# Experiments with Detecting and Mitigating AI Deception

arXiv: https://arxiv.org/abs/2306.14816 · HTML: https://mlsanity.com/paper/2306.14816/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 13 of its 17 references that are in the map.
- First posted on arXiv 26 Jun 2023.

## Builds on (most cited first, 13 of 13)

- [GPT-4 Technical Report](https://mlsanity.com/paper/2303.08774/) (2023)
- [Safe Reinforcement Learning via Shielding](https://mlsanity.com/paper/1708.08611/) (2017)
- [Is Power-Seeking AI an Existential Risk?](https://mlsanity.com/paper/2206.13353/) (2022)
- [Safe Multi-Agent Reinforcement Learning via Shielding](https://mlsanity.com/paper/2101.11196/) (2021)
- [Safe Reinforcement Learning via Shielding under Partial Observability](https://mlsanity.com/paper/2204.00755/) (2022)
- [Towards Formal Definitions of Blameworthiness, Intention, and Moral Responsibility](https://mlsanity.com/paper/1810.05903/) (2018)
- [Discovering Agents](https://mlsanity.com/paper/2208.08345/) (2022)
- [Path-Specific Objectives for Safer Agent Incentives](https://mlsanity.com/paper/2204.10018/) (2022)
- [Reasoning about Causality in Games](https://mlsanity.com/paper/2301.02324/) (2023)
- [Shielding Atari Games with Bounded Prescience](https://mlsanity.com/paper/2101.08153/) (2021)
- [Correct-by-Construction Runtime Enforcement in AI -- A Survey](https://mlsanity.com/paper/2208.14426/) (2022)
- [Dynamic Shielding for Reinforcement Learning in Black-Box Environments](https://mlsanity.com/paper/2207.13446/) (2022)
- [Definitions of intent suitable for algorithms](https://mlsanity.com/paper/2106.04235/) (2021)

## Abstract (arXiv)

How to detect and mitigate deceptive AI systems is an open problem for the field of safe and trustworthy AI. We analyse two algorithms for mitigating deception: The first is based on the path-specific objectives framework where paths in the game that incentivise deception are removed. The second is based on shielding, i.e., monitoring for unsafe policies and replacing them with a safe reference policy. We construct two simple games and evaluate our algorithms empirically. We find that both methods ensure that our agent is not deceptive, however, shielding tends to achieve higher reward.