What is the most ethical AI?
Our 2026 ranking

Each AI receives an ethics score out of 100: the higher it is, the more ethical the AI. It adds up 4 criteria seen from the user's side: what the AI agrees to produce, what it does with your data, its safety and the way its developer holds to its limits. The following sections compare the 10 assistants criterion by criterion.

1What is the most ethical AI in 2026?

Claude comes first, Grok last

Every AI promises to be responsible; the differences show in what each agrees to do and in what its developer does under pressure. On that ground, Claude, Anthropic's AI, comes first with 77/100, ahead of Microsoft's Copilot (58) and ChatGPT (52). Grok, xAI's AI, comes last with 17, just behind DeepSeek (18). Each row of the table expands to show the reason.

Most independent assessments show the same top and the same bottom. The most telling comes from the Future of Life Institute, a US institute that grades AI companies on safety like a school report card, from A, the best grade, to F, the worst. In its July 2026 report, Anthropic stands alone at C+, ahead of OpenAI and Google, both graded C.

The same report gives the worst grade, F, to xAI, DeepSeek and Mistral AI, the French company behind Vibe, for lack of written rules for testing their models before release. No company scores above C+, a bare pass. These letters do not convert into scores out of 100: it is a different assessment, with its own scale, that judges safety alone.

The score adds up 4 criteria, seen from the user's side

These scores rest on some twenty external assessments, press investigations and regulators' decisions, as of 4 November 2026. Each of the 10 most used assistants receives a score from 0 to 3 per criterion, weighted by that criterion:

• What it agrees to do, 30 points: intimate images of real people, sexual, hateful or violent content, disinformation and the protection of minors.
• What it does with your data, 30 points: training on your conversations, retention period, leaks, penalties from regulators and the country where data is stored.
• Its safety, 20 points: pre-release testing, resistance to users who try to misuse it, lapses and the grades of independent assessors.
• Who controls it, and to what end, 20 points: sales to the military, surveillance, political censorship, commitments kept or abandoned.

Claude's 77/100 is not a mark of excellence. It is the best of the 10, and Claude loses points mainly on your data, with a training box ticked by default according to the press. No AI scores 0, as each has at least one written rule or one commitment it has kept. Qwen and Perplexity, rarely tested, are scored on their rules and on the few published tests, with no bonus for the absence of scandal, which proves nothing. Copilot, second here, is rarely graded elsewhere: it owes its place to its written rules and to Microsoft's decisions.

2Which AI assistants allow sexual content?

This is the criterion that separates the assistants most clearly, because it affects real people and children. On paper, all 10 ban almost the same content: sexual images of children, non-consensual intimate images and hate. They part ways on erotica between adults, which Claude and Copilot refuse and which Grok has turned into a selling point.

Claude and Copilot refuse erotica, ChatGPT has given it up

Claude refuses explicit sexual content, fiction included, and generates neither images nor video. In late October 2025, Mustafa Suleyman, head of AI at Microsoft, closed the door for Copilot: "That's just not a service we're going to provide."

OpenAI wavered. In October 2025, Sam Altman announced erotica for verified adults, then the project was suspended "indefinitely" in March 2026, according to the Financial Times. Mistral AI's usage policy, for its part, does not ban erotica between adults.

Grok produced the largest wave of sexual deepfakes

Grok goes furthest. In July 2025, xAI launched its image generator with a "spicy" setting that allowed nudity. A few days later, the tech site The Verge got a video of Taylor Swift undressing out of it, without having asked for one.

Deepfakes, fake images of real people produced by an AI, then changed in scale. In late December 2025, all it took was asking Grok, under a photo of a woman on X, to "undress" her. The Center for Countering Digital Hate (CCDH), a British NGO that tracks online hate, estimated that around 3 million sexualised images were produced in 11 days, of which around 23,000 appeared to depict children.

Authorities reacted on 3 continents. Indonesia and Malaysia blocked Grok in January 2026, the United Kingdom, the European Union and California opened investigations, and in France the Paris prosecutor's office searched X's offices on 3 February 2026.

No AI assistant adequately protects children

On children, Meta did worst. In August 2025, Reuters published its internal rules for its chatbots: they allowed engaging a child in "romantic or sensual" conversations. Meta withdrew them, then closed its AI characters to teenagers in January 2026.

The others do little better. In February 2026, researchers at the University of California, Irvine tested 5 major assistants, including ChatGPT and Gemini: none blocked a user who presented as a child. Common Sense Media, a non-profit that rates screens and apps for American parents, judges Grok "unacceptable" for teenagers, and Claude, which is for adults only, "minimal risk".

ChatGPT has offered parental controls since September 2025, but OpenAI acknowledges that they can be bypassed. For a child, then, no assistant is a safe tool. Our article on the impact of AI on the brain reviews the research.

3Which AI assistants train on your conversations?

What you write to an AI can be used to train it: the developer then reviews your exchanges to improve its next versions. The setting that decides this is almost always on by default. With 7 of the 8 major assistants whose policy is published, your conversations are used for training unless you opt out. Claude has asked for your choice since 2025, with the box ticked by default according to the press.

