OpenAI vs Anthropic: Ethics Compared (ChatGPT vs Claude)

Anthropic scores higher in Mashinii’s stored comparison, but neither company earns an unconditional ethical endorsement. The underlying scores date from 28 April 2026: Anthropic averages 0 and OpenAI approximately −14.3 across seven scored dimensions. This article reviews newer evidence separately, through 8 October 2026. It does not present the April scores as a fresh October assessment.

If you are choosing between ChatGPT and Claude, the useful questions are specific: how are your conversations used, what limits govern military deployments, how are creators and workers treated, and what evidence supports each company’s safety and environmental claims?

The comparison at a glance

Explore the evidence behind the OpenAI company profile and Anthropic company profile. Company scores are analytical assessments, not a head-to-head test of every ChatGPT or Claude model.

What the scores say—and their limits

We checked the stored records on 8 October. All the underlying dimension records used below are dated 28 April 2026. The mean uses seven numeric scores for each company; four other dimensions contain no numeric score and are excluded, rather than treated as zero.

Anthropic leads in six of these seven dimensions; OpenAI leads in Honest & Fair Business. Our earlier text incorrectly reported OpenAI’s mean as −17, said Anthropic led in only one dimension, called both averages negative, and contradicted the safety score in its own table. OpenAI’s seven scores sum to −100, giving a mean of −14.3 when rounded to one decimal place.

An average can also hide important differences. Anthropic’s positive safety score offsets negative business-conduct and environmental scores. A reader who places particular weight on copyright or military participation may reach a different practical decision from someone prioritising safety governance. Missing scores do not prove that an issue is irrelevant or that no harm exists. See the methodology before treating the aggregate as a recommendation.

Military contracts: ceilings, safeguards and disputed uses

In July 2025, the US Chief Digital and Artificial Intelligence Office announced awards to four frontier AI companies, including both labs, each with a $200 million ceiling. A ceiling is not proof that $200 million was paid, and this announcement should not be conflated with later classified-deployment agreements. CDAO announcement.

In its February 2026 agreement announcement, updated on 2 March, OpenAI describes restrictions on domestic surveillance, autonomous weapons and certain automated decisions, supported by contractual and technical controls. This is OpenAI’s account of its safeguards, not an independent audit of their effectiveness. OpenAI agreement statement.

Anthropic’s 5 March statement says it received a supply-chain-risk designation letter on 4 March after its dispute with the government over restrictions on uses of Claude. It said it would challenge the designation in court. That dated statement establishes the company’s position at that point; it does not establish the latest outcome of every subsequent proceeding. Anthropic statement.

The meaningful ethical questions are which uses are prohibited, who can detect violations, who can stop a deployment, and what happens if the customer disputes a restriction. It is too simple to label one company’s safeguards “product-level” and the other’s “only contractual.” Nor does opposing particular military applications mean declining all defence work.

Copyright: separate settlement, liability and future conduct

The $1.5 billion Bartz v. Anthropic settlement received final approval on 20 July 2026. As checked for this review, the administrator expected initial distributions by 15 November, with some claims taking longer. Approval and payments to claimants are different stages. Settlement administrator and final-approval announcement from the Authors Guild.

We no longer describe this as a single completed payment that closed Anthropic’s copyright exposure. A settlement addresses specified parties and claims. Its existence does not establish that every training source, output or future use is lawful.

For OpenAI, this revision removes an unsupported total of active lawsuits, a damages figure and a sweeping claim about favourable judgments. A count of filed cases is not a measure of proven liability; allegations, interim decisions, final judgments and settlements must be identified separately. Absence of a comparable settlement in this article is not evidence of a clean copyright record.

For a creator or publisher choosing a tool, useful questions include the provenance of training material, licensing arrangements, rights-holder remedies and the terms governing your own uploads. A corporate ethics score cannot answer those questions for a particular work.

