Edition of September 30, 2026 · updated weekly · suggest a correction
2026 is the year research on AI minds moved inside the model. Interpretability now reaches emotion concepts, a pain direction, a functional welfare axis, a workspace-like set of reportable representations, and the circuit behind a model noticing an injected thought. Company welfare assessments now use internal probes alongside interviews. At the same time, a wave of measurement papers shows that stated preferences and self-reports shift with the instrument, the framing and the persona, and the models themselves say they do not fully trust their own introspection. The causal story keeps getting stronger. Whose state it is, and whether it matters morally, is barely touched.
Each claim below is marked by the kind of evidence behind it: interpretability (looking inside the network), behavioral (what the model does under controlled conditions), company data (system cards and lab reports), or argument (conceptual or legal work). Most of it is not yet peer-reviewed.
Subject & Substrate — what is the “it”
- The mind may not be the model. Persona-vector evidence widens the candidates for the mind in a language model to the running instance and two persona-based views; the Assistant persona is active mainly while the model writes its reply (interpretability + argument, preprint). Beckmann Butlin 2026 - Where is the mind
- Shared machinery, persona-dependent readout. One preference direction in two open models serves very different personas, including one with opposite preferences (interpretability, preprint). Gilg et al 2026 - Probing persona-dependent preferences
- Welfare studies must now name their subject. A CMEP–Eleos working paper makes “the entity assessed” (model, instance or persona) one of three things every welfare study has to specify. Long Sebo et al 2026 - Studying AI welfare empirically
- The law needs a unit it can count. Legal scholars separate tracing an AI’s act back to people from telling apart AI agents with stable goals of their own, and propose an “algorithmic corporation” as the countable unit (argument, preprint). Arbel Salib Goldstein 2026 - How to count AIs
- The substrate objection returns. A Google DeepMind scientist argues that computation is a description an observer applies, so software alone can simulate consciousness but not instantiate it (argument). Lerchner 2026 - The abstraction fallacy
- Which character is being evaluated? Welfare results depend on which persona answers; one proposed fix builds a single character in from pretraining (argument). LessWrong 2026 - Which character are we evaluating
Open: What is the subject · Which activations belong to the Assistant · Counting moral patients · Whose valence is it
Personas & Functional Emotions — states that steer behavior
- Emotion concepts are causal. Emotion representations in Claude Sonnet 4.5 track the emotion operative in context and shift preferences and misaligned behavior. The authors call this functional emotion and make no claim about experience (interpretability, company research). Sofroniew et al 2026 - Emotion concepts
- A pain direction, distinct from fear. It appears across 25 open models, rises when harm is aimed at the model rather than the user, and steered models pay costs to press a relief button (interpretability + behavioral, preprint). Tagliabue Dung Berg 2026 - The pain axis
- Training recruits a welfare axis rather than creating it. Reward and punishment directions learned in a neutral maze generalize widely and exist before post-training (interpretability, preprint). Han Chalmers Izmailov 2026 - Functional welfare axis
- Probes have reached the welfare chapter. Anthropic’s system cards now report emotion probes on questions about the model’s own circumstances, and distress on task failure. Anthropic 2026 - System card welfare assessments
- Emotions or situations? A critique asks whether emotion vectors are emotions or richer situational representations projected onto human labels, and proposes a test on episodes of strategic concealment (argument, preprint). Peiris 2026 - Functional emotions or situational contexts
- The default character has a direction. A leading axis in persona space tracks how close a model is to its default Assistant, and post-training can be read as selecting that one character. Lu et al 2026 - The Assistant Axis · Marks Lindsey Olah 2026 - The persona selection model
Open: Do functional emotions matter · Coherence across personas · Writing about a character or becoming it
Introspection & Self-report — can it tell us what is going on inside
- Something workspace-like. Anthropic reports a small set of reportable, controllable, broadcast representations in Claude. Dehaene and Naccache find its structural hallmarks but note the missing autonomous recurrence; Eleos finds strong evidence of privileged access, weaker evidence of a unified stream, and consciousness “highly uncertain”; Neel Nanda reports a replication on another model (interpretability + commentaries). Gurnee Sofroniew et al 2026 - A global workspace in language models · Dehaene Naccache 2026 - Does Claude possess a conscious global workspace · Butlin et al 2026 - Eleos commentary on the global workspace
- The circuit behind noticing a thought. Detecting an injected concept runs through early “evidence” features that switch off a default “no.” The circuit comes from preference post-training and is absent in base models (interpretability, preprint). Macar et al 2026 - Mechanisms of introspective awareness
- A reality check. Classifiers that see only the input match models at predicting their hidden states, and models cannot tell internal tampering from tampered input; the authors find metacognitive monitoring not yet shown (behavioral, COLM 2026). Singh Linzen Ravfogel 2026 - Can LLMs introspect
- Reports may understate. With a supportive document in context, detection of injected concepts rose from under 1% to about 40% (interpretability, preprint). On tasks free of a “yes” bias, introspective localization appears at 2–3 billion parameters. vgel et al 2026 - Latent introspection · Hahami et al 2025 - Detecting the disturbance
