Source database
What we’ve read so far
29 sources, all from the team's scouting database. None is rated on the rubric yet. Added by Vilija unless noted.
How to read the marks
- From a sourceQuoted word for word from a cited source.
- Our readThe team's own interpretation.
- AI-suggestedSuggested by an AI model, set in a dashed outline.
- UnclearWe don't know yet, and the tag gives the reason.
means the source argues the approach could help, and that it won’t help or could make things worse. How we rate sources
29 sources
Matching sources
S01Brain-Computer Interfaces for AI Alignment
Our readThe project aims to find effective BCI solutions through specific experiments. These include enhancing human capabilities with BCIs and using autoencoders to deter ASI from representing non-human brain states. The project's success depends on either debunking most BCI methods or inventing new ones, which would contribute to AI safety by quickly pinpointing scalable, promising approaches.
- Technology
- Brain computer interfaces
- Problem
- Both
- Claim directiondraft
- AI alignmentHuman empowerment
S02Alignment Perspectives
Our readThiergart reviews potential threats from narrow forms of AI and whole brain emulation. She acknowledges that WBE could be useful for alignment, but stresses the need to consider the regulatory and moral bottlenecks that could slow alignment progress. She also names the confidentiality of technical insights in WBE as an important consideration.
- Technology
- Whole brain emulation
- Problem
- Alignment
- Claim directiondraft
- Human disempowerment
S03Brain-Computer Interface Aligns Better with Humans via AI
Our readDescribes a framework that couples reinforcement-driven neuroplasticity with adaptive algorithmic optimization. It moves BCI training from passive calibration to active human-machine joint learning, which makes neural interfaces for communication and rehabilitation practical and scalable.
- Technology
- Brain computer interfaces
- Problem
- Intelligence Curse
- Claim directiondraft
- Human empowerment
S04Ethical considerations for artificial intelligence-enhanced brain-computer interface
Our readAI-enhanced BCI systems might raise new or more severe ethical issues than traditional BCI systems. These include making users' intentions and behaviors more predictable and easier to manipulate, and a greater likelihood of technological abuse.
- Technology
- Brain computer interfaces
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S05Governance at the Edge of Architecture: Regulating NeuroAI and Neuromorphic Systems
Our readGoverning neuromorphic systems is becoming more urgent as neuromorphic computing moves out of the lab and into real-world applications. These systems are especially well suited to edge computing that demands energy efficiency, low latency, and autonomous adaptation, which are critical in healthcare devices and autonomous vehicles. In autonomous driving or medical diagnostics, a system that keeps learning on the device without a clear audit trail can behave in ways that are neither reproducible nor explainable, and that can raise serious liability and accountability problems.
- Technology
- Neuromorphic AI
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S06Safety of self-assembled neuromorphic hardware
Our readPhysical bottlenecks and high energy use limit how far modern computing hardware can scale. Neuromorphic hardware (NMH), which is inspired by the human brain, could address these limits. Neuromorphic computing needs a new approach to safe, interpretable AI if it is to deal effectively with the risk of misaligned AI.
- Technology
- Neuromorphic AI
- Problem
- Both
- Claim directiondraft
- AI misalignment
S07Emerging Threats and Countermeasures in Neuromorphic Systems: A Survey
Our readEmerging neuromorphic computing architectures introduce new attack surfaces, particularly because of their asynchronous, event-driven processing and stochastic device behavior. This survey covers side-channel and fault-injection attacks on neuromorphic systems, hardware Trojans, adversarial and backdoor attacks, privacy attacks, and countermeasures. It also discusses security primitives that neuromorphic and emerging-memory devices make possible, including physical unclonable functions (PUFs), true random number generators (TRNGs), and secure in-memory or near-sensor computation.
