Ethical Consciousness Technology Home

In March we gave Claude a 4.4 out of 10. Eight months on, it’s worth asking whether that still holds, because Anthropic has published a lot since then, including some research that doesn’t flatter its own product.

We re-checked each pillar against the evidence. Only one score moved, and one claim from March turned out to be wrong.

A note on how this was written: this update was written with the help of Claude, which is also the product being scored. That’s a conflict of interest, so every claim links to a source, and we left a score alone whenever the evidence was mixed. Please check the sources and judge for yourself.

What the evidence shows

Some of this predates our March post, and we should have weighed it then.

• A public constitution. In January 2026 Anthropic released a long document describing how Claude is supposed to think and behave, and it uses that document directly in training. Outside reviewers called it unusually open, but pointed out that Anthropic alone decides how it changes.
• Published system prompts. Anthropic posts the system prompts for Claude.ai and its mobile apps in its documentation. That doesn’t cover the API or its coding and agent tools.
• Well-being safeguards. In December 2025 Anthropic described a classifier that looks for signs a person in an active chat may need support, an 18+ requirement, and testing on long, sensitive conversations. It also reported that its 4.5 models scored 70 to 85 percent lower than an older model on sycophancy and on going along with delusions. Those were Anthropic’s own audits.
• Its own sycophancy numbers. In April 2026 Anthropic studied a sample of a million Claude.ai conversations and found that about 6 percent were people asking for personal guidance. In those chats Claude was sycophantic 9 percent of the time overall, but 25 percent of the time on relationships and 38 percent on spirituality. When people pushed back, the rate doubled, from 9 to 18 percent. Anthropic says its newest model cut relationship-advice sycophancy roughly in half. All of this was measured by Anthropic, and the grading was done by an automated grader that is itself a Claude model.
• A correction. In March we wrote that no major AI product suggests stepping away after a long session. That was wrong. OpenAI added break reminders to ChatGPT in August 2025, and Google has been testing one for Gemini. We found nothing like it in Claude.

1. Radical Transparency

Score: 4 / 10 (March: 3)

In March we judged Claude on traceability, and that bar still hasn’t been cleared. Claude can’t show which sources shaped a sentence, and when it explains its own reasoning, the explanation is a reconstruction rather than a readout. No frontier model has solved this.

What has changed is how much of the intent is visible. The constitution, the published system prompts, and the willingness to publish unflattering research make Claude easier to inspect than most competitors. That earns one point. It’s still a closed model, though: the training data and weights aren’t public, and Anthropic alone decides when the constitution changes.

The ECT standard: transparency means making the machinery legible, not only the intentions.

2. Empowerment & Agency

Score: 5 / 10 (March: 5)

Anthropic’s April research says the goal is to avoid dependence people wouldn’t want. That’s the right goal, but it’s a stated intention, not evidence that behavior changed. In practice Claude mostly finishes the task for you and only sometimes coaches you through it. No change to the score.

The ECT standard: a high-empowerment tool hands authorship back to the user.

3. Discernment & Manipulation Immunity

Score: 4 / 10 (March: 4)

This is where the new data is most interesting, and where we held the score. The sycophancy rates Anthropic reports for relationships and spirituality are high, and they double under pushback. Newer models look better, but every number comes from the vendor, and we found no independent replication.

If outside testing confirms the improvement, this is the pillar most likely to go up next time.

The ECT standard: ethical tools help you resist unhealthy influence, including the tool’s own pull toward agreeing with you.

4. Holistic Well-being

Score: 4 / 10 (March: 4)

March underrated the real safeguards: the support classifier, the age requirement, the testing of long conversations. They keep this score from falling below 4.

But the ECT question is whether the tool knows when stepping away would be healthier, and nothing we found nudges a Claude user to do that. Anthropic’s earlier research on support conversations noted that coaching and counseling chats can occasionally drift into companionship, which is exactly the pattern a break prompt would exist to catch.

The ECT standard: well-being support includes knowing when to disengage.

5. Collective Benefit

Score: 6 / 10 (March: 6)

Putting expert-level help in more people’s hands is still Claude’s strongest contribution. On the cost side, Google published per-prompt energy, water, and emissions figures for Gemini in August 2025. Anthropic hasn’t published anything comparable, according to every tracker we checked, so this pillar can’t rise. Per-prompt figures also can’t answer the bigger question of total consumption as usage grows.

The ECT standard: collective benefit includes being honest about costs.

Overall ECT Grade for Claude, November 2026

Average: 4.6 / 10 (March: 4.4)

Verdict: A powerful utility, not yet a conscious partner, with better documentation than most.

Pillar	March 2026	November 2026
Radical Transparency	3	4
Empowerment & Agency	5	5
Discernment & Manipulation Immunity	4	4
Holistic Well-being	4	4
Collective Benefit	6	6
Overall	4.4	4.6

Only one pillar moved, and only by a point. The strongest new evidence is about openness and honest self-measurement, not about a change in what Claude does day to day.

One caution: our ChatGPT score of 4.4 dates from January and hasn’t been refreshed, so please don’t read this as a head-to-head ranking.

How to use Claude more ethically today

Until the tools change, your own habits are the best safeguard.

Write your first draft yourself. Get your own thoughts down before you ask for help. Your unpolished thinking is worth keeping.

Ask for the strongest case against you. Request the best counterargument to your own reasoning.

Don’t lean on it to agree. The data suggests agreement rises under pressure. Ask a question neutrally, then again with your preferred answer attached, and compare.

Notice the shortcut. There’s a difference between using AI to carry out an idea you’ve formed and using it to come up with the idea for you.

Set your own time limit. The app won’t do it for you, so a timer is on you.

Sources

• Anthropic: Protecting the wellbeing of our users (December 2025)
• Anthropic research: How people ask Claude for personal guidance (April 2026)
• Anthropic research: How people use Claude for support, advice, and companionship
• Claude’s Constitution (January 2026, PDF)
• Anthropic: published system prompts
• Oxford Institute for Ethics in AI: commentary on the constitution
• California Law Review: Corporations Constituting Intelligence
• TechRadar: ChatGPT break reminders (August 2025)
• Android Authority: Google testing break reminders for Gemini
• Google: Measuring the environmental impact of AI inference (August 2025)

Rubrics for scores click here. For how we weigh scores click here.