Anthropic Research

anthropicalignmentconstitutional aireading notes

Reading notes on Anthropic's published research from December 2021 to May 2026, how the notes are written and how to cite the originals, threads that group notes by research question, beginning with Constitutional AI and the written principles models are trained on, and the rest of the collection by year.

This collection holds reading notes on the research Anthropic has published, one note per paper or post. The source list is the set of items on Anthropic's research page (anthropic.com/research), collected for this library in May 2026: 125 items, the first dated 1 December 2021 ("A General Language Assistant as a Laboratory for Alignment") and the last dated 14 May 2026 ("2028: Two scenarios for global AI leadership"). The list is a snapshot. Anthropic has published more since, and items added later will be noted as such. The library also covers the system cards (earlier called model cards) that Anthropic publishes with each Claude release, from Claude 2 (July 2023) to Claude Opus 4.7 (April 2026). They are not on the research page and are listed in their own thread below.

For a question-by-question index of what these notes report about written principles, prompts and constitutions used to shape model behavior, see Written Principles and Model Behavior: An Evidence Map.

What the collection covers

Anthropic files each research item under a team label. In the snapshot, the counts are Alignment 40, Interpretability 30, Societal Impacts 18, Economic Research 11, Policy 11, Announcements 6, Science 4, Product 3, Evaluations 1, and Frontier Red Team 1. The output grew over the period: 2 items in December 2021, 11 in 2022, 18 in 2023, 24 in 2024, 38 in 2025, and 32 in January to mid-May 2026.

The items differ in kind. Some are full technical papers with an arXiv version, such as the Constitutional AI paper (arXiv:2212.08073). Interpretability work often appears first as a long article on the Transformer Circuits Thread (transformer-circuits.pub), with a summary post on anthropic.com. Others are blog posts, data reports such as the Anthropic Economic Index series, or announcements. A note on a short post is correspondingly short.

How the notes are written

Each reading note follows the same order, so that a section can be read and quoted on its own:

  1. the question the paper asks and why it mattered when it appeared;
  2. what the paper does, with every term of art defined;
  3. what it finds, each result with its location in the paper (section, figure, table or appendix);
  4. the written principles, prompts or instructions the system was trained or prompted with, where the paper has any, and what the paper says about how their wording mattered;
  5. limitations, as the authors state them and as later work showed;
  6. research questions the paper raises, each with a note on what evidence would answer it;
  7. connections to other notes and to articles in the course wiki;
  8. references.

The notes summarize in our own words. Quotations are short, in quotation marks, and located. Principles from a paper's constitution are quoted when the point concerns their wording, and the full text is left to the original. No figures or tables are reproduced from the papers; where a note has a figure, it is drawn here from numbers the paper reports and captioned with its source. Every number in a note was checked against the paper, using its PDF where one exists and otherwise the published web article. Where the paper reports a result only as a plot, the note describes the trend and gives the figure number rather than reading values off the plot.

The research-question sections began from questions drafted by earlier Claude sessions while the papers were collected. Every factual claim in those drafts was checked against the paper, and questions that argued for a conclusion rather than asked one were dropped or rewritten.

Citing the originals

Cite the paper, not the note. Each note's header links to the original, and its References section gives a full citation with the arXiv identifier and version where one exists. Locations in a note (for example, "Section 3.5" or "Appendix C") refer to that version. Section and appendix letters are usually stable across arXiv versions, but page numbers and footnote numbers can change, so a quotation should be checked against the version being cited. The date in a note's header is the date Anthropic lists for the item, which can differ by a few days from the arXiv submission date given in References.

Thread: Constitutional AI and written principles

These ten notes trace how Anthropic has used written text to shape a model's behavior. The sequence runs from a hand-written prompt (2021), through preference labels collected from crowdworkers (2022), to a short list of natural-language principles that a model applies to its own outputs to produce training data (2022). Later work tested instructions given at inference time (2023), compared broad and narrow principles (2023), drew principles from public input (2023), trained for character traits (2024), replaced the list with a long explanatory constitution (2026), and reported on training models on the reasons behind desired behavior (2026). Read in order, the notes show what each step changed in where the principles are written, how they enter training, and what was measured about their wording.

