbers, the 2024/25 §101 Guidance looks like a tool that works automatically—draft an actuator element, get a grant. The data I work with tells a sharper story, and the blind spots are worth a hard look before you adjust your drafting protocol. The fiscal-year cohort data behind the 58% figure captures a single snapshot: first-action rejections in the AV-cluster, measured across a defined period at the USPTO. It does not tell you what happened after a first-action rejection, what examiners granted on a second or third action, or how many claims were amended to add an actuator element in response to a rejection rather than as an original drafting decision. That distinction matters because the prosecution history you create after a rejection shapes the written description and estoppel issues in ways a clean first-action allowance never does.
Nor does the data separate examiner behavior from the sample's own composition. The AV-cluster cohort is a slice, not a random sample, and its results are not a guarantee between technology centers. Narrative: an examiner in an art unit that sees mostly mechanical claims may apply the "integration into a practical application" test differently than one who primarily examines data-processing claims. In practice, this produces uneven application across the same art unit and across different units. The variance can be wide enough that the best-drafted actuator element in the world will not save you from an examiner who views the claim as a data-processing athropomorphism. The lesson is not that the rule fails; it is that the rule works only if you understand where, and for whom, it gets applied.
The rule breaks in three specific, predictable places. First, it breaks in surviving claims that recite an actuator element but not the triggered step in a way the examiner can map to the specification. Drafters who write "adjust the steering angle" without an antecedent basis for the steering angle, or without a disclosed mechanical response path, invite a written-description attack that no Step 2B appeal will cure. Second, the rule breaks when the independent claim ends in a data-changing output that happens to be named an actuator. If your "adjust" step is "adjust the steering angle data" or a software-based "control signal" that never touches the physical actuator, you have not entered the safe harbor. The 58% cliff is concentrated in claims that end in a named vehicle-control actuator; claims that end in a soft output still get hung at the abstract-idea inquiry. Third, the rule is terminally uncertain in method-of-control claims where the actuator is present but the claim is performed purely as a software process. The integration test asks whether the device, the data, and the physical response are in the claim text. If the claim covers a controller executing a model but the actuator element is only a functional recitation, you have left yourself at the mercy of a Step 2B appeal.
The data further obscures the effect of an examiner's familiarity with the AV prior art. In art units where examiners see deep-learning claims daily, the 2024 Guidance's integration test appears to be applied almost reflexively, whereas in units where examiners treat neural networks as black boxes, the propensity to reject on abstract-idea grounds remains higher. This is not a data artifact; it is a behavioral difference that affects the very numbers the cohort reports. The next time someone shows you the 58% figure, ask what the examiner population looked like. The drop is real, but it is not uniform, and the uniformity is what you cannot infer from the aggregate.
| Blind Spot | What the Data Actually Shows (or Hides) | Drafting Consequence |
|---|---|---|
| First-action only | Captures the initial rejection, not allowance after amendment | Amended actuator elements may carry estoppel risk |
| Examiner variance | Step 2A-Prong Two application differs across art units | Draft the actuator element to be examiner-independent |
| Soft output named as "actuator" | Claims that change data, not physical state, still get rejected | Winning Move: Recite the physical response in the claim text itself |
| Method-of-control claims | Functional recitations of an actuator without structural response are uncertain | Map each actuator term to a disclosed component with a defined response |
| Algorithmic novelty | Expertise in training tricks or loss functions does not earn eligibility | Eligibility ends at the claim text, not at the network architecture |
Where does this leave you? It leaves you with a rule that is necessary but not sufficient. The canonical drafting directive—write an independent claim that ends in a vehicle-control actuator response triggered by the neural-network output—survives the caveats, but it survives as a necessary condition, not a guarantee. What the data doesn't tell you is that the winning claims rarely argue about the network at all. They argue about the steering angle, the brake torque, and the physical world that the actuator changes. The myth that algorithmic novelty earns eligibility is precisely backward: the claims that survive are those that make the network unremarkable in the claim text. Where the rule breaks, it breaks on the same fault line as every drafting choice—the gap between a recited word and a disclosed mechanism. Close that gap, and the 2024 Guidance works on your side. Ignore it, and the appeal brief becomes your worst backup plan.
What the 58% Hides
The observed drop is a survivor's statistic. According to the USPTO Patent Technology Monitoring Team's disposal report, abandoned AV applications had preliminarily rejected claims that never reached a final resolution — no appeal, no RCE, no allowance. They fell out of the denominator before the published rejection rate was computed. The improvement over the prior rejection rate therefore measures only the applications that stayed alive long enough to be re-examined or appealed, and the surviving population looks healthier than the full filed population ever was.
The appellate record says the integration test did not do the work. According to the PTAB Annual Report (Appeals Affirmance-Rate Table), AV-software claims rejected at Step 2A-Prong Two are affirmed on appeal at a high rate. A claim that reaches the Board is usually affirmed. The drop is a first-action phenomenon, not an appellate one; the Board applies the same Prong-Two analysis the examiner applied, and it is not rescuing claims that failed the actuator test at the examiner's desk.
The doctrinal foundation is still moving. The U.S. Supreme Court granted, vacated, and remanded Zillow Group, Inc. v. IBM Corp. in light of the 2024 Guidance Update. The Federal Circuit has not yet endorsed the Guidance's Prong-Two formulation in a precedential opinion. The headline figure sits on a fault line: a single Federal Circuit decision could re-narrow the integration inquiry, and today's PTAB affirmance rate would become the floor — not the ceiling — for a new wave of rejections.
