Talking to AI, and Catching Its Lies

Generative AI: Foundations and Applications · TCE

The prompt template (steal this)

# ROLE        You are a <specific persona with a stake in quality>.
# TASK        <one job, precisely: verb + object + length>
# CONTEXT     <the facts it needs — it remembers nothing>
# FORMAT      Reply ONLY with <JSON schema / bullet cap / word cap>.
# EXAMPLES    <2–5 input→output pairs; they pin format + edge cases>
# CONSTRAINTS If a fact isn't stated above, don't invent it.
              If unsure, say "I am not sure."

Task is mandatory; the rest are dials. One prompt = one job (chain calls for multi-step work).

The three power moves

1. Few-shot: show 2–5 examples instead of describing the format. The model imitates patterns — in-context only, no learning.
2. Step-by-step: "Solve step by step, then give the final answer." Visible, auditable working — errors surface.
3. Grounding: "Answer only from the facts above." The single line that suppresses invented awards, policies, citations.

Hallucination in one line

Trained for plausible, not true — so lies arrive fluent, confident, and welded to real facts. Tone tells you nothing. Only checking detects it. (Real stakes: a US lawyer was sanctioned over 6 invented citations; Air Canada was held liable for a policy its chatbot made up.)

The eval harness (the artifact that makes you an engineer)

def norm(s): return re.sub(r"[^a-z0-9 ]", "", s.lower())

def run_eval(template, tests):
    hits = 0
    for t in tests:
        ans = ask(template.format(q=t["q"]), temperature=0.0)
        hits += norm(t["expected"]) in norm(ans)
    return hits / len(tests)

Rules: expected = the key fact only · temperature 0 · run 3×, report average + spread · read every ✗ (model wrong? scorer too strict? question ambiguous?) · change one thing per iteration · a test the model always passes teaches nothing.

Scorer trade-offs

Scorer	Good	Trap
Exact match	unambiguous	fails correct answers over punctuation
Normalized contains	robust, simple	generic expected strings false-positive
LLM-as-judge	handles paraphrase	verbosity/position/self bias — audit it

The loop (memorize)

Write → eval → read failures → fix ONE thing → re-run. "It worked when I tried it" is not evidence — the demo is the best case, the eval is the expected case.

Carry forward: this harness grades your capstone (S6). Next session: photos on your phone.