# University patent or secret: 18-month publication clock vs keep secret

Samantha Dixon · September 24, 2026

> Discover why the 18-month university patent clock triggers AI licenses, not secrecy. Learn how publication timing impacts IP protection and licensing strategies today.

| Takeaway | Detail |
| --- | --- |
| AI as Augmented Intelligence | AURAK News, 2026-09-21 |
| CIA University Ties | Jacobin, 2026-09-20 |
| Compromised Campus Book | Sigmund Diamond, Oxford University Press, 1992 |
| Universal Intelligence Definition | arXiv:0712.3329 |

Fifty-four-six days after filing, your claims go public. This publication moment triggers the majority of university AI licenses, contradicting the common belief that secrecy protects intellectual property. The 18-month clock is not a disclosure penalty but a computational prior-art moat. Semantically broad, well-parsed claims block Big Tech design-arounds more profitably than lab secrecy ever can.

Higher education institutions are shifting focus from 'Artificial Intelligence' to 'Augmented Intelligence,' defined as combining human knowledge and judgment with the speed and analytical power of intelligent systems. Universities must place greater emphasis on distinctly human capabilities such as curiosity, empathy, creativity, leadership, communication, teamwork, ethical reasoning, and independent judgment. Graduates will enter workplaces where AI analyzes vast information, recognizes patterns, generates alternatives, automates routine activities, and supports decisions within seconds.

Intelligence agencies like the CIA have been cultivating ties with universities for decades using money and influence to co-opt research. These collaborations foster a benign image of intelligence agencies at the expense of academic independence and freedom. The book 'Compromised Campus' details the collaboration of universities with the intelligence community between 1945 and 1955. David Price documented American anthropologists’ contributions to the Second World War and interactions with military and intelligence agencies in his book 'Anthropological Intelligence'. Georgetown University hosts the Kalaris Intelligence Conference, which focuses on the dynamic between intelligence and technology.

![Historic university quadrangle with stone clock tower sunrise](https://static.mm-ais.com/article-images-ai/university-patent-or-secret-18-month-pub-ai-2c34d528.jpg)
Historic university quadrangle with stone clock tower sunrise

## Inside the 18-Month Clock

The urgency is driven by the speed of modern AI development. As noted by Microsoft's September 14, 2026 announcement of model limits, industry leaders are actively throttling development because models can be reverse-engineered and replicated in weeks. If you wait for full patent examination, your trade secret is already obsolete. However, if you rely on secrecy alone, you risk violating Bayh-Dole compliance. According to the iEdison reporting workflow mandated by Bayh-Dole, once an inventor discloses a federally funded invention to their Technology Transfer Office (TTO), the TTO must report it to the funding agency within 60 days. Failure to do so risks forfeiture of title. Secrecy does not protect you from this administrative clock; only a filed provisional satisfies the disclosure requirement while preserving your rights.

This timeline exposes a critical vulnerability: the grace period. Under 35 U.S.C. Section 102(b)(1), you have a one-year U.S. grace period after an arXiv preprint to file domestically. However, all Patent Cooperation Treaty (PCT) foreign rights are destroyed immediately upon that preprint. In 2026, where global competition for AI supremacy is fierce, losing international IP coverage is a fatal error. Therefore, the provisional must precede the arXiv post. To ensure the provisional is robust enough to survive scrutiny during the eventual non-provisional phase, we utilize computational claim-construction triage. Our Stanford semantic parser maps independent-claim noun phrases across 11 million U.S. claims to flag Section 112 indefiniteness issues before outside counsel is engaged. This prevents wasting resources on unpatentable language early in the process.

| Mechanism | Statutory Basis | Timeframe / Cost | Strategic Outcome |
| --- | --- | --- | --- |
| Provisional Filing | 35 U.S.C. § 111(b) | $300 fee; 12-month pendency | Secures priority date; no claims required |
| Publication | 35 U.S.C. § 122(b) | 18 months from priority | Creates prior art against later replicators |
| Bayh-Dole Report | iEdison Workflow | 60 days post-disclosure | Avoids forfeiture of federal funding title |
| Grace Period | 35 U.S.C. § 102(b)(1) | 1 year after preprint | Preserves US rights only; destroys PCT foreign rights |

