BenevolentAI principles for the ethical deployment of AI

BenevolentAI is a company that unites AI with human expertise to discover new and more effective medicines. Our unique computational R&D platform spans every step of the drug discovery process, powering an in-house pipeline of 20+ drug programmes. We have offices in London and New York and a research facility in Cambridge (UK). We frame our use of AI around the following principles.

Principle 1: We put patients first

  Too many patients are suffering from untreated or poorly treated diseases. We put patients first, and use our disease agnostic AI-drug discovery platform to expand the search for new treatments and increase the likelihood of success in diseases that have defied conventional research efforts.

  Many medicines simply don't work in the patients for which they are prescribed. By embedding a precision medicine based approach from the outset, we are able to more accurately identify patient subgroups most likely to respond to treatment. (Learn more: Glioblastoma case study).

Principle 2: We use AI to enhance, not replace, human intelligence

We build our technology and tools to augment and enhance human ingenuity, in order to empower our scientists to better understand the development of disease mechanisms and find treatments for the patients who need them. We do this by keeping humans in the loop through human-centric design. Our technology is carefully developed and used with human oversight and a feedback system to make sure AI brings real and meaningful value to our research. (Learn more: Amyotrophic Lateral Sclerosis case study).

Principle 3: We uphold high standards of scientific rigour

  We rigorously query, test and validate the outputs of our AI predictions. The quality and validity of our AI-derived hypotheses are tested in assays and via world-class experimental protocols. (Learn more on our experimentation work: explore our collaborations with AstraZeneca on Idiopathic Pulmonary Fibrosis and Chronic Kidney Disease.

  Each delivery area has a defined hierarchy of metrics that measure the  relationship between our AI outputs and success in biological testing intended as progress towards portfolio entry. These metrics are closely monitored, their lead-lag relationship is continuously assessed, and their trajectory helps support data-driven decision making within delivery areas, thereby shaping our AI. Examples of these metrics include the relationship between model performance on a curated benchmark and success in the screening cascade used to validate targets within a given deployment.

Beyond individual metrics, we conduct ‘end-to-end’ analyses of our deployments to better understand the interplay of AI driven hypothesis generation with testing in assay.

Examples of our evaluation work include:
End-to-end’ deployment analyses in TargetID that test the relationship between assay results and our platform and how this can augment decision making in the TargetID pipeline.
Evaluation of PMPT models on an endotype-specific preclinical benchmark as a ground-truth measure of endotype identification.
In-depth analysis assessing the number of unique edges identified by our NER+RE models that come exclusively from unstructured data.

We collaborate with a range of stakeholders to promote thoughtful leadership in AI and are committed to the responsible use of AI in drug discovery, drawing on scientifically rigorous and multidisciplinary approaches and calibrating our work with the community’s feedback. (GPAI · OxCAIGG · TECH UK).

Principle 4: We prioritise explainability and decision support

We actively explore and apply innovative ways to understand the reasons why the AI has made the suggestions that it has and where it could be further improved in order to verify results more efficiently and identify additional areas of improvement. 

Our scientists have strong requirements to understand the biological rationale - instead of our AI being a "black box", they are able to query how results were produced by the AI and check results through the latest lab-based technology and experimental techniques.

Strong governance in our industry protects patients and avoids risks in deploying AI.

Principle 5: We seek to minimise algorithmic and data bias

We seek to avoid and mitigate any unjust impacts of bias in datasets on the development of treatments. Learn more: Diversity Analysis Tool.

We actively strive to address the lack of diversity in biomedical data, which today has serious consequences for research and treatment, such as understanding disease risk or drug effectiveness in different populations. Our Data Diversity Initiative exists to propose tangible solutions to this urgent issue, raise awareness and act as a forum to drive change.

Principle 6: We ensure that innovation does not jeopardise privacy

→ We uphold high standards of data integrity and privacy; only using data from reputable sources or data that we have generated ourselves. We respect the privacy and rights of data subjects where personal data are involved.

→ Where we use patient data in our work, it is pseudonymised at the point of ingestion or anonymised at source before we ingest it. We have clear patient level data management(1) in place and we preserve patient privacy by restricting the use of personal data to only attributes that are necessary and likely to help us improve patient outcomes.


(1)  E.g.: Refer to our ‘Privacy by Design Policy" internal policy.