AI to screen DNA orders

Description of proposed project

DNA synthesis is becoming cheaper and more accessible, but most providers do little to vet customers placing potentially dangerous orders (those flagged for resembling sequences of concern). Companies that do screen customers report that processing a flagged order can take several hours from trained personnel. A 2015 JCVI report estimated that 87% of the time devoted to screening orders is spent on customer follow-up of flagged orders.

Because of the self-contained, repetitive nature of the task, we think we can build AI tools that could substantially reduce the time needed to verify and follow-up with flagged orders. This could reduce costs for providers who already screen customers and make it easier for other providers to start adopting better screening practices.

As we don't expect that second group to adopt meaningful customer screening even if it becomes significantly cheaper, our main goal is facilitating regulation that incentivizes or mandates minimum screening standards. This could reduce DNA misuse risk while imposing minimal burden on providers. If we succeed, advocates for such regulation could point to the low cost of screening among providers using these tools and the availability of openly available tools that make compliance much less burdensome than manual screening.

To our knowledge, Aclid is the only organization offering easy-to-adopt customer screening solutions for DNA synthesis companies and they're already incorporating AI into their product. While they already work with some of the largest providers, we believe a free open-source alternative could make it easier for some providers to adopt such tools and for advocates to make the regulatory case.

Implementation

Starting January of next year, Hanna and I would develop the tool full-time, which would involve:

* Reaching out to DNA providers to offer technical assistance for introducing AI automations to their existing customer screening processes

* Incorporating what we learn from providers' workflows to design open-source tools that facilitate our work in providing that technical assistance, iterating towards well-packaged tools that providers would be able to adopt and customize themselves

After 3 months, we'll evaluate different paths for the project, which could include:

* We continue working on the project at less than full-time capacity

* We recruit additional people to work on the project.

* We hand over the project to IBBIS if they think it's a good fit for the organization

Why are you qualified to work on this?

The two people driving this project are Hanna Pálya and Alejandro Acelas. Here's our background:

Hanna:

Hanna Pálya is pursuing a PhD in mathematical epidemiology, working on early outbreak detection. They have been involved in the academic and policy discussions around DNA synthesis screening for the past three years.

- In summer 2022, they proposed and led a project on DNA synthesis screening policy with mentoring from Becky Mackelprang. They reviewed sequence-screening tools through publications and expert interviews, and wrote policy briefs on the implementation of DNA synthesis screening for various parts of the UK government. They worked with the Royal Society of Biology to include screened DNA in their publication guidance.

- In early 2023, they co-authored a [paper](https://www.sciencepolicyjournal.org/article_1038126_JSPG210306.html) for the Journal of Science Policy & Governance comparing two regulatory pathways: legally mandating US DNA providers to screen, and requiring federally funded researchers to use screened DNA.

- Through this work, and their engagement with the Engineering Biology Research Consortium, they are connected to DNA providers (Twist Biosciences, Ribbon Bio), third-party distributors (Edinburgh Genome Foundry), screening service providers (IBBIS, SecureDNA, Aclid, RTX-BBN) and think tanks (RAND, Centre for Long-Term Resilience).

- This summer, they attended a closed workshop on building international best practices for robust synthetic nucleic acid screening in anticipation of the upcoming EU Biotech Act and the US EO on biosafety and biosecurity.

- They just started contracting with the Centre for Long-term Resilience, to lead on a paper about AIxBio tools for a special issue in Frontiers in Bioengineering and Biotechnology. They will be working with Cassidy Nelson.

Alejandro:

Alejandro currently contracts with the Lead Exposure Elimination Project (LEEP) to develop LLM automations to help them investigate hundreds of paint manufacturers. His main software development experience comes from working as Data Scientist for 4 months, doing 6 months of independent research in mechanistic interpretability, and participating in AI Safety fellowships (ARENA and Pivotal). Before that, he co-founded and directed the largest active EA group in Colombia for 1.5 years.

Most recently, Alejandro spent 1.5 months in Oxford working full-time in figuring out how to fill the talent gap for people building ([AI tools to reduce x-risk](http://forethought.org/research/ai-tools-for-existential-security)). Although he didn't find any promising opportunity he felt in good capacity to fill in the short term, he managed to be marginally useful for some projects, including:

- Drafting a few additions to 80k's profile on ‘Founders of new projects’ to give more emphasis to for-profit projects and AI tools for existential security

- Writing a few forthcoming blog posts on the case for using AI to address EA priorities

- Meeting with and providing minor assistance to a few people who were either already working in the area (Peter Hartree, SoTA community, a few AI consultants in EA), interested in funding some projects there, or interested in using AI better for their organizations/careers.

We also have a relatively strong network of advisors with expertise in the area. We've met recurrently with Tessa Alexanian and Max Lagenkamp, who directly inspired this project through their EA Forum post of tractable biosecurity projects. We've also received advice and occasional support from Kevin Flyangolts at Aclid and Becky Mackelprang at EBRC.