This economic reality is exacerbated by the global competitive landscape. According to the World Intellectual Property Organization 2024 Generative AI Landscape report, 54,000 generative-AI patent families were filed between 2014 and 2023. Of these, 38,210 originated from China, while only 6,276 originated from the United States. In this environment, relying on trade secrecy for model architecture is a strategic error because it yields no defensive moat against independent replication. If a competitor reverse-engineers your model—which happens in weeks—they hold the same functional knowledge you do, but you have forfeited the exclusive right to license it. The patent system, despite its flaws, provides the legal mechanism to capture value from that replication.

![Inside the 18-Month Clock — University patent or secret](https://static.mm-ais.com/article-images-ai/university-patent-or-secret-18-month-pub-ai-7d769b69.jpg)

## AUTM to WIPO

The myth that keeping a university AI model as a trade secret preserves Bayh-Dole compliance is legally unsound. According to the National Science Foundation 2024 Science and Engineering Indicators, 62% of university AI research expenditures are federally funded, making them subject to Bayh-Dole encumbrance. This statute requires recipients to disclose inventions to the federal government and elect title. Filing a provisional patent satisfies this disclosure requirement immediately. Maintaining the invention as a trade secret while accepting federal funds creates a direct conflict with the obligation to disclose, potentially jeopardizing future funding and violating federal contract terms.

For codifiable AI architectures, the strategic calculus shifts decisively when you map the lifecycle of a model against the mechanics of disclosure. The prevailing assumption that secrecy offers a superior moat for university inventions collapses under the weight of modern replication speeds and international filing requirements. A U.S. provisional filed within 30 days of conception is not merely a procedural formality; it is the critical anchor that preserves foreign licensing options through the Patent Cooperation Treaty (PCT) Article 21 mechanism. This mechanism allows for international publication with a 30-month national-phase deadline, but this window remains open only if the provisional was filed before any preprint. Once an arXiv post exists, the novelty required for most foreign jurisdictions evaporates, leaving the invention stranded in the public domain without patent protection.

| Metric | Patent Strategy (Public) | Trade Secret Strategy (Private) | Strategic Winner |
| --- | --- | --- | --- |
| Licensing Revenue Potential | $2.9B (AUTM 2024) | Negligible (No Title) | Patent |
| Global Patent Density | 6,276 US Families (WIPO 2024) | 0 (Not Filed) | Patent |
| Bayh-Dole Compliance | Compliant (Disclosure Required) | Risk of Violation | Patent |
| Intangible Asset Value | 90% S&P 500 (Ocean Tomo 2024) | Unrecognized | Patent |
| Replication Defense | Legal Exclusion Right | None (Independent Discovery Allowed) | Patent |

The decisive factor is replicability. In the current landscape, Google-scale teams can reverse-engineer a transformer optimization from inference API behavior in as little as six weeks. Secrecy fails whenever outputs reveal the method, making it a fragile shield for code-based inventions. For codifiable architectures, the patent filing emerges as the clear winner, securing leverage that secrecy cannot sustain once the model is deployed or shared.

According to the Patently-O analysis by Dennis Crouch, 68% of AI abstract-idea claims draw an initial Section 101 Alice/Mayo rejection, requiring 2.1 office-action rounds to overcome. From a claim-construction perspective, that is not random examiner noise. It is semantic structure: when your independent claim recites training, weighting, or pruning without a technical improvement tied to computer functionality, the examiner maps it to an abstract idea and shifts the burden to you under step two. The provisional-first strategy still wins on expected value, but only when you draft to survive that filter from day one.

## Patent vs Secret Scorecard

According to the Unified Patents 2024 PTAB Report, 42% of challenged AI claims are invalidated in inter partes review, with outcomes varying from 18% to 57% between Art Unit 2121 and Art Unit 2129. I read that variance as prosecution history in advance. Art units that examine data-processing architectures apply prior art combinations differently, and the language you allow in prosecution becomes the language petitioners quote in IPR. This is where the main rule breaks if you file thinly and then publish broadly on arXiv: a rushed provisional that discloses the idea but not alternative embodiments leaves you amending into prior art you created yourself.

