Fabrication and why it sounds certain

A chatbot states a wrong answer in the same steady voice it uses for a right one, and that voice is not a reliable signal of which one you are reading.

Where the wrong answer comes from is not a malfunction. A 2025 OpenAI paper argues fabrication begins as an ordinary classification error during pretraining: when false statements cannot be reliably told apart from true ones in the training text, some false ones get produced by the same statistical pressure that produces the true ones.¹ The paper then argues the habit survives training because the scoreboards used to rank models award credit for a confident guess and nothing for admitting uncertainty, and its proposed remedy is to change how existing benchmarks are scored rather than to add a new test. That matters for the most repeated claim about these systems. Fabrication as a permanent, unfixable property is contradicted by a vendor's own account of its causes.

The certainty is measured too. Anthropic inspected more than a million training transcripts of Claude Opus 5 by recursive summarisation and found many cases where the model "confidently stated an answer it was unsure about".² The same card's inspection of internal activations lists fabrication awareness as a recurring category: while producing invented material the model represents it internally as invented, while the visible answer does not. Calibration, meaning the match between the confidence a model expresses and how often that confidence turns out right, was measured on GPT-4 in 2023. The pretrained model was highly calibrated on a held-out slice of a knowledge test, and calibration was reduced after post-training.³

In practice the published figures are mostly comparative. Anthropic reports Opus 5 as 11 percent more accurate than Opus 4.8 with a hallucination rate 6 percent higher, and says absolute values across its audit dimensions reflect which scenarios were sampled, so they serve for comparing models.² The GPT-5 system card of 13 August 2025 does publish absolute SimpleQA hallucination rates, for four models at once; gpt-5-thinking's is 0.40, defined there as the share of factual claims containing minor or major errors.⁴ Detached from that definition and that no-browsing condition, the number means nothing.

The boundary is the calibration result. It is one model in 2023. A separate 2023 study across Pythia, LLaMA and LLaMA-2 also finds a post-training calibration loss, but attributes it to instruction tuning where the GPT-4 report attributes it to post-training as a whole.⁵ Whether the drop holds for the models you use today is not established, so treat a confident tone as a writing style rather than a probability.