Anthropic is open about the fact that its models sometimes make things up.
Resolved true The record shows it happened or was so.
Factual: Statements about the present or the past, checked against the record: did they hold up?
The evidence is the person's or their organisation's own records.
Said for an organisation they lead or work for; it counts on both records.
What would count as true or false
Judged as of 2026. True: Anthropic's public materials from that time (model cards, system cards, user-facing disclaimers in Claude, help-centre pages, research posts) openly say that its models can hallucinate, confabulate or produce false information. Mostly true: such statements exist but are mostly in technical documents, with only minor gaps. Mixed: the acknowledgments are present in some channels but missing or played down in prominent ones. Mostly false: acknowledgments are rare or buried, while Anthropic's main messaging presents the models as reliable. False: the evidence shows Anthropic denied or hid that its models make things up.
Set before the research began.
Evidence (4)
-
E1 Why do AI models hallucinate? · Claude Academy Anthropic (Claude Academy) · 1 January 2026 Own record
“But sometimes AIs still make things up. We call these errors hallucinations.”
-
E2 Claude está fornecendo respostas incorretas ou enganosas. O que está acontecendo? | Anthropic Help Center Anthropic Help Center · 5 March 2026 Own record
“Na tentativa de ser um assistente útil, Claude pode ocasionalmente produzir respostas que são incorretas ou enganosas.”
-
E3 Reduce hallucinations Anthropic (Claude Developer Platform docs) · 21 October 2026 Own record
“Even the most advanced language models, such as Claude, can sometimes generate text that is factually incorrect or inconsistent with the given context.”
-
E4 Model Card and Evaluations for Claude Models - Anthropic Anthropic · 8 August 2026 Own record
“Claude models still confabulate – getting facts wrong, hallucinating details, and filling in gaps in knowledge with fabrication.”
Reasoning
Anthropic's published materials from 2026 say openly, in several channels, that its models make things up. The user-facing Claude Academy tutorial says AIs "sometimes still make things up" and calls these hallucinations (E1). Two technical documents say the same: the developer docs say Claude can generate factually incorrect text (E3), and the model card says Claude models "still confabulate" and fill gaps with fabrication (E4). Because the claim is about what Anthropic itself publishes, these published materials are the direct record, and they cover both user-facing and technical channels, as the criteria for "true" require.
Cite this claim
Daniela Amodei, “Anthropic is open about the fact that its models sometimes make things up.” (said 2026). Verdict: resolved true. Reliability API, claim c_8af95caa8ce6, method 0.1.0+28da144b, dataset release v1.1.2. https://epistacklabs.com/claims/c_8af95caa8ce6
This claim is in dataset release v1.1.2, with the same verdict.
Reply to this claim
More factual claims by Daniela Amodei
-
Mostly true
“In the lab, Claude designed new protein binders from scratch.”
In laboratory work, Anthropic's Claude designed new protein binders from scratch.
-
Mostly true
“including open-source code much of the internet runs on, before attackers could exploit them.”
-
Resolved true
“They also showed that more than 6,000 women had come in for prenatal care, but only 1,762 had delivered with a healthcare worker present, a discrepancy that the team is investigating and addressing.”
-
Resolved true
“Those dashboards flagged a spike of 25,000 pneumonia cases, so the team can anticipate and plan for future surges.”
-
Resolved true
“Sometimes Claude is wrong.”
Verdicts are made automatically; how verdicts are made.