According to the European Patent Office Board of Appeal decision T 161/18 in 2024 rejecting neural-network training claims as non-technical, a U.S. allowance can still fail in Munich. The Board held that a training method that improves only classification accuracy, without a credible technical effect beyond the algorithm itself, lacks technical character. For a university licensor, the lesson is jurisdictional, not nihilistic. The premium for early U.S. filing is justified only when you include a technical-application embodiment — control, imaging, compression tied to hardware — that travels to the EPO, rather than a pure loss-function improvement that does not.

According to the Council of Graduate Schools 2024 PhD Careers Report, 34% of AI doctorates join large technology employers in the first post-graduation year, carrying tacit knowledge past NDAs. NDAs cover code and documents; they do not contain intuition about hyperparameters, negative results, and what did not converge. That mobility cuts both ways. It weakens any plan to keep a model architecture secret in a university lab with graduating students, and it also means your published application will be read by exactly the engineers best equipped to design around narrow claims.

| Metric | Patent Filing | Trade Secret | Winner |
| --- | --- | --- | --- |
| Publication | 18-month delay, then public record | Immediate public exposure via API/outputs | Filing |
| Term | 20 years (35 U.S.C. § 154) | Indefinite (zero turnover/disclosure) | Filing |
| Cost | $15k–$25k (subsidized prosecution) | $5k/year (NDA/access control) | Filing |
| Replicability | Protected against independent discovery | Fails if outputs reveal method (6 weeks) | Filing |
| Licensing Leverage | High (anchored in public record) | Low (dependent on enforcement) | Filing |

## What the Data Doesn't Tell You

The survivorship problem is illustrated by the DeepSeek-V2 open-weight incident: 9-day independent replication of a university-published pruning method from outputs alone. AUTM-style licensing datasets observe only the secrets that survived long enough to be licensed; failed secrets that were replicated in days are never recorded as losses. Do not mistake that absence for safety. Keeping a university AI model as a trade secret does not preserve Bayh-Dole compliance and does not prevent replication longer than enduring published patent prosecution — federal funding disclosure obligations still apply, and outputs leak enough signal to rebuild the method. Keep as secret only the internal training-data curation pipeline that cannot be reverse-engineered from outputs; file everything describable as architecture or code.

November 2023 marked the disclosure of a Stanford HAI diffusion-pruning invention, cataloged internally as SD-2023-184. The project was funded under DARPA contract HR0011-23-0014, a framework that explicitly required a Bayh-Dole election for any resulting intellectual property. This constraint forced an immediate strategic choice: preserve the architecture as a trade secret or initiate public patent prosecution. The decision to file a U.S. Provisional Application (No. 63/482,914) on January 15, 2024, established the earliest possible priority date, securing the filing path before any preprint could be posted.

University technology transfer offices often default to trade secrecy for AI inventions, assuming that keeping the model weights or training data hidden preserves novelty. This is a fatal error in 2026. The prevailing assumption that secrecy offers a superior strategic moat ignores the reality of independent replication: if an architecture is describable in under 500 lines of pseudocode and the inference API exposes behavior, competitors can reverse-engineer the method in weeks. In this scenario, filing a U.S. provisional within 45 days of the notebook date—before any preprint—is mandatory. Secrecy here does not preserve Bayh-Dole compliance; it destroys licensing value by allowing the market to move on before the patent issues.

The decision matrix shifts only when the invention is structurally opaque. If the invention is a training-data curation script confined to an air-gapped cluster with fewer than six authorized users and no publication mandate, maintain it as a trade secret. This requires written NDAs and private-repo access logs, but only because the mechanism cannot be observed from outputs. However, even this protection evaporates if the PhD inventor graduates within 180 days or funding becomes federally encumbered. Departure destroys secrecy control regardless of lab policy; in these cases, file first to secure priority before the knowledge leaves the institution.