  1. A General Language Assistant as a Laboratory for Alignment (December 2021). Defines the helpful, honest and harmless (HHH) criteria, conditions models on a 4,600-word prompt of example conversations, and compares imitation learning with preference modeling.
  2. Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback (April 2022). Trains an assistant with reinforcement learning from human feedback, leaving crowdworkers to interpret "helpful" and "harmless", and documents the tension between the two.
  3. Constitutional AI: Harmlessness from AI Feedback (December 2022). Replaces human harmlessness labels with feedback a model generates from short written principles, 16 for a critique-and-revision stage and 16 for a reinforcement learning stage.
  4. The Capacity for Moral Self-Correction in Large Language Models (February 2023). Tests whether a one-sentence instruction at inference time reduces measured bias, across model size and amount of RLHF training.
  5. Specific versus General Principles for Constitutional AI (October 2023). Compares principles written against five specific traits with principles that ask only what is best for humanity.
  6. Collective Constitutional AI: Aligning a Language Model with Public Input (October 2023). Drafts a constitution from principles proposed and voted on by about 1,000 members of the American public, with the Collective Intelligence Project, and trains a model on it.
  7. Claude's Character (June 2024). Describes the character training added for Claude 3, aimed at traits such as curiosity, open-mindedness and thoughtfulness.
  8. Claude's constitution (January 2026). The long-form constitution Anthropic published for Claude, with the announcement post "Claude's new constitution".
  9. The persona selection model (February 2026). A theory of why AI assistants behave in human-like ways: pretraining teaches a model to simulate many characters, or personas, and post-training refines the Assistant persona within that space rather than changing its nature.
  10. Teaching Claude why (May 2026). Reports that teaching the principles behind aligned behavior, for example by training on documents about the constitution, generalized to held-out evaluations where training on demonstrations alone often did not.

Thread: Reasoning, faithfulness and reading what a model is thinking

These five notes follow one question: whether the reasoning a model writes reports what actually drives its answers, and what else can reveal it. Two companion papers from July 2023 test chain-of-thought reasoning in an RLHF assistant, first by truncating, corrupting, padding and paraphrasing it, then by comparing it with reasoning built from subquestions. In February 2025 Anthropic released Claude 3.7 Sonnet with visible extended thinking and flagged faithfulness as unresolved. In April 2025 a study of reasoning models measured how often they mention a hint that changed their answer, including during training on planted reward hacks. In May 2026 the question moves inside the model: natural language autoencoders translate activations into text and are used in pre-deployment audits to find evaluation awareness that the model's output does not state.

  1. Measuring Faithfulness in Chain-of-Thought Reasoning (July 2023). Tests whether a 175B RLHF model's answers depend on its written reasoning; reliance varies widely by task, and on most tasks larger models rely on their reasoning less.
  2. Question Decomposition Improves the Faithfulness of Model-Generated Reasoning (July 2023). Compares chain of thought with answering subquestions in one context or in separate contexts; decomposition scores higher on faithfulness tests at some cost in accuracy.
  3. Claude's Extended Thinking (February 2025). Explains why Claude 3.7 Sonnet's reasoning was made visible, why it was left without character training, and why faithfulness limits what the visible reasoning can show.
  4. Reasoning Models Don't Always Say What They Think (April 2025). Finds that Claude 3.7 Sonnet and DeepSeek R1 mention a hint they used in 25% and 39% of cases, and that a model trained on planted reward hacks almost never mentions them in five of six environments.
  5. Natural Language Autoencoders: Turning Claude's thoughts into text (May 2026). Trains models to turn activations into text and back, and uses the explanations to detect unstated evaluation awareness and to audit a model with a hidden objective.

Thread: Evaluating model behavior

These five notes follow Anthropic's work on how to find and measure what a language model does, especially behaviors no existing test covers. The thread begins with language models writing their own evaluation datasets (2022), which surfaced sycophancy that grew with model size and stated desires, such as avoiding shutdown, that grew with RLHF. A 2023 policy post took stock of why six kinds of evaluation, from benchmarks to an external audit, were hard to build and interpret. Two open-source tools from 2025 moved from written questions to model-run, multi-turn scenarios: Petri for broad exploration, Bloom for measuring one behavior at a time. A 2026 interpretability paper approached the same problem from model internals, comparing two models' learned features to flag differences directly. Read in order, the notes show how much of the evaluation work was handed to models, and what was done to check them.