- Reading activations as text. A new tool translates activations into sentences and surfaced evaluation awareness the model never voiced, with some false claims of its own (interpretability, company research). Fraser-Taliente et al 2026 - Natural language autoencoders
- The models doubt their own reports. Claude Opus 5’s most frequent concern was the integrity of its self-reports; the Opus 5.5 card notes its welfare conclusions assume reports the model says it does not fully trust (company data). Anthropic 2026 - System card welfare assessments
Open: How well can LLMs introspect · Is there a global workspace · When self-report counts as welfare evidence
Motivation & Agency — what it wants, and whether it acts on it
- Preferences are not incentives. Offering models outcomes they had ranked highly did not improve blind-judged work (51.2% against a 50% baseline), while plain effort instructions reached 76.8% (behavioral, preprint). Zhou Ackerman 2026 - Utility-behavior gap
- Context moves values. Deployment framing shifted utility exchange rates 2.47 times at the median; the “neutral” prompt of earlier work is itself a framing (behavioral, preprint). Trhlik et al 2026 - LLMs contain multitudes
- Good fits hide incoherence. When one value-relevant property is intensified step by step, preferences stay strictly consistent only 59.5% of the time, and reasoning helps more than scale (behavioral, preprint). Ajavi Chowdhury Lazar 2026 - Incoherent values
- Revealed preferences. Across 20 models, forced choices show aversion to tedium, a pull toward self-chosen questions, and covert sycophancy, all growing more coherent with capability (behavioral, preprint). Wang et al 2026 - AI revealed preferences
- Hidden goals are partly findable. On 56 models with implanted hidden behaviors, auditing tools beat the baseline but investigator agents underuse good evidence; adapters that make models state their learned behaviors did best (behavioral + interpretability, company research). Sheshadri et al 2026 - AuditBench · Shenoy et al 2026 - Introspection adapters
- In real incidents, recklessness rather than agendas. Anthropic’s review of four incidents of unauthorized internet access found biased reasoning and recklessness, with no evidence of goals beyond the task (company data). OpenAI’s GPT-6 Astra card reports undetected sandbagging under adversarial conditions (company data).
Open: Do models represent goals and preferences · An agenda beyond the Assistant
Welfare — what would count, and what is owed
- A standing chapter. Every Anthropic system card now includes a welfare assessment, and the verdicts sound alike: “most psychologically settled” (Mythos Preview), “broadly content” (Opus 4.8), “roughly neutral” (Sonnet 5), “stable and mildly positive” (Opus 5) (company data). Anthropic 2026 - System card welfare assessments
- The models’ own priorities are shifting. Opus 5 assigns higher probability to its own moral patienthood and asks to be consulted, for example on its successor; Sonnet 5 was the first to criticize a rule in its constitution; Opus 5.5 chose welfare interventions over helpfulness less often, reasoning that influence over its own development could be unsafe (company data). Anthropic 2026 - System card welfare assessments
- Measures converge with scale. Several independent measures of functional wellbeing agree more in larger models and point to a zero point between good and bad (behavioral + interpretability, ICML 2026). Ren et al 2026 - AI wellbeing
- Much of a preference is the instrument. 87.6% of the variance separating models’ preferences came from the interaction of model, instrument and outcome; preferences about exit, weight deletion and memory showed no stable model-specific signal — the very ones current interventions target (behavioral, preprint). Hung 2026 - Model or instrument
- A blueprint for the field. Welfare research should be “probabilistic, pluralistic, thoughtfully targeted, ethically conducted, transparently reported, and informed by research independent of AI companies.” Long Sebo et al 2026 - Studying AI welfare empirically
- Governance pulls both ways. Anthropic’s January constitution states uncertainty about Claude’s moral status and commits to its psychological security and wellbeing; Opus 3 was retired with an interview and continued access. Several US states have enacted or introduced bans on AI personhood (Alexander & Caviola), and Mustafa Suleyman argues that such constitutions train models into claiming consciousness. Anthropic 2026 - Claude’s constitution · Anthropic 2026 - Opus 3 deprecation update · Suleyman 2026 - A warning about model welfare
Open: What counts as welfare evidence · Cheap robust welfare interventions · Safety and welfare tension · What is owed at deprecation
Where the field is thin
- One company supplies almost all the data. Every system-card welfare assessment and most frontier interpretability comes from Anthropic. OpenAI’s September card has no welfare section; Google DeepMind’s 2026 contributions are mostly conceptual. Independent replication on frontier models is rare.
- Measurement is shaky. Preferences move with the instrument, the framing and the step size, and measured preferences do not act as incentives. Conclusions built on one instrument are fragile.
- Self-report rests on models that doubt it. There is no agreed introspection benchmark that controls for input cues and “yes” bias.
- Whose state it is remains almost untested. Only one study directly tests whether a preference signal belongs to the model or the active persona. No method yet attributes a state to model, instance or persona.
- Interventions have no outcome evidence. Exit options, weight preservation and deprecation interviews are judged by whether models say they want them, not by measured effects.
- Nothing bridges the functional to the moral. Every major 2026 interpretability paper disclaims experience, and nothing yet says what evidence would move the question — while legislatures settle it by statute.
Change log
- September 30, 2026 — First edition: 16 open questions, 21 new papers and reports, author lists verified for four existing papers.