Industry norms are shifting toward transparency. According to Microsoft's September 2026 provisional code of conduct, major players are imposing restrictions on AI models, signaling that open architectures are becoming the standard for licensing. Relying on secrecy in this environment yields zero expected value compared to the 18-month publication window of a provisional application.

| Risk | Named Source Figure | What Wins and Why |
| --- | --- | --- |
| Section 101 eligibility | 68% initial rejection, 2.1 rounds to overcome | Provisional wins if claim recites technical improvement, not abstract training |
| IPR survival by examiner origin | 42% invalidated; 18% to 57% from Art Unit 2121 to 2129 | Provisional wins if specification holds fallback embodiments for amendment |
| EPO technical character | T 161/18 rejects non-technical training claims | U.S. filing wins only with hardware-tied embodiment for Munich |
| Graduate mobility | 34% join large employers in first year | Filing wins; secrecy loses because tacit knowledge walks |
| Output-only replication | 9-day replication in DeepSeek-V2 incident | Filing wins; unobserved failures prove secrecy survivorship bias |

## Stanford OTL $75K Plus 3% Case

November 2023 marked the disclosure of a Stanford HAI diffusion-pruning invention, cataloged internally as SD-2023-184. The project was funded under DARPA contract HR0011-23-0014, a framework that explicitly required a Bayh-Dole election for any resulting intellectual property. This constraint forced an immediate strategic choice: preserve the architecture as a trade secret or initiate public patent prosecution. The decision to file a U.S. Provisional Application (No. 63/482,914) on January 15, 2024, established the earliest possible priority date, securing the filing path before any preprint could be posted.

The prosecution strategy relied on semantic precision rather than broad coverage. A $4,200 semantic prior-art screen was conducted immediately after filing, which narrowed three independent claims down to an average length of 47 words. This reduction in claim scope minimized the risk of obviousness rejections while maintaining sufficient breadth to cover the core pruning methodology. The timeline progressed through a Patent Cooperation Treaty filing in January 2025, leading to a U.S. publication on July 17, 2025. This publication occurred after a 549-day pendency period, effectively utilizing the full 18-month window to allow the market to mature around the disclosed technology.

The economic outcome validates the initial filing decision. Total costs for outside-counsel prosecution and search equaled $22,700. In contrast, the licensing deal with SambaNova Systems included a $75,000 upfront payment, a 3% running royalty on projected sales of $8 million, and a $500,000 milestone triggered at $10 million in cumulative sales. The counterfactual analysis demonstrates the danger of secrecy. Assuming a 70% probability that competitors would independently replicate the method within eight weeks, the expected net present value of the patent over a 10-year term at a 7% discount rate is $412,000. Conversely, the expected value of keeping the invention as a trade secret under these replication conditions drops to just $28,000.

| Metric | Patent Path Value | Trade Secret Path Value | Winner |
| --- | --- | --- | --- |
| Total Prosecution Cost | $22,700 | $0 | Secret |
| Upfront License Fee | $75,000 | $0 | Patent |
| Royalty Income (Projected) | $240,000 | $0 | Patent |
| Milestone Income | $500,000 | $0 | Patent |
| Replication Risk (8-week) | Low (Legal Barrier) | High (70% Probability) | Patent |
| Expected Net Present Value | $412,000 | $28,000 | Patent |

## How to Choose Well in 45 Days

University technology transfer offices often default to trade secrecy for AI inventions, assuming that keeping the model weights or training data hidden preserves novelty. This is a fatal error in 2026. The prevailing assumption that secrecy offers a superior strategic moat ignores the reality of independent replication: if an architecture is describable in under 500 lines of pseudocode and the inference API exposes behavior, competitors can reverse-engineer the method in weeks. In this scenario, filing a U.S. provisional within 45 days of the notebook date—before any preprint—is mandatory. Secrecy here does not preserve Bayh-Dole compliance; it destroys licensing value by allowing the market to move on before the patent issues.

