ChatGPT, Claude, Gemini, Mistral, Perplexity:
The fact-based comparison

How far back does the training data go?

AI models are trained on data up to a specific date — the so-called Knowledge Cutoff DateKnowledge CutoffThe date up to which an AI model was trained. Events or information after this date are unknown to the model from its own knowledge — it then needs a web search to provide current information.. Events or information after this date are not part of the model's training. This means: for current topics, a web search is often necessary, even if the AI has access to one.

Knowledge cutoff dates for all 5 assistants

What does this mean in practice?

• ChatGPT (GPT-5.2): Knows events up to end of August 2025. Everything after requires a web search.
• Claude (4.6 Sonnet): Distinguishes between "reliable knowledge" (Aug 2025) and "training data cutoff" (Jan 2026). Most reliable knowledge up to August 2025.
• Gemini (3.0): Oldest cutoff (January 2025) — a web search is needed for anything from February 2025 onwards.
• Mistral: July 2024 — comparatively outdated for current events.
• Perplexity: Uses various models (GPT, Claude, Mistral) and combines them with web search — therefore the most current.

Who are they — and what do they offer?

How providers use your chat history

Most AI assistants use chat histories to improve their models — if you do not object (opt-outOpt-out vs. Opt-inOpt-out = you must actively object, otherwise your data is used. Opt-in = your data is only used if you actively consent. Opt-in is more privacy-friendly.). The details vary considerably:

The details compared

ChatGPT (OpenAI):

• Default: conversations are used for training.
• Opt-out: Settings → Data Controls → turn off "Improve the model for everyone".
• After opt-out: 30 days retention for safety/abuse detection.
• Temporary Chat: stores nothing, but must be activated manually each time.
• Enterprise/APIEnterprise & APIBusiness customer access and programming interfaces. Stricter privacy rules apply here: no training on user data, contractually guaranteed. Costs more, but your data stays private.: automatically no training, contractually guaranteed.

Claude (Anthropic):

• Default: NO training without explicit consent (opt-in).
• Training only if: you actively give feedback (thumbs up/down) or join the Development Partner Programme.
• Retention: 30 days for safety (the "Unbreakable Safety LoopSafety LoopAnthropic's safety mechanism: all chats are retained for 30 days to detect misuse — even if you opt out of training. After that they are deleted."), even with opt-out.
• If opted in: 5 years retention for training purposes.
• Policy change Sept 2025: Anthropic introduced an opt-in/opt-out hybrid. Users who did not respond to the policy update were treated as having consented — a move that attracted controversy.

Gemini (Google):

• Default: conversations are stored for 18 months and used for training.
• Opt-out: complicated — via "Gemini Apps ActivityActivity ControlsGoogle's settings for data collection. Controls which data Google stores and uses for training. Difficult to find and comes with feature restrictions when turned off." in Google Account Settings. Can be adjusted to 3/36 months or turned off entirely.
• Problem: if Activity Control is turned off, some features (Gmail, Docs integration) become restricted.
• Human review: conversations can be reviewed by Google employees.
• Google ecosystem: data is linked with search history, YouTube and Gmail.

Mistral:

• Default: training ON for the consumer versionConsumer versionThe free or low-cost version for private users (e.g. via the website or app). Unlike enterprise/API access, chat data is often used for training here..
• Opt-out: straightforward via Settings → disable training.
• Enterprise/API: Zero Data RetentionZDRZero data retention: the provider stores no chat histories and uses no data for training. Standard for enterprise and API access, but not for free versions. (ZDR) — no use for training, contractually guaranteed.
• EU servers: available for GDPR compliance.
• Open source: models can be hosted locally (complete data control).

Perplexity:

• Default: training ON, including for Pro users.
• Opt-out: Settings → AI Data Usage → turn off.
• Incognito modeIncognitoA mode in Perplexity where no chat history is stored and no data is used for training. Must be activated manually for each new chat — not a permanent default.: stores nothing, but must be activated manually.
• Problem: Perplexity uses APIs from OpenAI/Anthropic (Claude) — data passes through their systems.
• API: no training for API users.

Privacy ranking

1. Claude: No training without opt-in, transparent 30-day retention.
2. Mistral: EU-based, GDPR-compliant, simple opt-out, open-source option.
3. Perplexity: Simple opt-out, incognito mode, but data passes through OpenAI/Anthropic APIs.
4. ChatGPT: Opt-out possible, but training is on by default. Temporary Chat available.
5. Gemini: Complicated opt-out, 18-month retention, Google ecosystem linkage.

