The Confident MachinePre-print

An interactive AI explainable

Why language models make things up, and what to do about it.

Large language models write with total confidence, whether they are right or inventing facts from thin air. In 14 short, visual pages, this explorable builds up an intuition for what hallucinations are, why they are a structural consequence of how these models work, and how researchers, engineers, and everyday users can navigate them.

Start reading →

Contents

Part I: Foundations

What language models actually do, and why that matters for truth.

  1. 1The Confident MachineA fluent, authoritative, entirely fabricated answer, and why it matters. What this explorable covers and how to read it.
  2. 2One Token at a TimeLLMs are next-token predictors. Watch a model build a sentence probability-by-probability.
  3. 3Rolling the DiceTemperature, top-p, and how sampling turns a probability distribution into words.
  4. 4Plausible ≠ TrueThe core insight: the training objective rewards plausibility, not truth. Fluency and factuality come apart.

Part II: Anatomy of a Hallucination

What hallucinations are, what they look like, and how to classify them.

  1. 5Defining HallucinationWhich failures count as hallucinations, a working taxonomy sorted from real examples, and why 'lying' is the wrong mental model.

Part III: Why Models Hallucinate

The statistical, architectural, and incentive-driven roots of confabulation.

  1. 6The Limits of Model MemorySingletons, the monofact rate, and why compressed parameters blur rare facts into plausible neighbors instead of storing them.
  2. 7Why the Model Answers AnywayThree reinforcing incentives for confident wrongness: scoring rules that reward guessing over abstaining, sycophancy and leading questions that bend answers toward the user, and RLHF training that rewards sounding sure over being honest.
  3. 8The Snowball EffectOne wrong token becomes a paragraph of fiction: exposure bias and error cascades in autoregressive generation.

Part IV: Measuring & Detecting

Benchmarks, uncertainty, and catching hallucinations in the act.

  1. 9How Do We Know It's Wrong?Benchmarking against known answers (TruthfulQA, SimpleQA, HaluEval, FActScore), what aggregate scores hide, and whether a model's own token probabilities can warn you first.
  2. 10Ask It Five TimesSampling-based detection: if answers disagree with each other, trust none of them.

Part V: Mitigation & Engineering

Grounding, retrieval, abstention, and building systems that fail safely.

  1. 11Grounding with EvidenceDetection flags a claim; retrieval supplies evidence; generation reads it; citations make the connection checkable, and all four steps can still fail.
  2. 12Defense in DepthHow real products layer retrieval, verification, uncertainty, and UX to ship trustworthy systems.

Part VI: Living With Confident Machines

Practical judgment for users, builders, and society.

  1. 13Use, Verify, or Walk AwaySeven habits that work from inside an ordinary chat window, a two-question heuristic for how much to trust an answer, and when the combination means it's time to close the tab.
  2. 14Will Hallucinations Ever Go Away?The frontier: what's improving, what's fundamental, why an honest machine still isn't a solved machine, and the whole argument, in three sentences.

How to read it

Pages stand alone but build in order. Every figure is live: sliders re-derive the curves, diagrams answer clicks, and the demos sample from a real model when one is available; canned transcripts otherwise.

It is written for undergraduate students and other learners meeting these ideas for the first time; no prior machine-learning background is assumed.

Authors

Shashank Bangalore Lakshman*†

Graduate Student, M.S. in Artificial Intelligence

Department of Computer Science, The University of Texas at Austin

Shravan Joshi†

Undergraduate Student, B.S. in Cognitive Science

Department of Letters and Sciences, University of California Los Angeles

* Corresponding author.† These authors contributed equally.

Cite this work

@misc{sblsj2026theconfidentmachine,
  title  = {The Confident Machine},
  author = {Bangalore Lakshman, Shashank and Joshi, Shravan},
  year   = {2026},
  note   = {Preprint. Submitted to the VISxAI Workshop at IEEE VIS; submitted for peer-review on 10th Aug 2026.},
  url    = {https://the-confident-machine.vercel.app/},
}