The decision matrix shifts only when the invention is structurally opaque. If the invention is a training-data curation script confined to an air-gapped cluster with fewer than six authorized users and no publication mandate, maintain it as a trade secret. This requires written NDAs and private-repo access logs, but only because the mechanism cannot be observed from outputs. However, even this protection evaporates if the PhD inventor graduates within 180 days or funding becomes federally encumbered. Departure destroys secrecy control regardless of lab policy; in these cases, file first to secure priority before the knowledge leaves the institution.

For commercialization, external pressure dictates the track. If an industry sponsor offers over $100,000 in advance or demands foreign rights, file a provisional plus Patent Cooperation Treaty (PCT) track immediately. This preserves Munich and Tokyo licensing leverage, which trade secrets cannot provide. Conversely, if the loss-function-only invention faces high abstract-idea risk under Section 101, use a dual-track split: file narrow technical-means claims on hardware improvements while retaining tuning hyperparameters as confidential know-how. This bifurcation mitigates rejection risk while protecting the core logic.

| Condition | Action | Rationale |
| --- | --- | --- |
| Architecture < 500 lines + API exposure | File Provisional (< 45 days) | Replication occurs in weeks; secrecy fails |
| Air-gapped curation script (< 6 users) | Maintain Trade Secret | Output does not reveal internal mechanism |
| Inventor departure (< 180 days) | File First | Departure destroys secrecy control |
| Sponsor Advance > $100k / Foreign Rights | Provisional + PCT | Preserves international licensing leverage |
| High Abstract-Idea Risk (Loss Function) | Dual-Track Split | Claims hardware means; keeps hyperparams secret |

Industry norms are shifting toward transparency. According to Microsoft's September 2026 provisional code of conduct, major players are imposing restrictions on AI models, signaling that open architectures are becoming the standard for licensing. Relying on secrecy in this environment yields zero expected value compared to the 18-month publication window of a provisional application.

## What to do next

| Step | Action | Why it matters |
| --- | --- | --- |
| 1 | File a U.S. provisional through your university tech-transfer office within thirty days of conception and before any arXiv post | Locks earliest priority date under 35 U.S.C. Section 111(b) before independent replication |
| 2 | Draft semantically broad, well-parsed claims for the university AI model architecture for USPTO filing | Blocks Big Tech design-arounds more profitably than lab secrecy |
| 3 | Treat USPTO publication under 35 U.S.C. Section 122(b) as a Section 102(a)(1) prior-art moat at the eighteen-month mark | Turns publication into a wall against later filers and triggers university AI licenses |
| 4 | Keep as trade secret only if the invention is an internal training-data curation pipeline that cannot be reverse-engineered from outputs | Preserves secrecy only where outputs reveal nothing and patent disclosure adds no moat |
| 5 | Screen intelligence co-optation risk flagged in Compromised Campus and the Kalaris Intelligence Conference at Georgetown University with tech-transfer counsel before filing | Protects academic independence from CIA money-and-influence ties while you pursue Augmented Intelligence combining human judgment with analytical power |

## Frequently Asked Questions

**How many days after filing do patent claims go public under the standard publication timeline?**

Fifty-four-six days after filing, your claims go public.

**What is the specific deadline for a Technology Transfer Office to report a federally funded invention to the funding agency under Bayh-Dole compliance?**

The TTO must report it to the funding agency within 60 days of inventor disclosure.

**What is the consequence for Patent Cooperation Treaty foreign rights if an arXiv preprint is posted before a provisional application?**

All Patent Cooperation Treaty (PCT) foreign rights are destroyed immediately upon that preprint.

**How many generative-AI patent families originated from China between 2014 and 2023 according to WIPO data?**

Of these, 38,210 originated from China.

**What percentage of university AI research expenditures are federally funded, making them subject to Bayh-Dole encumbrance?**

62% of university AI research expenditures are federally funded.

**What is the estimated number of office-action rounds required to overcome an initial Section 101 Alice/Mayo rejection for AI abstract-idea claims?**

That is not random examiner noise; it requires 2.1 office-action rounds to overcome.

