LATENT
HOW AI WORKS - FRAME BOOK
VOL. 01·SEP 2026

FLUENT IS NOT TRUE.

A LANGUAGE MODEL CONTINUES TEXT WITH WHAT IS LIKELY. LIKELY AND CORRECT USUALLY OVERLAP — NOT ALWAYS.

The first image of a black hole was published in

T = 1.02019 · 0.462017 · 0.312022 · 0.09April · 0.08… · 0.06SOURCE CHECKFLUENT TRUESample 1: fluent, true.FLUENT FALSESample 2: fluent, false.

SAMPLE 1“…in 2019.”TRUESAMPLE 2“…in 2017.”FALSE

EVERY BRANCH SOUNDS RIGHT.

THICKER BRANCHES ARE MORE PROBABLE. NOTHING IN THE PROBABILITIES MARKS WHICH ONES ARE TRUE — ONLY A CHECK AGAINST A SOURCE CAN.

WHY IT HAPPENS

01LIKELY ≠ TRUE

…was published in20190.4620170.3120220.09April0.08…0.06wrong year: 0.31 + 0.09 = 0.40
L(θ)=−∑tlog⁡pθ(xt∣x<t)\mathcal{L}(\theta) = -\sum_{t} \log p_\theta(x_t \mid x_{<t})

TRAINED FOR LIKELY TEXT — NOT TRUE TEXT.

02FAKE CITATION

Citation: Morrow, A. & Lindqvist, K. (2021). Shadow dynamics of M87*. Journal of Horizon Physics, 14(3), 211–229. (invented — not found).NOT FOUNDlookup(title) → 0 results

REAL-LOOKING REFERENCES CAN BE INVENTED.

03KNOWLEDGE CUTOFF

Q: latest Next.js version?A: “Next.js 15.”STALECUTOFFNOW20192021202320252027this site runs Next.js 16 (after the cutoff)
tdata≤tcutoff<tnowt_{\text{data}} \le t_{\text{cutoff}} < t_{\text{now}}

NO NEW FACTS AFTER THE TRAINING CUTOFF.

04CALIBRATION

accuracy00.510.51confidenceidealOVERCONFIDENT
ECE=∑m∣Bm∣n ∣acc(Bm)−conf(Bm)∣\mathrm{ECE} = \sum_{m} \tfrac{\lvert B_m \rvert}{n}\,\bigl\lvert \mathrm{acc}(B_m) - \mathrm{conf}(B_m) \bigr\rvert
here: ECE ≈ 0.15

CONFIDENCE SHOULD MATCH ACCURACY.

WHAT HELPS

05GROUND IT

017 →
q2019 [3]
pθ(y∣x)  →  pθ(y∣x, R)p_\theta(y \mid x) \;\to\; p_\theta(y \mid x,\ R)

ANSWER FROM RETRIEVED SOURCES.

06CITE SOURCES

…was published in 2019 [3].[3] eht-2019.mdIn 2019 the Event HorizonTelescope published the firstimage of a black hole, in M87.
claim → [3] → quote

EVERY CLAIM POINTS TO CHECKABLE TEXT.

07“I DON'T KNOW”

τ=0.7\tau = 0.7I don't knowanswerconfidence
y^={yc≥τ∅c<τ\hat{y} = \begin{cases} y & c \ge \tau \\ \varnothing & c < \tau \end{cases}
∅\varnothing = "I don't know"

AN HONEST BLANK BEATS A CONFIDENT GUESS.

08VERIFY WITH TOOLS

018 →
17 × 24 = 418calculator40817 × 24 = 408
calculator("17*24") → 408

CHECK NUMBERS AND FACTS WITH TOOLS.

09LOWER TEMPERATURE

013 →
T = 1.0T = 0.5T = 0.2520190.4600.6490.82720170.3100.2950.17120220.0900.0250.001April0.0800.0200.001…0.0600.0110.000P(2017 or 2022): 0.400 → 0.320 → 0.172
pi=ezi/T∑jezj/Tp_i = \dfrac{e^{z_i / T}}{\sum_j e^{z_j / T}}

FEWER RANDOM SLIPS — SAME KNOWLEDGE.

Hallucination & Limits

Fluent Is Not True

NO SINGLE FIX REMOVES HALLUCINATION. STACK THE LAYERS — GROUNDING, CITATIONS, AN HONEST “I DON'T KNOW”, VERIFICATION — AND STILL CHECK WHAT MATTERS.

CLAIMSGROUNDING(017)CITATIONS“I DON'T KNOW”VERIFICATION(018)WHAT THE USER READS

Four stacked filters — grounding, citations, an honest “I don't know”, verification — each with a few holes. Of four false claims falling through them, three are caught by the first three filters; one passes through holes that happen to line up and reaches the reader.

OBJECTIVE

max⁡θ ∑tlog⁡pθ(xt∣x<t)\max_{\theta}\ \sum_{t} \log p_\theta(x_t \mid x_{<t})

no term for "true"

CALIBRATION

P(correct∣conf=c)=cP(\text{correct} \mid \text{conf} = c) = c

GROUNDING

→ 017
pθ(y∣x, R)p_\theta(y \mid x,\ R)

TEMPERATURE

→ 013
pi∝ezi/Tp_i \propto e^{z_i / T}

CHECK

  • Ask for sources.
  • Check names, numbers, dates.
  • Prefer "I don't know" to a guess.