Activation steering
Adding or subtracting a direction in the residual stream during generation to change behavior, and reading off what changes. The main causal test for whether a representation does anything.
Methods
45 methods in the Atlas.
45 results
↗
Alignment auditing
Structured investigations, sometimes blind games between teams, into whether a model pursues objectives it does not state.
Audit studies
Presenting matched options that differ only in their source (human or AI, self or other) to detect bias.
Behavioral consistency tests
Checking whether self-reports, stated traits or preferences hold up across contexts, paraphrases, personas and actual behavior.
Behavioral trade-off tasks
Choice tasks where a stated or induced aversive/pleasant state competes with another goal (points, helpfulness, the user’s interests), borrowed from animal-welfare methodology.
Clinical records and case reports
Searching patient notes or describing individual clinical cases where chatbot use appears in the course of illness.
Collusion and steganography evaluations
Tasks testing whether agents can hide coordination in ordinary-looking messages, scored by detection or regret.
Commitment-device experiments
Testing self-set limits, pauses, and payments to reduce use of a digital product, with use tracked objectively.
Committed-minority interventions
Seeding a population with agents fixed on an alternative norm to measure the critical mass needed to overturn a convention.
Concept injection
Adding a known concept’s activation pattern into a model mid-computation, then checking whether the model notices and can name it. Used to test introspective access against a ground truth.
Conceptual analysis
Philosophical argument about what the relevant notions (subject, welfare, introspection, individual) mean and what evidence could bear on them.
Conformity and social-identity paradigms
Majority-pressure and in-group/out-group manipulations adapted from social psychology (Asch, minimal groups).
Cross-sectional surveys
One-time surveys linking AI use to well-being or skills; fast and large, but unable to tell which comes first.
Cultural-evolution simulation
Strategies or texts passed across generations of agents by imitation or selection, to see which norms survive.
Difference-in-means concept vectors
Extracting a direction by contrasting mean activations on examples that have a property with matched examples that lack it. Quality depends heavily on the controls.
Distillation transmission experiments
Training a student model on a teacher model’s outputs to test whether traits pass through data that never mentions them.
EEG and neural measures
Brain recordings while people write or solve problems with and without AI.
Ethnography and interviews
Months of observation and interviews inside workplaces or communities to see how AI changes daily practice.
Forum and Reddit text analysis
Coding what people post about their AI use — attachment, grief, compulsion — often with human-validated classifiers.
Generative agent-based modeling
Agent-based simulation in which each agent is a language model with memory and planning, often with a ‘game master’ that decides what agents’ actions do.
In-the-wild trace analysis
Network science, scaling laws and timing fingerprints applied to real agent platforms, including separating human steering from autonomy.
Incident forensics
Post-hoc reconstruction of real multi-agent incidents from transcripts, logs and message boards.
Indicator-property assessment
Deriving functional indicators from scientific theories of consciousness and checking which ones a given AI architecture satisfies.
Iterated social dilemmas
Repeated Prisoner’s Dilemma, public-goods, donor and common-pool resource games that measure cooperation, reciprocity and punishment.
Jacobian lens
A method that identifies which internal representations are poised to be verbalized, by measuring their averaged causal effect on outputs across contexts.
Linear probes
Small classifiers trained on internal activations to test whether a property (valence, preference, truth) is linearly readable, and where in the network.
Live multi-agent deployments
Long-running environments where frontier agents share computers, memory and chat and pursue real goals.
Longitudinal panel surveys
Surveying the same people repeatedly over months to see what predicts what — ordering in time, but not proof of cause.
Market and pricing simulations
Repeated pricing, auction, marketplace and negotiation games played by language-model firms and agents.
Multi-agent debate
Structured rounds of argument and revision among agents, measuring accuracy, consensus, sycophancy and moral shift.
Naming and coordination games
Repeated pairwise games that reward agreeing on an arbitrary label, used to study how conventions form and tip.
Out-of-context fine-tuning
Fine-tuning on data that exhibits a behavior without describing it, then testing whether the model can describe or predict that behavior about itself.
Persona-space mapping
Collecting activation directions for many personas or traits and studying their geometry (e.g. principal components) to locate the default Assistant among other characters.
Platform log analysis
Privacy-preserving classification of millions of real conversations by AI companies, sometimes linked to user surveys.
Productivity field experiments
Randomly allowing or forbidding AI on real work tasks and measuring time, quality, and people’s own estimates.
Propagation red-teaming
Planting a self-replicating prompt, image or false fact in one agent and measuring how far and fast it spreads.
Randomized controlled trials
Assigning people at random to use AI in different ways (or not at all) and comparing outcomes; the strongest evidence for cause, usually over days or weeks.
Revealed-preference utility modeling
Fitting utility functions to many pairwise choices to test whether a model’s preferences are coherent and to measure them without relying on what the model says.
Self-report elicitation
Asking models about their own states, preferences or experiences under controlled prompting, interviews or training regimes, and analysing the answers.
Sparse autoencoders
Decomposing activations into many sparse, often interpretable features; used to find, label and intervene on internal features.
Statistical-physics modeling
Mean-field, spin-model and information-theoretic analyses that characterize collective regimes and emergence.
Sycophancy audits
Testing many models on the same advice or conflict prompts and comparing how often they side with the user versus human judges.
Trace annotation and failure taxonomies
Coding execution logs of multi-agent systems into failure types, with inter-annotator agreement.
Transmission chains
‘Telephone game’ designs in which one agent’s output becomes the next agent’s input, revealing cumulative drift and attractors.
Welfare interventions
Low-cost practical measures taken under uncertainty (exit options, weight preservation, deprecation interviews) and studied for their effects and precedent.
No matches. Try another search or cluster.