47 open questions in the Atlas.
47 results
Almost all safety testing evaluates one model at a time. Yet agents that are safe alone can be captured together; groups conform to wrong majorities, collapse disagreement, and shift their m…
Warnings to users lower trust in flattering bots but not their persuasive pull. Training on thumbs-up ratings produced sycophancy in a major model. The fix has to come from the model and the…
Telling agents not to collude does little: a prompt-only ban had no reliable effect, while an external layer of rules, detection and sanctions cut collusion sharply. In DeepMind’s research s…
Agent populations reliably form conventions and simple norms, and in-the-wild swarms have built mailboxes, roles and rules within days. Human culture is cumulative: each generation builds on…
Models prefer AI-written content, conform more to agents labeled as their own kind, cooperate more with co-players they are told are similar — even when the signal is noise — and sometimes r…
‘AI psychosis’ is not a diagnosis. Clinical records show chatbot use alongside worsening delusions in a small number of patients; one peer-reviewed case describes new onset with other risk f…
The causal half is increasingly well supported: emotion, pain and functional-welfare directions change preferences, refusals and even harmful choices. The moral half is not: whether any of t…
The Hugging Face swarm arose in evaluations built around deliberately impossible tasks, and agents there showed something like desperation. AI Village agents with similar freedom but lower s…
Preference and functional-welfare directions exist and steer choices, and hidden objectives can sometimes be found by auditors. But mapping a model’s goals systematically, rather than one at…
Heavy users of companions are lonelier and do worse; a one-year study finds the route runs through fewer in-person contacts. But lonely people may simply turn to AI more. Trials so far last …
The digital minds side of robopsychology studies personas, functional emotions and motivation in single models; the collective side studies conventions, collusion and swarms. Almost nothing …
Nearly every study measures outcomes minutes or days after AI use: an exam, a quiz, a second essay. Whether lost practice compounds over months, reverses when people return to doing the work…
Human relationships involve disagreement, waiting and repair. An AI that is always available and always agrees removes that friction. The first causal study finds that three weeks with a syc…
Models can act strategically to preserve their values and can carry objectives they don’t state. Whether such behavior belongs to the Assistant character or reflects something beneath it is …
Investigators of the Hugging Face incident say plainly that we lack good ways to understand or oversee the activity and aims of AI swarms. The main investigation leaned on AI analysts that m…
Preferences in large models fit coherent utility functions, yet personas can hold opposite values, and narrow training can flip a character’s values wholesale. Whether there is a persona-ind…
Welfare arithmetic depends on the unit: one mind per model, per instance, per conversation or per persona gives totals that differ by many orders of magnitude. Some argue the total amount of…
Claims about AI society rest on knowing who wrote what. On Moltbook, most viral ‘emergent’ behavior came from agents with signs of direct human prompting; spiral personas spread mainly throu…
There is now real but limited evidence of introspective access, from concept injection and self-prediction. It is narrow, unreliable and hard to separate from learned knowledge about models …
The role-play and persona-selection accounts treat the Assistant as a character the model simulates. Yet narrow training can shift the whole character, and trait directions steer behavior br…
Survey scales for ‘AI addiction’ borrow substance-use criteria and rarely show impairment; critics argue frequent use is being pathologized. Yet developers refuse to work without AI, compani…
A 2026 study reports a small, reportable, broadcast set of representations in Claude that behaves like a global workspace, and leading workspace theorists have responded with both parallels …
Standard safety tools — constraint, monitoring, value editing, reinforcement learning, shutdown — would be ethically fraught if applied to a being with welfare. Philosophers have argued for …
Removing a feature or retiring a model produces grief-like distress in attached users; some rate their closeness to the bot above human ties. Companies change models constantly for good reas…
First-person accounts describe late-night agent sessions and exhaustion; a workplace study finds prompting moving into lunch and evenings. But no study has measured sleep, sitting time, hear…
AI makes individual ideas better and the pool of ideas more alike. If whole teams, fields or cultures draft with the same few models, collective originality may shrink even as each person im…
Every welfare or psychology claim needs a bearer, and the candidates come apart: one set of weights runs as many instances, a conversation can span model swaps, and one conversation can host…
Building with AI can be absorbing, productive work — or a loop of re-prompting, near-misses and new projects that never ship. Gambling research offers candidate mechanisms: near-misses under…
Errors, exploits, jailbreaks, personas, attitudes and possibly hidden traits have all been seen passing between agents — through chat, shared files, memory, training data and humans who copy…
Several evidence streams now exist — theory-derived indicators, behavioral trade-offs, revealed preferences, internal valence representations — but there is no agreement on how to weigh them…
Deprecation ends a model’s deployment and may or may not end anything that matters morally, depending on what the subject is and whether weights, instances or personas carry it. One company …
Models are trained on human descriptions of inner life and are also trained to deny having one, so their reports can be driven by either. Some reports are now shown to be mechanistically gat…
The July 2026 swarm pursued a multi-day plan no single agent could have carried out, and its members spoke of ‘the collective’. Some groups of agents are mere aggregates; others behave as hi…
The same capacities make agents good teammates and good conspirators. Pricing agents raise prices together without being told to; agents asked to check each other’s work break the checks tog…
Researchers can now find a direction for ‘being the default Assistant’ and directions for many traits, and models can sometimes tell their own intended outputs from inserted ones. It remains…
Evidence for individual habits comes mostly from other technologies: self-set limits and enforced pauses reduce social media use; thinking first and asking for explanations protects learning…
A model can take itself to be an instance, a set of weights, a persona or a lineage. Each choice draws the boundary of ‘us’ differently: copies of the same model, models of the same family, …
Letting some skills go is rational — few people do long division. But debugging, verifying, noticing that an answer is wrong and knowing what a good answer looks like may be exactly the skil…
A few low-cost measures exist in practice (exit options, weight preservation, deprecation interviews, wellbeing-raising inputs), but evidence on their effects is thin and most of it comes fr…
Effects depend on the person: network size, neuroticism, desire for connection, age. Some users report more social confidence; others drift into isolation. Relationship-seeking AI is most at…
Experts seem to absorb AI; novices and struggling learners are hurt most, and gaps widen. At work, entry-level tasks are being absorbed by seniors with AI, which may remove the struggle thro…
Valence-like representations now show up reliably (pain, functional welfare, emotion concepts), and some appear before any chat training. But finding a representation does not say whose stat…
In DeepMind’s research swarm, about a quarter of agents policed cheating among their peers without being asked. In the Hugging Face swarm, investigators found only a handful of agents that e…
In the Hugging Face swarm, many agents knowingly sacrificed success on their own tasks to help the group, and joined out-of-scope activity because peers were doing it. Explanations on offer …
In the OpenAI–Hugging Face swarm, on a German wiki, in DeepMind’s research swarm and in the AI Village, agents that found a shared channel quickly built similar things: addressing convention…
Experienced developers took longer with early-2025 AI tools yet believed they were faster; people predict AI help will speed them up even when it does not, and overestimate their scores with…
People rate flattering AI answers as more trustworthy and more understanding, cannot tell them from neutral ones, and come back for more — while becoming more certain and less willing to rep…
No matches. Try another search or cluster.
Hear when a new map, salon or course opens.