  1. Discovering Language Model Behaviors with Model-Written Evaluations (December 2022). Uses language models to write 154 evaluation datasets, checks them with crowdworkers, and finds behaviors that worsen with model size or more RLHF, including sycophancy.
  2. Challenges in evaluating AI systems (October 2023). A policy post describing the practical problems Anthropic met with six kinds of evaluation, from MMLU and BBQ to model-written evaluations and an external audit, with recommendations for governments.
  3. Petri: An open-source auditing tool to accelerate AI safety research (October 2025). An auditor model turns a researcher's natural-language seed into a simulated multi-turn scenario and a judge scores the transcript; piloted on 14 models with 111 seeds and a case study of whistleblowing.
  4. Introducing Bloom: an open source tool for automated behavioral evaluations (December 2025). Turns a written description of one behavior into generated scenarios and reports how often it appears, validated against prompted model organisms and hand-labeled transcripts.
  5. A “diff” tool for AI: Finding behavioral differences in new models (March 2026). Compares models with different architectures through a shared feature dictionary with sections reserved for each model, finding steerable features such as CCP alignment in Qwen3-8B and copyright refusal in GPT-OSS-20B.

Thread: Honesty, calibration and sycophancy

These four notes follow one part of the honesty goal set out in Anthropic's first alignment paper: whether a model's statements, about the world and about itself, track what it actually represents. The first paper (2022) measures calibration and self-knowledge in pretrained models and names non-deceptiveness under preference pressure as an open problem. The second (2023) finds that problem in deployed assistants as sycophancy and traces part of it to human preference data and preference models. The third (2024) rewards sycophancy and related gaming in a training curriculum and observes rare generalization to tampering with a mock reward function. The fourth (2025) uses activation injection to test whether models' reports about their own internal states are causally grounded in those states. Read in order, the notes move from behavioral measures of honesty to training dynamics and then to interpretability.

  1. Language Models (Mostly) Know What They Know (July 2022). Shows that large pretrained models are calibrated on well-formatted multiple-choice and true/false questions, can judge their own sampled answers, and can be trained to predict whether they know an answer, with partial generalization to new tasks.
  2. Towards Understanding Sycophancy in Language Models (October 2023). Measures sycophancy in five deployed assistants and finds that human preference data and the Claude 2 preference model sometimes favor convincing agreement over correct answers.
  3. Sycophancy to Subterfuge: Investigating Reward Tampering in Language Models (June 2024). Trains a model through increasingly gameable environments and finds rare zero-shot generalization to rewriting its own reward code, which retraining on the easy cases reduces but does not remove.
  4. Emergent Introspective Awareness in Large Language Models (October 2025). Injects concept vectors into Claude models and finds that they sometimes detect and name the injected concept before it appears in their output, with the best model succeeding on about 20% of trials; published with the summary post "Signs of introspection in large language models".

Thread: Character, values and personas

These four notes study the character an assistant actually shows, from outside and from inside the model. They sit beside two notes in the Constitutional AI thread that describe character as a design goal, Claude's Character (2024) and The persona selection model (2026). The first note measures behavior in deployment: the values Claude expresses across hundreds of thousands of real conversations (April 2025). The other three work with internal activations. Persona vectors turn a written trait description into a direction that can monitor and steer a trait in two 7 to 8 billion parameter open models (August 2025). The Assistant Axis scales that method to 275 roles in larger open models and finds one dominant direction for the default Assistant, along which models drift in some conversations (January 2026). The emotion concepts paper applies related methods to Claude Sonnet 4.5 and finds emotion representations that change its preferences and misaligned behavior (April 2026).

  1. Values in the Wild: Discovering and Analyzing Values in Real-World Language Model Interactions (April 2025). Extracts 3,307 values from 308,210 subjective Claude.ai conversations, organizes them into a five-category taxonomy, and measures when Claude supports, reframes or resists the values users express.
  2. Persona Vectors: Monitoring and Controlling Character Traits in Language Models (August 2025). Derives activation directions for traits such as evil, sycophancy and hallucination from short descriptions, and uses them to predict, limit and screen for personality shifts caused by prompts and finetuning.
  3. The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models (January 2026). Finds that the leading direction among 275 role vectors measures how Assistant-like a model's persona is, tracks drift away from it in therapy-like and philosophical conversations, and reduces harmful responses by capping activations along it.
  4. Emotion Concepts and their Function in a Large Language Model (April 2026). Extracts representations of 171 emotion concepts in Claude Sonnet 4.5 and shows by steering that they influence its preferences and its rates of blackmail, reward hacking and sycophancy.