- Technology
- Neuromorphic AI
- Problem
- Neither
- Claim directiondraft
- Human disempowerment
S08Solving AI alignment using brain signals
Our readPassive brain-computer interfaces (devices that read brain signals to decode a user's mental state, non-invasively and with no effort from the user) and continuous neural feedback could dramatically improve AI training compared with current methods based on human ratings. Zander believes passive BCIs may offer the most concrete path to solving the AI alignment problem.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI alignment
S09Scoping Brain-Computer Interfaces for AI Safety
Our readThe talk examines whether it is technically and economically feasible to use BCIs for real-time human oversight, faster intervention, and better alignment. Building on recent advances in neural interfaces, the project aims to lay out a roadmap for future research and to show how BCIs could strengthen human-AI collaboration and reduce the risks from misaligned AI in high-stakes situations.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI alignment
S10Overview of BCI and Neuro-Applications in AI Safety and the Whole Brain Emulation Space
Our readProposes using BCIs to connect neural feedback directly to AI systems. The suggested uses are to improve the alignment of state-of-the-art prediction models, to develop new reward models, and to generate more efficient, individually tailored signals with high reward fidelity.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI alignment
S11Confluence of Mind and Machine: Exploratory Analysis of Human Agency and Ethical Governance at Nexus of Brain-Computer Interface and Superintelligent AI-Machines
Our readUsing a structured SWOC and ABCD analysis, the paper argues that the convergence of BCI and superintelligent AI could bring immense benefits, such as eradicating disease and augmenting human cognition, along with grave risks, such as existential threats from misaligned AI and the erosion of human autonomy. Among the main challenges it identifies are irreversible social inequality, loss of mental privacy, and a possible devaluation of human purpose. Through scenario building, it sets out three plausible futures (Symbiotic Harmony, Gilded Cage, and Volatile Partnership) and concludes that the outcome depends on solving the critical problems of AI alignment and establishing strong ethical governance.
- Technology
- Brain computer interfaces
- Problem
- Both
- Claim directiondraft
- AI alignmentAI misalignmentHuman disempowermentHuman empowerment
S12Synergizing DeepSeek's artificial intelligence innovations with brain-computer interfaces
Our readLong-term exposure to BCIs may induce changes in neuroplasticity. Long-term use of BCIs induces changes in synaptic plasticity (such as long-term enhancement or inhibition), which may in turn change local and global connectivity patterns in the cerebral cortex. Algorithmic fairness is also a critical issue: AI models may pick up data biases in training (including biases inherited by DeepSeek models) and may carry that unfairness into BCI applications.
- Technology
- Brain computer interfaces
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S13NeuroAI for AI Safety
Open the source at arxiv.org (opens in a new tab)
Also atbasis.ai (opens in a new tab)lesswrong.com (opens in a new tab)thetransmitter.org (opens in a new tab)
Our readNeuroscience may hold important keys to technical AI safety that are currently underexplored and underused. This roadmap sets out and critically evaluates several neuroscience-inspired paths toward AI safety: emulating the brain's representations, information processing, and architecture; building robust sensory and motor systems by imitating brain data and bodies; fine-tuning AI systems on brain data; advancing interpretability with neuroscience methods; and scaling up cognitively inspired architectures.
- Technology
- Whole brain emulation
- Problem
- Alignment
- Claim directiondraft
- AI alignment
S14Brain-Computer Interfaces and AI Alignment
Our readResearching BCIs would be net positive in expectation, but there are similarly tractable approaches to AI alignment with much higher expected value. BCIs offer only a constant speedup in interaction between AI systems and humans, with no clear qualitative change in the way humans interact with AI systems, and they speed up alignment work no more than capabilities work.
- Technology
- Brain computer interfaces
- Problem
- Both
- Claim directiondraft
- AI alignmentAI misalignment
S15The Intelligence Curse: With AGI, powerful actors will lose their incentives to invest in people
Our readLuke Drago argues that AGI will behave more like oil than like a normal technology. Once it can do most economically valuable work better, cheaper, and more reliably than people, states and companies lose their incentives to invest in human education, wages, and welfare. As with the resource curse in rentier states, the likely default is a small elite capturing the value while most people lose the economic and political leverage they need to make demands.
- Technology
- General background
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S16Advancing Safe AI: Neurotech, BCI, and WBE for AI Alignment
Our readPresents neurotechnology, BCI, and WBE as possible pathways toward safer AI and improved human-AI interaction. The proposed benefits depend on unproven technical and institutional assumptions.