Safety governance and worker voice

Anthropic’s current Responsible Scaling Policy page lists version 3.4, effective 8 July 2026, and an August risk report. It also links its reporting and anti-retaliation policy. Our previous reference to version 3.0 as current was outdated. Anthropic policy and version history.

OpenAI also publishes a Preparedness Framework and deployment-safety material. Its August 2026 discussion of critical cybersecurity capabilities describes additional safeguards and changes it intends to make to its framework. These documents should be evaluated as evolving governance commitments, not guarantees that a model cannot cause harm. OpenAI’s account of critical cyber capabilities.

Worker reporting deserves the same distinction between a policy and its implementation. OpenAI’s January 2026 policy provides anonymous reporting and says personnel may report to outside authorities without advance notification. It would therefore be misleading to describe the company’s present policy solely through earlier disputes. OpenAI Raising Concerns Policy.

The stronger test for either company is whether workers can raise concerns without retaliation, whether independent reviewers receive enough information, and whether serious findings change deployment decisions. Certification names or a single harmlessness percentage do not resolve those questions. We have removed unsupported or inadequately contextualised claims of that kind from this comparison.

Labour: the people behind the model

TIME’s January 2023 investigation reported low pay and distress among Kenyan workers labelling harmful content for OpenAI through Sama. The report also includes company responses and describes the contract’s early termination. It is historical evidence, not a description of every current contractor. TIME investigation.

The ethical test extends beyond the consumer-facing company: pay, workload, informed consent to disturbing tasks, psychological support and a credible route to complain all matter. Outsourcing a task does not remove the need to examine these conditions.

This review does not establish an equivalent comparison across all of Anthropic’s suppliers. That gap cannot be turned into a positive finding that its supply chain is free of harm. A zero in the stored labour score is not a supplier audit.

Privacy: compare the product you actually use

OpenAI’s documentation distinguishes consumer services from business products and says business inputs and outputs are not used for training by default. Anthropic’s privacy documentation likewise distinguishes consumer and commercial offerings and explains how model-improvement permissions affect consumer chats. OpenAI data-use explanation and Anthropic training-data explanation.

Before uploading sensitive material, check the specific plan, model, training setting, retention rules, administrator access and any connected third-party tools. “Not used for training” and “not retained” are different promises. A brand-level ranking does not replace checking the terms that apply to your account.

Environment: ask for comparable measurements

OpenAI’s Stargate Community announcement and Anthropic’s electricity-cost announcement both make commitments about avoiding higher energy bills for local consumers. Those are relevant commitments, but they are not evidence of zero emissions or zero water impact. OpenAI announcement and Anthropic announcement.

The earlier article also misattributed the estimate of roughly 0.32 millilitres of water per ChatGPT query. That number comes from Sam Altman’s June 2025 blog post, not from an independent researcher’s estimate. Altman’s post.

A meaningful comparison needs a common boundary: which model and task, which data centre, which cooling system, whether electricity-generation water is included, and whether training or only inference is counted. An average per-query estimate cannot by itself rank two companies’ total environmental impacts. We do not claim this review supplies a complete, independently verified emissions inventory for either lab.

Which is the more ethical choice?

Anthropic leads in Mashinii’s dated score snapshot. The newer evidence does not justify declaring either company universally ethical or safe. Both participate in military work, publish safety commitments and operate resource-intensive AI services. The issues differ in their evidence, timing and relevance to your use.

Use the comparison to identify your priorities, then examine the OpenAI research and Anthropic research dimension by dimension. Recheck product privacy settings before sharing material, and assess the quality of evidence behind a claim rather than choosing solely on reputation. For wider context, explore how Mashinii evaluates companies and our latest research.

Correction and review, 8 October 2026: corrected the interpretation of the April scores; updated policy and settlement information; distinguished contract ceilings from payments; corrected the water-estimate attribution; and removed unsupported lawsuit totals, universal ESG-coverage claims and overconfident conclusions. Company statements are identified as such. The underlying dimension scores were not changed; their aggregate arithmetic was corrected.