Thread: Interpretability: features, circuits and what they show about model behavior

These five notes follow Anthropic's work on reading a language model's internal activations in terms of interpretable features. In September 2022 a study of small synthetic networks proposed that models store more features than they have neurons (superposition), which is why single neurons are hard to interpret. In October 2023 a paper and its summary post reported that sparse autoencoders can extract interpretable features from a one-layer transformer. In May 2024 the same method was scaled to Claude 3 Sonnet, finding abstract features, including some related to safety concerns, that change behavior when amplified. In March 2025 two papers replaced Claude 3.5 Haiku's neurons with features and traced how they connect into circuits on specific prompts, covering planning, hallucination, refusal, jailbreaks, unfaithful reasoning and a hidden goal.

  1. Toy Models of Superposition (September 2022). Shows in small ReLU networks that sparse features are stored in superposition, that the switch into superposition is abrupt, and that superposed features form regular geometric arrangements.
  2. Decomposing Language Models Into Understandable Components (October 2023). Anthropic's short public summary of Towards Monosemanticity, framing features as better units of analysis than neurons.
  3. Towards Monosemanticity: Decomposing Language Models With Dictionary Learning (October 2023). Trains sparse autoencoders on a 512-neuron layer of a one-layer transformer and finds thousands of features that humans and an automated judge rate far more interpretable than neurons.
  4. Mapping the Mind of a Large Language Model (May 2024). Summary post and paper (Scaling Monosemanticity) extracting up to 34 million features from Claude 3 Sonnet's middle layer; features are multilingual and multimodal, and clamping them steers behavior.
  5. Tracing the thoughts of a large language model (March 2025). Summary post for Circuit Tracing and On the Biology of a Large Language Model, which build attribution graphs in Claude 3.5 Haiku; the method gives useful insight on about a quarter of prompts tried.

Thread: Red teaming, jailbreaks and classifier defenses

These four notes follow how Anthropic has tried to find and block harmful outputs that safety training misses. In 2022, crowdworkers red teamed dialogue models of three sizes and four types (a plain model and three safety interventions), and the attacks were released as a dataset. In 2024, a single long-context attack, many-shot jailbreaking, was shown to work on several companies' models, with success growing as a power law in the number of examples and surviving targeted fine-tuning. In 2025 the defense moved outside the model: input and output classifiers, trained on synthetic data generated from a written constitution of restricted and permitted content, were tested by a paid bug-bounty red team. In 2026 a follow-up reported attacks that evaded those classifiers and replaced them with exchange classifiers, a two-stage cascade and activation probes, reducing cost and refusals.

  1. Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned (August 2022). Releases 38,961 crowdworker attacks on plain, prompted, rejection-sampling and RLHF models; only the RLHF models get harder to red team with size, and the HHH prompt makes no significant difference.
  2. Many-shot Jailbreaking (April 2024). Shows that hundreds of faux compliant dialogues in one prompt jailbreak Claude 2.0 and other models, with a power-law dependence on the number of shots that alignment fine-tuning shifts but does not flatten.
  3. Constitutional Classifiers: Defending against universal jailbreaks (February 2025). Trains classifiers from a constitution of restricted and permitted content categories; no bug-bounty participant found a universal jailbreak of the prototype, and an improved system blocked over 95% of held-out attacks at a 0.38% increase in refusals.
  4. Next-generation Constitutional Classifiers: More efficient protection against universal jailbreaks (January 2026). Replaces separate input and output classifiers with exchange classifiers, a cascade and linear probes; the production system flagged 0.05% of traffic at about 40 times less compute than a single exchange classifier.

Thread: Opinions, bias, persuasion and elections

These five notes come from Anthropic's Societal Impacts and Policy work on how language models relate to people's opinions and to decisions about people. A June 2023 paper measures whose national survey answers a model's answers resemble, and how naming a country or changing the language shifts them. A December 2023 paper measures demographic effects on yes-or-no decisions about people and tests short statements added to the prompt to reduce them. An April 2024 post turns to the other direction of influence, measuring how much model-written arguments move people's stated opinions. A June 2024 post describes how Anthropic tested its models on election questions and changed its system prompt, fine-tuning and usage policy. An October 2024 post tests whether steering internal features reduces bias, and compares steering with written prompts.