Hallucinations: AI confidently makes things up

What are hallucinations? AI models "hallucinateHallucinationWhen an AI confidently invents false information — such as non-existent studies or court rulings. The model does not know it is lying, because it has no concept of truth." when they confidently present false information — invented court rulings, non-existent studies, fabricated facts. The problem: they sound completely convincing while doing so.

Case 1: Mata v. Avianca (May 2023)

New York lawyer Steven Schwartz filed a brief in which 6 of the 9 cited court rulings were fabricated — all generated by ChatGPT. When the judge asked him to produce the rulings, Schwartz asked ChatGPT: "Are these cases real?" ChatGPT: "Yes, these are real cases."

Schwartz submitted the "full texts" — all invented by ChatGPT, including fabricated quotes and internal references. The judge called it "unprecedented" and imposed a $5,000 fine.

979 documented cases worldwide

French researcher Damien Charlotin maintains a database of lawyers worldwide who have been caught out by AI hallucinations. From 2 cases per week (spring 2025) to 2–3 cases per day (end of 2025). The numbers are rising exponentially.

Stanford study: 17–33% hallucination rate

Stanford University RegLab (May 2024) tested specialised legal AI tools (Lexis+ AI, Westlaw AI) — tools marketed as "hallucination-free". The result:

• Lexis+ AI: 17% hallucination rate
• Westlaw AI: 33% hallucination rate
• Even in tools explicitly built for legal work with access to real legal databases.

Why do AI models hallucinate?

AI language models work probabilisticallyProbabilisticProbability-based: the model calculates which next word is statistically most likely — based on patterns in billions of pages of text. It does not understand what it says. — they calculate which word is most likely to come next, based on patterns in training dataTraining dataBillions of texts from the internet, books, Wikipedia and other sources on which an AI model is trained. The quality and scope of this data determines what the AI can do and knows.. They have no understanding of "true" or "false". When you ask a difficult question for which there are no good answers in the training data, the AI invents a plausible-sounding answer.

What does this mean in practice?

• Never trust blindly. Always verify sources, especially for facts, studies, statistics and quotes.
• Perplexity is better for research: always provides sources that can be verified.
• Claude & ChatGPT: with web search enabled they often cite sources — but verify those too.
• Creative writing = fine. Facts = CAUTION.

Pentagon vs. Anthropic: ethics or pragmatism?

On 27 February 2026, a conflict between the US Department of Defense and Anthropic (makers of Claude) escalated. The Pentagon gave Anthropic a deadline of 5:01 pm: remove the restrictions on mass surveillance and autonomous weapons, or lose the $200 million contract.

Anthropic CEO Dario Amodei refused: "We cannot in good conscience agree." The company insisted that Claude must not be used for mass surveillance of US citizens or fully autonomous weapons systemsAutonomous weaponsWeapons systems that can independently select and engage targets without human decision-making. Highly controversial internationally, as they remove human control over decisions to kill..

President Trump ordered: "All federal agencies must immediately stop using Anthropic." Defence Secretary Pete Hegseth declared Anthropic a "Supply Chain RiskSupply Chain RiskAn official US classification normally reserved for hostile foreign companies (e.g. Huawei). Applying it to a US company is unprecedented." — a designation normally reserved for foreign adversaries.

Hours later: OpenAI CEO Sam Altman announced that his company had signed a deal with the Pentagon — reportedly with the same "red lines" (no mass surveillance, no autonomous weapons).

What is the difference?

Unclear. OpenAI claims the restrictions are identical to Anthropic's demands. Anthropic says: the Pentagon reserved the right to override the rules "when necessary". Anthropic wanted binding guarantees. OpenAI accepted the Pentagon's promise to adhere to "existing laws and policies".

Context: OpenAI & the military

In January 2024, OpenAI removed the ban on military use from its terms of service. Since then, OpenAI has been actively working with the Pentagon. Yesterday (27 Feb 2026) that arrangement was formalised.

What does this mean?

This development reveals different priorities:

• Anthropic: willing to forfeit $200 million and the US government market in order to uphold its ethical principles.
• OpenAI: accepts a Pentagon deal with "the same" restrictions, but without binding guarantees.

For privacy-conscious users the question is: do I trust a company that puts money above principles — or one that sacrifices money for principles?