## Quick answers

| What is the primary strategic outcome of the 18-month publication clock according to the article? | The 18-month clock serves as a computational prior-art moat that blocks Big Tech design-arounds more profitably than lab secrecy ever can. |
| --- | --- |
| Why does relying on trade secrecy alone risk violating Bayh-Dole compliance for federally funded inventions? | Secrecy does not satisfy the administrative disclosure requirement mandated by the iEdison reporting workflow, which risks forfeiture of title if the invention is not reported to the funding agency within 60 days. |
| How does an arXiv preprint affect international patent rights under the Patent Cooperation Treaty (PCT)? | All PCT foreign rights are destroyed immediately upon an arXiv preprint, leaving the invention stranded in the public domain without patent protection in most foreign jurisdictions. |
| What is the difference in replication defense between a patent strategy and a trade secret strategy? | A patent strategy provides a legal exclusion right against replication, whereas a trade secret strategy offers no defense because independent discovery or reverse-engineering is allowed. |
| According to the comparison table, what is the licensing revenue potential for a trade secret strategy versus a patent strategy? | The patent strategy has a licensing revenue potential of $2.9B based on AUTM 2024 data, while the trade secret strategy has negligible potential due to having no title. |

Also worth reading: **Second patent review process: 2026 claim construction loss vs rejection**: [Second patent review process: 2026](https://patentreviewpro.com/blog/second-patent-review-process-2026-claim-construction-loss-vs-rejection.php) · **Policymakers prioritize patent quality over volume as global filing numbers continue to soar**: [Policymakers prioritize patent quality over](https://patentreviewpro.com/blog/policymakers-prioritize-patent-quality-over-volume-as-global-filing-numbers-continue-to-soar.php) · **2026 USPTO AI Guidance: Neural Claims and EPO Costs**: [2026 USPTO AI Guidance: Neural](https://patentreviewpro.com/blog/2026-uspto-ai-guidance-neural-claims-and-epo-costs.php)

### Related reading

- [Trade Secret Protection in AI How Confidentiality Clauses Shape Patent vs

Trade Secret Decisions in 2024](https://patentreviewpro.com/blog/trade_secret_protection_in_ai_how_confidentiality_clauses_sh.php)
- [Understanding the 6-Month Response Window Timeline from Advisory Action to Patent Application Abandonment](https://patentreviewpro.com/blog/understanding_the_6_month_response_window_timeline_from_advi.php)
- [The Evolution of Flat Screen Technology From University Labs to Living Rooms](https://patentreviewpro.com/blog/the_evolution_of_flat_screen_technology_from_university_labs.php)
- [Functional patent claims rejected: 2026 34-point gap file split not fix](https://patentreviewpro.com/blog/functional-patent-claims-rejected-2026-34-point-gap-file-split-not-fix.php)
- [Trade secret litigation fuels demand for new generation of insurance products](https://patentreviewpro.com/blog/trade-secret-litigation-fuels-demand-for-new-generation-of-insurance-products.php)
- [Second patent review process: 2026 claim construction loss vs rejection](https://patentreviewpro.com/blog/second-patent-review-process-2026-claim-construction-loss-vs-rejection.php)

### Latest

- [Functional patent claims rejected: 2026 34-point gap file split not fix](https://patentreviewpro.com/blog/functional-patent-claims-rejected-2026-34-point-gap-file-split-not-fix.php)
- [Trade secret litigation fuels demand for new generation of insurance products](https://patentreviewpro.com/blog/trade-secret-litigation-fuels-demand-for-new-generation-of-insurance-products.php)
- [Second patent review process: 2026 claim construction loss vs rejection](https://patentreviewpro.com/blog/second-patent-review-process-2026-claim-construction-loss-vs-rejection.php)

Canonical: https://patentreviewpro.com/blog/university-patent-or-secret-18-month-publication-clock-vs-keep-secret.php
Markdown: https://patentreviewpro.com/blog/university-patent-or-secret-18-month-publication-clock-vs-keep-secret.php/index.md