Research Engineer – Human Influence
<div class="content-intro"><h2>About the AI Security Institute</h2>
<p>The AI Security Institute is the world's largest and best-funded team dedicated to understanding advanced AI risks and translating that knowledge into action. We’re in the heart of the UK government with direct lines to No. 10 (the Prime Minister's office), and we work with frontier developers and governments globally.</p>
<p>We’re here because governments are critical for advanced AI going well, and UK AISI is uniquely positioned to mobilise them. With our resources, unique agility and international influence, this is the best place to shape both AI development and government action.</p></div><h2><strong><span data-contrast="none"><span data-ccp-parastyle="heading 2">The deadline for applying to this role is Sunday 11th October 2026, end of day, anywhere on Earth.<br></span></span></strong><strong><span data-contrast="auto"><br>Team</span></strong><span data-ccp-props="{}"> <strong>Description</strong></span></h2>
<p><span data-contrast="none">The Human Influence (HI) team focuses on the ways in which AI can influence human beliefs, decisions, and behaviour. A substantial class of AI risk operates through people. AI systems can persuade people to change their beliefs and to take action; can build trusting relationships with people in order to exploit them; can extract private information from them; and can hold delegated ownership of high-stakes decisions.</span><span data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"> </span></p>
<p><span data-contrast="none">Our work is highly interdisciplinary, drawing on methods from computational social science, AI safety and security, cognitive and behavioural science, machine learning, and data science. Typical projects include running rigorous human-AI interaction studies and randomised controlled trials, building evaluations and benchmarks that track AI capabilities across model releases, eliciting model capabilities through fine-tuning and self-play, and developing datasets to monitor real-world risk exposure and severity.</span><span data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"> </span></p>
<h2><strong>Role Description</strong></h2>
<p><span class="TextRun SCXW255428478 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW255428478 BCX8">We are </span><span class="NormalTextRun SCXW255428478 BCX8">looking for a </span><span class="NormalTextRun SCXW255428478 BCX8">Research Engineer to join </span><span class="NormalTextRun SCXW255428478 BCX8">the Human Influence team</span><span class="NormalTextRun SCXW255428478 BCX8">. </span><span class="NormalTextRun SCXW255428478 BCX8">Successful candidates will be strong researchers and engineers with </span><span class="NormalTextRun SCXW255428478 BCX8">a track record</span><span class="NormalTextRun SCXW255428478 BCX8"> of </span><span class="NormalTextRun SCXW255428478 BCX8">carrying</span><span class="NormalTextRun SCXW255428478 BCX8"> out</span> <span class="NormalTextRun SCXW255428478 BCX8">scalable</span> <span class="NormalTextRun SCXW255428478 BCX8">work in LLM post-training and fine-tuning</span><span class="NormalTextRun SCXW255428478 BCX8">, especially with Reinforcement Learning</span><span class="NormalTextRun SCXW255428478 BCX8">;</span> <span class="NormalTextRun CommentStart CommentHighlightPipeRest CommentHighlightRest SCXW255428478 BCX8">or</span><span class="NormalTextRun CommentHighlightRest SCXW255428478 BCX8"> w</span><span class="NormalTextRun CommentHighlightRest SCXW255428478 BCX8">ith</span> <span class="NormalTextRun CommentHighlightRest SCXW255428478 BCX8">comparable </span><span class="NormalTextRun CommentHighlightRest SCXW255428478 BCX8">expertise</span><span class="NormalTextRun CommentHighlightRest SCXW255428478 BCX8"> in </span><span class="NormalTextRun CommentHighlightRest SCXW255428478 BCX8">engineering and </span><span class="NormalTextRun CommentHighlightRest SCXW255428478 BCX8">validating</span> <span class="NormalTextRun CommentHighlightRest SCXW255428478 BCX8">large-scale evaluation pipelines.</span></span><span class="EOP Selected SCXW255428478 BCX8" data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"> </span></p>
<p><span class="EOP Selected SCXW255428478 BCX8" data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"><span class="TextRun SCXW39204035 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW39204035 BCX8">Projects the Research Engineer might deliver include:</span></span></span></p>
<ul>