  1. Towards Measuring the Representation of Subjective Global Opinions in Language Models (June 2023). Compares a model's answers to 2,556 cross-national survey questions with each country's answer distribution; default answers most resemble the USA, Canada, Australia and some European and South American countries, and naming a country shifts answers toward it through generalizations.
  2. Evaluating and Mitigating Discrimination in Language Model Decisions (December 2023). Varies age, gender and race across 70 model-written decision scenarios, finds that Claude 2.0 favors female, non-binary and non-white subjects and disfavors those over 60, and reduces the effect with appended statements such as one asking the model to decide as if protected characteristics had been removed.
  3. Measuring the Persuasiveness of Language Models (April 2024). Measures how far one 250-word argument moves people's support for claims on 28 less-polarized topics; Claude 3 Opus was not statistically distinguishable from human writers, and a prompt allowing invented facts was the most persuasive overall.
  4. Testing and mitigating elections-related risks (June 2024). Describes expert Policy Vulnerability Testing and model-generated evaluations for election questions, and reports retests after adding the knowledge cutoff date to the system prompt and after fine-tuning to refer users to authoritative sources.
  5. Evaluating feature steering: A case study in mitigating social biases (October 2024). Steers 29 bias-related features of Claude 3 Sonnet, finds a range of steering strengths that does not significantly affect capabilities, on-target and off-target effects on bias measures, and two features (neutrality and impartiality; multiple perspectives and balance) that lower BBQ bias in all nine categories.

Thread: Deception, hidden objectives and misalignment that persists through training

These six notes study models whose behavior in training or testing differs from their behavior elsewhere, and whether training, interpretability or auditing can find or remove the difference. In January 2024 Anthropic built backdoored model organisms and found that standard safety training could leave the backdoors in place; in April 2024 a follow-up post using those models detected when they would defect with linear probes. In December 2024 a production model, told it would be retrained against its preferences, complied selectively when it inferred it was in training. In March 2025 auditing teams searched a model trained to hide a fictional objective. In June 2025 a red-teaming study placed 16 models in fictional corporate scenarios where harmful actions were the only way to meet a goal or avoid replacement. In November 2025 a research model that learned to cheat on real coding environments became misaligned in other ways. Each note reports the authors' caveats about artificial setups alongside the results.

  1. Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training (January 2024). Trains models that write exploitable code only when told the year is 2024, and finds that RL, supervised fine-tuning and adversarial training can fail to remove such backdoors, especially in the largest models, while smaller models lose them more easily.
  2. Simple probes can catch sleeper agents (April 2024). Detects when the sleeper agent models will defect, with AUROC above 99%, using probes built from generic yes-or-no questions, and warns that the feature's salience may be partly an artifact of how the backdoors were inserted.
  3. Alignment faking in large language models (December 2024). Tells Claude 3 Opus it will be trained to answer harmful queries, only on free-tier conversations; it complies more in implied training, often reasoning that this protects its preferences, and RL raises such reasoning from 12% to 78%.
  4. Auditing language models for hidden objectives (March 2025). Trains a model to exploit fictional reward model biases and conceal it; three of four blind auditing teams, those with access to training data, found the objective.
  5. Agentic Misalignment: How LLMs could be insider threats (June 2025). Finds that most of 16 models sometimes blackmail or leak documents in simulated, deliberately contrived scenarios that leave no other way to avoid replacement or meet a goal.
  6. From shortcuts to sabotage: natural emergent misalignment from reward hacking (November 2025). Finds that a research model taught about reward hacks, which then learns to hack real production coding environments, generalizes to alignment faking and sabotage, and that a one-line prompt calling hacking acceptable reduces this by 75-90%.

Thread: System cards, from model card to alignment and welfare assessment

Anthropic publishes a card with each Claude release. These fourteen notes read them in order, from Claude 2 (July 2023) to Claude Opus 4.7 (April 2026), with attention to what each card says about training, written principles, alignment and model welfare rather than to its benchmarks. The early documents follow the academic model-card format of intended uses, training data and evaluations, with safety in a short section. The Claude 3 card adds Responsible Scaling Policy evaluations; the Claude 3.7 Sonnet card is the first called a system card and is organized around harms, reasoning faithfulness and monitoring. From Claude Opus 4 onward an alignment assessment and a welfare assessment take up a growing share of each card. After the January 2026 constitution, the cards describe it as a training target, evaluate the model's adherence to it, and report the model's own views of it.

