Before Fearn, I prosecuted and drafted patents at Morrison & Foerster (MoFo), one of the top Big Law firms. The economics were tough: limited client budgets and high billing rates meant we were at constant risk of going over-budget or getting our hours written off.

The core of IP practice — strategy and judgment — is where IP professionals add real value. Yet most of our time was consumed by formatting, boilerplates, dependency updates, and other mechanical tasks. We were expected to deliver airtight work at breakneck speed, with outdated tools that were never built for IP.

Worse, the cloud-based AI tools entering the market created new risks: inputting client inventions into ChatGPT or other public AIs can count as “public disclosure”, which can invalidate the patent you are trying to write. Confidentiality, the lifeblood of IP, was being compromised.

Why IP is uniquely difficult

In general, current legal AI tools are not attuned to the strangeness of intellectual property law. Such legal AI tools appear to fall in one of two buckets:

1. Bucket One: The AI tool is a general legal AI tool not designed for the unusual contours of IP law. Such tools use generic models that have been trained on all sorts of irrelevant materials, which result in outputs not suitable for IP law. These tools tend to serve lawyers that practice non-IP fields of law.
   
   
   

2. Bucket Two: The AI tool claims to be an IP-specific AI tool, but it is not actually designed for the unusual contours of IP law. Such tools still use generic models that have been trained on irrelevant materials. The tools in this bucket incur unique risks for their users that they otherwise wouldn’t have, had they served lawyers outside of IP law. For example, tools in this bucket share invention disclosure information to public AI servers and models, even when those models train on the disclosed information. A public model’s training data can often be tricked into regurgitating exact parts of its training data, such as the invention disclosures.

As alluded to by bucket two, current legal AI tools miss the general fact that IP law is strange. Drafting and prosecuting patents doesn’t require a JD, but does require a technical degree. Patent practitioners even take their own separate bar exam, as opposed to the commonly seen state bar exam (after all, patents are allowed only on a federal level—you get, for example, a US patent, and not a California patent). It’s within this strange IP world that privacy matters in a way that it simply doesn’t for other fields of law. Using the wrong AI tool isn’t just risky — it can destroy the very IP rights you’re trying to protect.

Building from the ground up

Building a proper AI tool for the IP legal community meant starting from scratch. Building on top of existing third-party AI models, such as ones from OpenAI, Anthropic, or Google, was a no-go for many reasons. First, such third-party models require cloud usage, because to run, they require seriously beefy and specialized computing hardware. The number of law firms (or even tech companies) that can run such general-purpose AI models is extraordinarily few. Second, these massive AI models are impossibly hard to steer, or “align”, as technical folks would say. The sheer size and complexity of such models means that fundamentally, even with very explicit prompt engineering, problematic AI behaviors are inevitable. Third, these massive AI models can verbatim regurgitate the content they were trained on. A clever user, when given freeform options via a chat interface, can trick the AI model into spitting out private training data. For the legal world, and especially the IP legal world, such regurgitation can be obviously disastrous. Fourth, massive AI models are prone to hallucinations. For the legal industry, and especially the IP legal industry, a silent hallucination in an IP matter can be a serious liability to both the draftsperson, and the integrity of the IP. 

For all these reasons and more, building a secure and private AI solution for the IP industry meant we needed to build our models from scratch. And not only did we need to build them from scratch, but we needed to build them in a way that would mitigate or eliminate the above issues. For this reason, at Fearn, we developed our own hypercompact foundation models built especially for IP law. By designing our models to be hypercompact, we address all the above issues: we circumvent the need for specialized computing hardware, we can eliminate the AI’s hallucinations, and we have total control over our models’ alignments. Having control over our alignments further meant that we could implement rules for preventing misbehaviors, such as content regurgitation. 

Building these hypercompact models was far from trivial. It required a specialized founding team: myself, a former Big Law patent prosecutor and Caltech PhD dropout, and Angela Gao, a Caltech PhD in computer science and math, with a deep research background in AI. Given our domain expertise and research backgrounds, we curated and labeled a set of high quality IP training data, and together with our team of PhDs from Caltech, MIT, and beyond, we trained our custom model architectures on our proprietary datasets. The results were spectacular. Our models outperformed every general-purpose AI on the market, when it came to tasks related to IP drafting. And it’s not just text: our image models also outperform competitors when generating patent-allowable figures.

AI Native — Not just a wrapper

Fearn isn’t a ChatGPT wrapper. We are truly AI native. We’ve leveraged our knowledge and experience to build our own in-house foundation models designed exclusively for the needs of IP law.

Our bespoke models allow for automating the mechanical layer of IP drafting, so lawyers can focus on what they do best: the creative core of their work, such as claims strategy and portfolio management. Fearn saves hours of repetitive work while protecting the integrity of innovation. And because our models can run fully on your infrastructure, we never need to touch the cloud. Your IP remains private and secure, always.

We’re just getting started. If you’re interested in bringing disclosure-safe AI into your IP practice, reach out — we’d love to work with you.

Fearn is not a law firm nor a substitute for an attorney or law firm. Communications between you and Fearn are protected by our privacy policy, but not by the attorney-client privilege or as work product. Fearn will provide certain technologies to help attorneys draft intellectual property. Fearn, however, does not and cannot provide any kind of advice, explanation, opinion, or recommendation about possible legal rights, responsibilities, remedies, defenses, etc.