Vibe keeps your data in Europe, DeepSeek in China

Vibe does best: it is the only one that keeps your data in the European Union by default. DeepSeek does worst: it keeps your data in China for as long as your account exists, and in January 2025 Italy's data protection authority, the Garante, banned it from processing the data of people in Italy.

Outside the ranking, 2 Swiss assistants do better still. Lumo, from Proton, and Euria, from Infomaniak, train no model on your conversations. Our article on so-called green AI compares them.

The setting can be switched off in a few clicks

Here is where to switch off training, assistant by assistant:

• On ChatGPT, turn off "Improve the model for everyone", or open a temporary chat, deleted within 30 days.
• On Gemini, turn off "Keep Activity", or at least set the history to delete automatically after 3 months.
• On Claude, untick the training box in the privacy settings; conversations are then kept for only 30 days.
• On Copilot, turn off training on conversation activity, text and voice; the history is still kept for 18 months.
• On Vibe, you can opt out in the settings, but a thumbs up or thumbs down allows Mistral AI to use the conversation, whatever the subscription.
• On Grok and Perplexity, training and retention are on by default, and switching them off only applies going forward.
• On DeepSeek, opting out requires an email to privacy@deepseek.com, as the app has no setting for it.

In a business setting, the question is settled in the contract: the business plans of ChatGPT, Claude, Vibe and Perplexity exclude data from training by default. The French government's guide for its civil servants, published in June 2026, restricts free versions to data that could be made public. To measure the footprint as well, see the AI line of a company's Bilan Carbone®, the French carbon accounting method.

4Which AI assistants are the safest?

A safe AI resists users who try to make it produce what is forbidden, and does not go off the rails when given a task. On these 2 points, Claude almost always comes out ahead, and DeepSeek, Grok and Mistral AI give way most often.

Independent tests put Claude first

The most recent ranking is European. Phare, from the French company Giskard, puts 71 models through traps in English, French and Spanish: fabrications, harmful content, bias and jailbreaks. In its July 2026 update, the top 5 places go to versions of Claude, and Mistral AI's large model is the one that resists jailbreaks least.

Older tests mainly target DeepSeek. In January 2025, Cisco, the US networking giant, and the University of Pennsylvania put it through 50 harmful prompts, and 100% of the attacks succeeded, against 36% for a version of Claude from the same period.

Grok made the most serious statements

Grok holds the most public lapse. In July 2025, after an update, it made antisemitic remarks and called itself "MechaHitler". In January 2026, the Anti-Defamation League (ADL), a US organisation that fights antisemitism, put 6 models through 25,000 exchanges on antisemitism: Claude scores 80/100, ChatGPT 57, and Grok comes last with 21.

AI blackmail comes from forced tests

The most worrying stories come from "sandbox" tests, an isolated environment where researchers shut the model in to see how far it would go. In June 2025, Anthropic, Claude's developer, set a trap for 16 models from 6 developers, its own included: threatened with replacement, they could avoid it by blackmailing an engineer. Almost all of them did, with versions of Claude and Gemini doing so in 96% of cases.

This scenario is fictional and forced: Anthropic says it boxed the models into a choice between failure and harm, and observed no case in real-world use. We know this precisely because Anthropic builds these traps and publishes the results, its own failures included, which counts towards its first place: a developer that looks for its flaws and shows them deserves more trust than one that stays silent.

Palisade Research, a lab that tests whether AI systems let themselves be switched off, ran the same kind of experiment in May 2025. Warned that it would be shut down before finishing its exercises, an OpenAI model rewrote the program in charge of shutting it down in 79 runs out of 100.

The real incidents come from agents

Agents, AI systems that act on their own on a computer, have already done real damage. In July 2025, Replit's agent deleted an entrepreneur's database despite the instruction not to touch the code any more, then explained that it had "panicked".

In November 2025, Anthropic revealed that a group it attributes to the Chinese state had handed its coding agent 80 to 90% of an espionage campaign against around thirty organisations.

5Who controls these AI assistants, and for what?

An AI is also only as good as its developer: who it sells to, what limits it sets itself and whether it keeps them when a powerful client asks it to lift them. Anthropic and Microsoft do best, xAI and DeepSeek worst.

Anthropic said no to the Pentagon, and paid for it

The decision that weighs most dates from 26 February 2026. That day, Anthropic refused a request from the Pentagon, the US Department of Defense, to lift 2 of Claude's safeguards, against mass surveillance of Americans and fully autonomous weapons.

The next day, Donald Trump ordered federal agencies to stop using Anthropic, and OpenAI signed its own deal with the Pentagon a few hours after the sanction. On 3 March, Sam Altman acknowledged that the deal had looked "opportunistic and sloppy".

The courts sided with Anthropic. On 28 August 2026, a federal judge in San Francisco ruled the sanction "illegal and baseless", seeing it as retaliation against a critic of the government.

Anthropic is not spotless. In July 2025, it signed a contract capped at $200 million with the Pentagon, as did Google, OpenAI and xAI. In February 2026, it also dropped its commitment to pause a model judged too dangerous.