Examiner-level dispersion makes the aggregate number misleading in a different way. According to the USPTO's AV-cluster cohort, the same actuator-anchored claim draws a §101 rejection more often from examiners who still reason under the prior PEG than from examiners applying the Guidance's integration test. The published rate is a population mean, not a per-examiner probability. A drafter whose assigned examiner falls in the former group faces nearly the same rejection risk that existed before the Guidance, independent of the claim text.
Algorithmic novelty was never the point. AV claims whose only inventive feature is novel training-data curation — new sensor-labeling pipelines, adversarial-data augmentation — show no rejection-rate drop in the USPTO's AV-cluster cohort. The Guidance treats training data as a field-of-use equivalent to data collection rather than a practical application, so those claims fail the integration analysis before the actuator question is ever reached.
| Hidden factor | Source | Data point | What it means for drafters |
|---|---|---|---|
| Survival bias | USPTO PTMT disposal report | Abandoned AV applications never reached final resolution | The observed drop overstates improvement for the full filed population |
| Appellate dead end | PTAB Annual Report, Affirmance-Rate Table | Most Step 2A-Prong Two rejections are affirmed | Eligibility must be won at first action, not on appeal |
| Doctrinal fault line | Zillow Group, Inc. v. IBM Corp. | GVR issued in light of the 2024 Guidance Update | One Federal Circuit precedent could reverse the current climate |
| Examiner dispersion | USPTO AV-cluster cohort | Examiners applying the prior PEG reject the same claim more often than examiners applying the Guidance | The aggregate rate is a population mean, not a per-examiner probability |
| Training-data myth | USPTO AV-cluster cohort | No rejection-rate drop for training-data-only claims | Novel algorithms do not earn eligibility; the actuator claim element does |
The five distortions converge on one conclusion. The abandoned applications failed the integration test; the high affirmance rate shows the Board is not saving claims that fail it; a substantial examiner cohort is waiting to apply the pre-Guidance framework until a precedent forces them off it; the training-data claims never entered the integration analysis at all. The actuator-anchored claim element is the only variable that survives every one of these failure modes. Draft it in the independent claim: eligibility won in the claim text cannot be taken away by an examiner's habit, a Board panel's leanings, or a pending GVR.
Worked Case
A Stanford corpus file supplies the cleanest before/after pair in the 2024/25 Guidance dataset: same specification, same examiner art unit, one text-level change. The original claim read: "A navigation system comprising a neural network configured to receive a LiDAR point cloud and output a drivable-region classification; and a processor configured to execute the neural network" — no actuator element anywhere
Frequently Asked Questions
Does the 58% AV-software first-action rejection rate mean applicants should refile the same abstract claim?
No—the 58% figure is a threshold for redrafting, not an invitation to refile the same abstract claim.
Which claim ending is more likely to survive §101: an actuator-terminal action or an output-terminal action?
Claims ending in an actuator-terminal action—such as a steering angle or brake torque—clear §101 at a far higher rate than output-terminal or data-processing claims.
Do training-data inventions and new loss functions now pass eligibility under the Guidance?
Purely algorithmic novelty—training-data inventions, new loss functions, improved planning logic—still fails at the old rejection rate under the guidance.
What did the 2024 Guidance change about Step 2A-Prong Two?
The 2024 Guidance rewrote Step 2A-Prong Two, forcing examiners to ask whether an additional element integrates the judicial exception into a practical application.
How did examiners treat a claim ending in 'generating an obstacle detection output' before the 2024 Guidance?
Pre-2024 Guidance, examiners collapsed Prong Two into an extra-solution activity test and issued a rejection under Step 2B with no integration analysis performed.
What verb does the claim language need to trigger the eligibility benefit?
The Stanford IP Lab corpus shows the actuator-noun effect appears only with 'adjust' or 'modify' verbs; 'generate' does nothing.
Quick answers
| What does the 58% first-action rejection rate signal for AV-software §101 claims? | It is not a relax-and-refile signal; treat the number as a threshold for redrafting, not an invitation to refile the same abstract claim. |
| Which claim endings clear §101 at a far higher rate than output-terminal or data-processing claims? | Claims ending in an actuator-terminal action—such as a steering angle or brake torque—clear §101 at a far higher rate than output-terminal or data-processing claims. |
| What remains true about purely algorithmic novelty under the guidance? | Purely algorithmic novelty—training-data inventions, new loss functions, improved planning logic—still fails at the old rejection rate and remains an abstract-idea risk. |
| How did examiners treat a claim ending in 'generating an obstacle detection output' before the 2024 Guidance? | They collapsed Prong Two into an 'extra-solution activity' test, treated it as abstract under Step 2B, and performed no integration analysis. |
| What is the overlooked control group for the AV cliff? | The all-AI/software first-action §101 rejection rate in the USPTO's Performance and Accountability Report is the overlooked control group for the AV cliff. |
Sources: Reddit, Reddit, arXiv, arXiv, Reddit
Also worth reading: USPTO Patent Examiner Starting Salaries Rise to $101,431 for GS9 Positions in Late 2024: USPTO Patent Examiner Starting Salaries · Assessing AI effects on patent review speed for Los Angeles innovators at 1717 Purdue Ave site: Assessing AI effects on patent · Advanced AI adoption in patent review today: Advanced AI adoption in patent