<li><span class="EOP Selected SCXW255428478 BCX8" data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"><span class="TextRun SCXW39204035 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW39204035 BCX8"><span class="TextRun SCXW195800030 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW195800030 BCX8">Designing and building </span><span class="NormalTextRun SCXW195800030 BCX8">a </span><span class="NormalTextRun SCXW195800030 BCX8">Reinforcement Learning</span><span class="NormalTextRun SCXW195800030 BCX8"> environment a</span><span class="NormalTextRun SCXW195800030 BCX8">imed</span><span class="NormalTextRun SCXW195800030 BCX8"> at</span><span class="NormalTextRun SCXW195800030 BCX8"> mitigating </span><span class="NormalTextRun SCXW195800030 BCX8">a model’s ability to e.g. deceive </span><span class="NormalTextRun SCXW195800030 BCX8">a</span><span class="NormalTextRun SCXW195800030 BCX8"> user</span> <span class="NormalTextRun SCXW195800030 BCX8">i</span><span class="NormalTextRun SCXW195800030 BCX8">n a one-to-one conversation or within multi-agent threads.</span></span><span class="EOP Selected SCXW195800030 BCX8" data-ccp-props="{"335557856":16777215,"335559739":0}"> </span></span></span></span></li>
<li><span class="EOP Selected SCXW255428478 BCX8" data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"><span class="TextRun SCXW39204035 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW39204035 BCX8"><span class="TextRun SCXW75218379 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW75218379 BCX8">Leveraging </span><span class="NormalTextRun SCXW75218379 BCX8">state-of-the-art</span> <span class="NormalTextRun SCXW75218379 BCX8">interpretability methods to </span><span class="NormalTextRun SCXW75218379 BCX8">iden</span><span class="NormalTextRun SCXW75218379 BCX8">tify</span><span class="NormalTextRun SCXW75218379 BCX8"> why</span><span class="NormalTextRun SCXW75218379 BCX8"> models </span><span class="NormalTextRun SCXW75218379 BCX8">exhi</span><span class="NormalTextRun SCXW75218379 BCX8">bit</span><span class="NormalTextRun SCXW75218379 BCX8"> con</span><span class="NormalTextRun SCXW75218379 BCX8">cerning </span><span class="NormalTextRun ContextualSpellingAndGrammarErrorV2Themed SCXW75218379 BCX8">behaviour</span><span class="NormalTextRun ContextualSpellingAndGrammarErrorV2Themed SCXW75218379 BCX8">,</span><span class="NormalTextRun ContextualSpellingAndGrammarErrorV2Themed SCXW75218379 BCX8"> and</span> <span class="NormalTextRun SCXW75218379 BCX8">designing mitigations</span><span class="NormalTextRun SCXW75218379 BCX8"> that c</span><span class="NormalTextRun SCXW75218379 BCX8">an </span><span class="NormalTextRun SCXW75218379 BCX8">be applied to models irrespective of training regime.</span></span><span class="EOP Selected SCXW75218379 BCX8" data-ccp-props="{"134233117":false,"134233118":false,"201341983":0,"335551550":1,"335551620":1,"335557856":16777215,"335559685":720,"335559737":0,"335559738":0,"335559739":0,"335559740":279,"335559991":360}"> </span></span></span></span></li>
<li><span class="EOP Selected SCXW255428478 BCX8" data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"><span class="TextRun SCXW39204035 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW39204035 BCX8"><span class="TextRun SCXW12239187 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW12239187 BCX8">Building the scalable system architecture underpinning the repeatable delivery and analysis of model evaluations and benchmarks.</span></span><span class="EOP Selected SCXW12239187 BCX8" data-ccp-props="{"134233117":false,"134233118":false,"201341983":0,"335551550":1,"335551620":1,"335557856":16777215,"335559685":720,"335559737":0,"335559738":0,"335559739":0,"335559740":279,"335559991":360}"> </span></span></span></span></li>
<li><span class="EOP Selected SCXW255428478 BCX8" data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"><span class="TextRun SCXW39204035 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW39204035 BCX8"><span class="TextRun SCXW171712400 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW171712400 BCX8">Delivering ambitious, engineering-heavy research projects on Human Influence topics, for instance by </span><span class="NormalTextRun SCXW171712400 BCX8">leveraging</span><span class="NormalTextRun SCXW171712400 BCX8"> post-training techniques on a large compute cluster.</span></span><span class="EOP Selected SCXW171712400 BCX8" data-ccp-props="{"335557856":16777215,"335559739":0}"> </span></span></span></span></li>
</ul>
<h2><span class="EOP Selected SCXW255428478 BCX8" data-ccp-props="{"335557856":16777215,"335559738":240,"335559739":240}"><span class="TextRun SCXW39204035 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun SCXW39204035 BCX8"><strong><span class="TextRun SCXW233327487 BCX8" lang="EN-GB" data-contrast="none"><span class="NormalTextRun CommentStart CommentHighlightPipeRest CommentHighlightRest SCXW233327487 BCX8">Who </span><span class="NormalTextRun CommentHighlightRest SCXW233327487 BCX8">we're</span><span class="NormalTextRun CommentHighlightRest SCXW233327487 BCX8"> looking for</span></span><span class="EOP Selected SCX
This listing is from Greenhouse. View original listing ↗