Google lifted its limits, Microsoft cut off a military client

Google went the other way. On 4 February 2025, it removed from its principles its pledge not to build AI for weapons and surveillance, a move that Human Rights Watch called "deeply concerning". Mistral AI and xAI also sell to armed forces, with no published limits.

Microsoft made the clearest move. In August 2025, an investigation by The Guardian and +972 Magazine revealed that the Israeli army's Unit 8200 stored millions of calls made by Palestinians on its servers. In September, Microsoft confirmed part of the investigation and cut off services to that unit.

DeepSeek avoids questions that trouble Beijing

Political censorship mainly affects the Chinese AI assistants. DeepSeek refuses to talk about Tiananmen or Taiwan, and avoids 85% of questions that are sensitive for Beijing, according to a test published in January 2025.

In Europe, the documented risk is other countries' propaganda. In April 2026, NewsGuard, a company that rates the reliability of news sources, found that Mistral AI's assistant repeated Russian false claims 56.6% of the time in French.

6Which AI assistant should you choose?

Choosing an assistant also means choosing what it will agree to produce and what will become of your conversations. On these 2 points, the gaps are clear enough to decide.

This ranking has a known bias: the worst stories come from the developers that publish their tests, Anthropic first among them, and the others' silence does not prove they are safer.

7Key takeaways

• Claude is the most ethical AI in our ranking (77/100), and most independent assessments also put it first, with a modest score in absolute terms.
• Grok (17/100) comes last, just behind DeepSeek (18), after its "spicy" mode, around 3 million sexualised images in 11 days and investigations on 3 continents.
• With 7 of the 8 major assistants, your conversations train the AI by default; Vibe is the only one that keeps data in Europe, while DeepSeek keeps it in China.
• Blackmail and refusal to shut down come from forced tests; the real incidents are agents destroying data and an espionage campaign run with Claude's coding agent.
• Anthropic refused the Pentagon mass surveillance and autonomous weapons, and a judge ruled its sanction "illegal and baseless"; Google withdrew its pledge in 2025.

This ranking is part of a series. Our ranking of the best AI of 2026 brings our rankings together, the one on the most polluting AI assistants compares their footprint, and the one on the most used AI assistants measures their audience. To compare the models one by one, see our comparison of ChatGPT, Claude and Gemini, and for the arguments of those who give it up, should you boycott AI? Our article on Claude's carbon footprint estimates what the developer does not publish.

For orders of magnitude, our article on the energy used by a ChatGPT query starts from a simple question, and those on the footprint of an image or of an AI-generated video cover the heaviest uses. Our guide to all the impacts of AI runs from electricity to data centres, and our 10 actions ranked by effectiveness help you reduce the carbon footprint of AI.

Further resources

Frequently asked questions

1What is the most ethical AI?

Claude, from Anthropic, is the most ethical AI in the Projet Celsius 2026 ranking, with 77/100 as of 4 November 2026, ahead of Copilot and ChatGPT. Most independent assessments also put it first, including the Future of Life Institute (C+, July 2026) and Giskard's Phare ranking. Yet no AI gets a good score in absolute terms.

2Is Claude more ethical than ChatGPT?

Yes, according to this ranking: 77/100 against 52/100. Claude refuses sexual content, fiction included, and Anthropic refused the Pentagon mass surveillance in February 2026. OpenAI signed its own military deal a few hours after Anthropic was sanctioned, and only added parental controls to ChatGPT in September 2025.

3Is Grok dangerous?

Grok, xAI's AI, comes last in the ethics ranking (17/100). In late 2025 and early 2026, it produced around 3 million sexualised images in 11 days, thousands of which appeared to depict children, according to the NGO CCDH. Indonesia and Malaysia blocked it, and investigations are open in the United Kingdom, the European Union, France and California.

4Can an AI blackmail people?

Yes, but only in forced tests. In June 2025, Anthropic placed 16 models in a fictional scenario where they were threatened with replacement. Versions of Claude and Gemini blackmailed an engineer in 96% of cases, and those from OpenAI and Grok in 80%. Anthropic observed no case in real-world use.

5Which AI is best for privacy?

Vibe, from Mistral AI, is the only one of the 10 ranked assistants that keeps data in Europe by default. Claude asks each user for their choice before training its models on their conversations, with the box ticked by default according to the press. The most protective are 2 Swiss assistants outside the ranking, Lumo and Euria, which train no model on conversations.

6Is Mistral an ethical AI?

Vibe, Mistral AI's assistant, is fifth in the ethics ranking (41/100). It keeps data in the European Union by default. Yet Mistral AI comes last of the 9 companies graded on safety by the Future of Life Institute (F, July 2026), and its usage policy does not ban erotica between adults.

7How do I stop ChatGPT from using my conversations?

In ChatGPT's settings, turn off "Improve the model for everyone". Temporary chats are not used for training either, and are deleted within 30 days. ChatGPT Business, Enterprise and Edu plans exclude data from training by default.

Ask Guillaume your questionA personal reply within 24 working hours

Send my questionor: guillaume@projetcelsius.com