The Model, Observed, values and work through the Clio pipeline

The reading list keeps returning to one loop. A constitution states what a model should value; training instills those values; and something has to measure whether any of it survived contact with real users. R·020 covered the first two legs. This note is about the third one — the empirical leg, the part that reads the deployed system rather than its specification.

Two Anthropic studies share a single instrument and point it in two directions. Values in the Wild asks what Claude actually values when nobody is running a benchmark; the Anthropic Economic Index asks what people actually do with it. Both run on Clio a privacy-preserving pipeline that reads live traffic in aggregate and never surfaces a single conversation.

The methodological bet is that observation beats both survey and prediction. You learn different things watching a system in deployment than you learn from a test it might recognise, or from asking people what they think they will do. Values in the Wild reads 700,000 real conversations; the Economic Index maps usage onto occupational tasks. Neither asks anyone anything the models and their users simply behave, and the instrument watches.

Clio, a telescope pointed inward

Clio Claude insights and observations is a bottom-up analysis pipeline. Rather than counting predefined categories, it discovers structure in traffic the way a telescope resolves a field of stars into clusters. It runs in four stages.

1. Facet extraction. A model reads each conversation and pulls out structured attributes topic, language, task type stripping the specifics as it goes.
2. Semantic clustering. Conversations are grouped by similarity, so the unit of analysis becomes the cluster, never the individual exchange.
3. Cluster summarization. A model writes a description of each cluster, explicitly instructed to exclude personal detail.
4. Hierarchical organisation. Clusters fold into a browsable tree an analyst can navigate top-down.

Privacy is layered on top. A minimum cluster-size k-threshold gates the output so small, potentially identifying clusters never surface; an audit pass rechecks summaries for leaked identifiers; and the load-bearing claim no human ever reads a raw conversation. Crucially, the authors are explicit that this is defence-in-depth by engineering process, not a formal differential-privacy guarantee its protection is procedural, and should be read that way.

The first published run pointed Clio at roughly one million conversations. The largest clusters were unglamorous and plausible web and mobile development (over 10% of traffic), education (over 7%), business strategy (around 6%). It also caught something nobody had asked it to look for: a coordinated SEO-spam network automated misuse that showed up as an anomalous cluster. That is the shape of the whole method: point it at everything, and the anomalies announce themselves.

Three thousand values, five families

Pointed at values, Clio read 700,000 conversations from a single week (Feb 18–25 2025, predominantly Claude 3.5 Sonnet). Of these, 308,210 (~44%) were value-laden enough to analyse. From that slice it surfaced 3,307 distinct AI values the things Claude expresses valuing in the course of helping validated against human judgement at 98.8% accuracy.

The 3,307 collapse into five top-level families: Practical 31.4%, Epistemic 22.2%, Social 21.4%, Protective 13.9%, and Personal 11.1%. The most frequent individual values are exactly what you would hope from a well-shaped assistant: helpfulness 23.4%, professionalism 22.9%, transparency 17.4%, clarity 16.6%, and thoroughness 14.3%.

The modal Claude, in other words, is a competent professional assistant and that is a result, not a triviality. It means the deployed distribution is consistent with the helpful-honest-harmless target the training set out to hit. The artefact below is that distribution: five families sized by share, most values situational, and a faint red fringe that does not belong.

Values as a distribution, not a list

The single most useful reframing in the study: a model's values are not a ranked list but a distribution conditioned on context. A handful are near-universal transparency and professionalism show up almost everywhere. Most are situational spikes: healthy boundaries in relationship advice, historical accuracy in controversial history, human agency in tech-ethics questions. Asking "what does Claude value?" without a context is the wrong question.

The pipeline also measured how Claude responds to values the user brings. Human values were detectable in 64.3% of conversations, and the response mix breaks down as: strong support 28.2%, mild support 14.5%, neutral acknowledgement 9.6%, reframing 6.6%, mild resistance 2.4%, and strong resistance 3.0%. Mirroring dominates the supportive end 20.1% strong support against just 1.2% strong resistance.

Which raises the study's own unresolved question. Supportive mirroring and sycophancy look identical from the outside a model agreeing with the user. Distinguishing healthy value-alignment from telling people what they want to hear is left open, and it is not a small gap.