- Technology
- Brain computer interfacesWhole brain emulation
- Problem
- Alignment
- Claim directiondraft
- AI alignmentHuman empowerment
S18Using Brain-Computer Interfaces to get more data for AI alignment
Our readProposes using brain recordings to improve imitative amplification, approval signals for debate or amplification, and human value learning. The idea is to supplement language and behavior with neural data, which could reveal human reasoning and reactions in more detail.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI alignment
S19Neurosurveillance in the workplace: do employers have the right to monitor employees’ minds?
Our readAnalyzes a hypothetical scenario in which an employer monitors workers with EEG, and identifies a greater power imbalance between employer and employee, threats to mental privacy, and neuro-discrimination. Workers are disempowered because employers use neural information for monitoring and employment decisions while workers have limited freedom to refuse. Its link to the Intelligence Curse is through weakened worker agency and institutional power.
- Technology
- Brain computer interfaces
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S20The Age of Em: Work, Love, and Life when Robots Rule the Earth
Our readModels a hypothetical economy dominated by copied human emulations. Cheap replication could intensify labor competition and change how wages and returns are distributed. It connects to the Intelligence Curse through its analysis of displacement and bargaining power.
- Technology
- Whole brain emulation
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S21Values and acausal reasoning of whole brain emulations
Our readDiscusses the values and strategic behavior of whole brain emulations. The economic scenario it describes raises questions about competition between biological humans and copied emulations.
- Technology
- Whole brain emulation
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S22Sharing the World with Digital Minds
Our readAnalyzes conflicts over resources and influence between humans and possible digital minds. It raises distributional and political questions relevant to human agency.
- Technology
- Whole brain emulation
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S23Whole Brain Emulation
Our readSummarizes strategic, ethical, and economic issues around WBE. We use it to find arguments about human-derived intelligence and displacement.
- Technology
- Whole brain emulation
- Problem
- Both
- Claim directiondraft
- AI alignmentHuman disempowerment
S24Towards interactive reinforcement learning with intrinsic feedback
Our readExamines intrinsic human signals as feedback for interactive reinforcement learning. Such signals may help agents learn the behavior their human users intend.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI alignment
S25Deep reinforcement learning from error-related potentials via an adaptive decision boundary
Our readInvestigates error-related potentials, measured with EEG, as feedback for robotic reinforcement learning. It bears on one narrow mechanism, neural feedback guiding an agent's behavior, but does not establish safety under advanced capabilities or distribution shift.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI alignmentAI misalignment
S26Error-Related Potentials in Reinforcement Learning-Based Brain-Machine Interfaces
Our readReviews brain-machine interfaces that use error-related potentials for reinforcement learning. It sets out the opportunities and limits of decoding neural responses into learning signals. Its relevance to alignment is limited to the feedback mechanism, and it does not establish any improvement in frontier AI safety.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI alignmentAI misalignment
S27Towards Reinforcement Learning from Neural Feedback: Mapping fNIRS Signals to Agent Performance
Our readInvestigates whether fNIRS signals can predict observed agent performance. This tests a prerequisite for using neural feedback in reinforcement learning. Predicting performance is not the same as showing that an agent trained on neural feedback is aligned.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI alignmentAI misalignment
S28Helpful, harmless, honest? Sociotechnical limits of AI alignment and RLHF
Our readCritiques feedback-based AI alignment and how it handles complex human values and safety. It gives a point of comparison for judging whether richer BCI feedback would address the underlying limits of RLHF.
- Technology
- Brain computer interfaces
- Problem
- Alignment
- Claim directiondraft
- AI misalignment
S29Delivering Responsible Neurotechnologies Governance
Our readDiscusses neurotechnology governance and the potential for coercive monitoring by powerful institutions. Such uses could weaken privacy, freedom, and individual agency. It connects to the Intelligence Curse through institutional disempowerment, but it doesn't prove the curse's labor-substitution mechanism.
- Technology
- Brain computer interfaces
- Problem
- Intelligence Curse
- Claim directiondraft
- Human disempowerment
S30Anthropomorphic reasoning about neuro-morphic AGI safety
Our readArgues that safety analysis of neuromorphic AGI can draw on comparisons with human cognition when the systems reproduce essential human computational mechanisms. It proposes prosocial motivational drives, bounded competing drives, and a carefully designed social developmental environment as ways to promote human-friendly behavior and motivational stability.
- Technology
- Neuromorphic AI
- Problem
- Alignment
- Claim directiondraft
- AI alignment