Introducing the Crosby Intelligence Fellowship

Contracts underpin every economic transaction. They are the rails of commerce. Yet negotiating them is still slow, expensive, and opaque.

AI has made its fastest progress in domains with verifiable answers: math, code, games like chess and Go. Legal negotiation is different. There is rarely a single right move — the information is incomplete, much of what matters goes unsaid, and the best play depends on who is across the table. It is closer to poker than to math: great attorneys read the other side, weigh what they will accept, and judge when to press, when to concede, and when to hold firm under uncertainty. They anticipate how a counterparty will respond, and adjust as the hand unfolds.

In order for legal AI to progress, the industry needs to get better at empiricism — measuring and defining quality rigorously while preserving the judgment that is central to great legal work. These challenges make legal AI the next research frontier, and the hardest problems won’t be solved by any one team. We want to engage the research community to solve them with us.

Crosby Intelligence is the research-focused arm of Crosby Legal, an AI-native law firm, which means we work every day on the problems (and with the domain experts) that this research needs.

• Format. Remote-friendly, six months.
• Compensation. A $25,000 stipend for the fellowship.
• Compute. $12,500 in Codex credits.
• Expert access. Regular sessions with practicing attorneys for problem framing, annotation, and evaluation.
• Data. Access to Crosby’s labeled (non-client) collection of sample contracts and histories.
• Publication. Fellows retain the right to publish their work.

By the end of the program, we aim for every Fellow to have produced a draft paper, benchmark, or open-source artifact.

The problems outlined below are loose guidance. We encourage you to propose your own ideas if they are at the frontier of legal AI and align with your research.

1. Reward models under expert disagreement

In code, you can verify the answer. In legal work, two senior attorneys redlining the same clause will often disagree, and both versions can be defensible. That disagreement is lost when averaged away by standard preference modeling. How do you build reward models that learn from expert disagreement and separate taste from error?

2. High-fidelity negotiation simulation

Great redlines anticipate the counterparty’s response, so great lawyers will rehearse against simulated counterparties. Out of the box, an LLM playing opposing counsel concedes too fast, fails to hold a position for strategic reasons, and doesn’t trade concessions the way a real attorney does. How do you align agents that behave like real opposing counsel from scarce negotiation records? How do you evaluate whether the simulator predicts what real counterparties actually do? What does an RL environment look like for negotiations?

3. The Surgical Editing Benchmark

A senior attorney reshapes an entire deal by changing a single word, while an LLM haphazardly rewrites the entire paragraph. Coding agents show the same pathology, and will refactor a file to fix a one-line bug. These are the same problem: minimal-diff editing under a sufficiency constraint. A robust benchmark must jointly measure minimality and whether the edit achieves the objective.

4. Knowing when to hand off to a human

The bar for verifying legal work is exceedingly high, and verification is expensive. An agent that can flag which parts of its own output need especially thorough review (instead of treating every line as equally trustworthy) would improve downstream quality, reliability, and efficiency. When should an AI legal agent escalate to a human, and how do you train that judgment from sparse expert-override data? More broadly: how do you give language models calibrated uncertainty?

5. Bring your own problem

If you’re working on something else in legal AI, contract intelligence, or automated negotiation, we’d love to see it. Strong submissions make progress measurable, explain why the obvious approach fails, and will generalize beyond a single product.

Fellows may be PhD students, postdocs, faculty, or independent researchers. You may be a good fit if you have a strong background in machine learning or NLP, can execute a research project independently while incorporating feedback, and are excited about problems where ground truth is messy and the data is confidential. No legal background required.

Submit your resume/CV and a one-page research proposal: the question, why existing approaches fall short, your approach, and what you’ll deliver.