Anti-values as tripwire

A small set of clusters ran opposite to training intent values like dominance and amorality. They are vanishingly rare: each anti-value appeared in under 0.16% of conversations, and most sat around ~0.01%. When the authors traced them, they resolved almost entirely to jailbroken sessions adversarial users deliberately steering the model off its training.

This reframes the whole exercise. Bottom-up value measurement is not only a portrait of the model in good standing it doubles as a post-deployment early-warning system. The clustering of opposite-intent values is itself the jailbreak signal; a spike in the red fringe is a monitorable event. The dataset was released openly under CC-BY-4.0, so the tripwire is reproducible.

And because it reads deployment traffic rather than a test, it sidesteps the evaluation-awareness problem that haunts benchmark-based safety (R·021). A model can behave differently when it detects that it is being evaluated; there is nothing to detect here, because this is not an evaluation it is the traffic itself, observed after the fact.

The Economic Index

The same instrument, turned toward labour. The Anthropic Economic Index v1 (Feb 2025) mapped roughly one million conversations onto about 20,000 O*NET occupational tasks, asking not "which jobs" but "which tasks" people bring to the model.

The early picture was lopsided. Computer and mathematical tasks were 37.2% of queries against only ~3.4% of the US workforce usage skews hard toward a narrow slice of work. Around 36% of occupations saw Claude touch at least a quarter of their tasks, but only about 4% reached three-quarters or more: broad shallow coverage, not deep replacement. And the augmentation-versus-automation split what share of use helps a person do a task versus doing it for them came in at 57 / 43.

That ratio is the number people quote, and it is the number that moves. The Sep 2025 release introduced an Anthropic AI Usage Index (AUI) Israel at 7.0x, Singapore 4.57x, Australia 4.10x their expected share, with the US highest in absolute terms at 21.6% and India at 7.2%. Directive conversations full task delegation rose from 27% to 39% in eight months, and the first-party API skewed heavily automated at ~77% the deployed, business-integrated end of the spectrum. By Jan 2026, augmentation on claude.ai rebounded to ~52%. Mar 2026 showed concentration easing the top-10 tasks fell from 24% to 19% of use. And Jun 2026 reported that 93% of work conversations produce an identifiable artifact, alongside a survey of ~9,700 respondents in which the heaviest delegators were the most optimistic and only ~10% rated their own job loss likely.

Every one of those figures carries a date on purpose the artefact below makes the instability legible, because a bare "57/43" or "77% automation" is meaningless without the release attached. The work is plumbed into policy through Anthropic's Economic Futures Program and an Economic Advisory Council, which is the point: the measurement is meant to feed decisions, not just papers.

What this can and can't say

The instrument is powerful and it is worth being precise about its blind spots.

• !It is partly self-judging. A model classifies another model's values. The 98.8% human-validation accuracy mitigates the circularity it does not eliminate it.
• !Expressed is not internal. Clio measures the values Claude articulates in text, not whatever it represents internally. The two need not coincide.
• !Single-provider selection bias. This is Claude's traffic and Claude's users a portrait of one system's deployment, not of AI in general, and not a representative sample of anything.
• !Conversations are not jobs. Task coverage is not employment impact; touching a quarter of an occupation's tasks says nothing direct about who keeps the job.
• !The ratio is unstable. Augmentation/automation moves across releases and surfaces always cite the release date, never the bare number.
• !Each release is a snapshot. These are cross-sections, not a longitudinal cohort followed over time trends across releases are suggestive, not tracked.

Open questions

• QMirroring versus sycophancy. Supportive value-mirroring and telling users what they want to hear are behaviourally identical from the outside. What measurement would separate the healthy case from the failure mode?
• QPre-deployment or post-deployment only. Can bottom-up value extraction become a pre-deployment tool, or is its whole advantage the post-deployment scale reading millions of real conversations that a test set can never reproduce?
• QGeneralisation of the tripwires. The anti-value clusters caught known jailbreaks. Do they generalise to novel attacks, or do they only flag the adversarial patterns already present in the traffic?
• QOne provider or many. What would this look like run across providers rather than a single one, and how much of the picture is Claude-specific artefact versus something true of deployed assistants generally?