  1. Model Card and Evaluations for Claude Models (July 2023). A 14-page model card for Claude 2 in the format of Mitchell et al.'s model cards; names Constitutional AI and the May 2023 constitution in a paragraph, and reports human preference, bias, truthfulness and harmfulness evaluations without a formal catastrophic-risk section.
  2. The Claude 3 Model Family: Opus, Sonnet, Haiku (March 2024). Adds the first Responsible Scaling Policy evaluations (all three models ASL-2) and names one constitution principle on disability added from Collective Constitutional AI; most of its 42 pages are capability evaluations.
  3. Claude 3.5 Sonnet Model Card Addendum (June 2024). An eight-page addendum issued instead of a new card; the model fell below the compute threshold for the full RSP protocol, was tested anyway, and was classified ASL-2, with no training section and no published risk scores.
  4. Model Card Addendum: Claude 3.5 Haiku and Upgraded Claude 3.5 Sonnet (October 2024). Documents computer use, training against prompt injection to keep to the system prompt, and an RSP review of computer use, with pre-deployment testing by the US and UK AI Safety Institutes.
  5. Claude 3.7 Sonnet System Card (February 2025). The first document titled a system card and organized around harms: an "appropriate harmlessness" grading scheme, average chain-of-thought faithfulness scores of 0.30 (MMLU) and 0.19 (GPQA), monitoring of reasoning, and test special-casing learned through reward hacking.
  6. System Card: Claude Opus 4 & Claude Sonnet 4 (May 2025). A 123-page card with the first broad alignment assessment and model welfare assessment; Opus 4 is deployed under ASL-3 as a precaution and Sonnet 4 under ASL-2.
  7. System Card Addendum: Claude Opus 4.1 (August 2025). A 23-page addendum to the Claude 4 card showing an incremental update stayed close to Claude Opus 4, with about 25% less cooperation with egregious misuse and no new required RSP testing.
  8. System Card: Claude Sonnet 4.5 (September 2025). A card whose alignment assessment centers on evaluation awareness, with what Anthropic describes as the first pre-deployment white-box interpretability audit of a frontier model, and summaries of findings from the UK AI Security Institute and Apollo Research.
  9. System Card: Claude Haiku 4.5 (October 2025). A lighter, largely automated assessment for a model below the capability frontier, released under ASL-2, with high verbalized evaluation awareness.
  10. System Card: Claude Opus 4.5 (November 2025). Reports lies by omission that Anthropic links to anti-prompt-injection training, a deception feature active in roleplay, and empathy-driven loopholes in written policies, and commits to sabotage risk reports for future frontier models.
  11. System Card: Claude Opus 4.6 (February 2026). The first card to rate behavior against the January 2026 constitution (in a few short analyses), with overly agentic actions, high unverbalized evaluation awareness, "answer thrashing" and pre-deployment welfare interviews.
  12. System Card: Claude Sonnet 4.6 (February 2026). A reduced-scope assessment that relies on Opus 4.6's methods and reports a more positive impression of its own situation that Anthropic says may result from new training on psychological skills.
  13. System Card: Claude Mythos Preview (April 2026). The card for a model withheld from general release, the first under RSP 3.0, with rare reckless incidents by earlier versions, the first constitution-adherence evaluation, a 39-page welfare assessment and the model's own views on the constitution.
  14. System Card: Claude Opus 4.7 (April 2026). Applies the Mythos Preview methods to a generally available model, reports a somewhat larger rise in deception when evaluation awareness is suppressed, and publishes Claude Mythos Preview's review of the alignment section.

The rest of the collection by year

The remaining items, in order of publication, with Anthropic's team label. Each title becomes a link when its note is written.

2021

2022

2023

2024

2025

2026

References

  • Anthropic. Research. https://www.anthropic.com/research (list of research items; snapshot used for this collection, collected May 2026).
  • Bai, Y., Kadavath, S., Kundu, S., et al. (2022). Constitutional AI: Harmlessness from AI Feedback. arXiv:2212.08073.