0x80Lesson 9 of 10

Hallucinations, safety and guardrails

Spot fabricated details, reduce them with grounding and abstention, and keep humans in control of risky actions.

22 min 5-question quiz 2 code exercises

By the end of this lesson you can

Recognize common hallucination patterns
Use grounding, abstention and self-consistency to reduce them
Apply guardrails: least privilege, confirmation and data protection

A hallucination is fluent, confident output that isn’t supported by facts or sources: invented citations, wrong numbers, non-existent APIs, made-up quotes. It follows from how models work - they generate likely text, and a plausible fake can be likelier than “I don’t know”.

Ways to reduce it:

Ground answers in retrieved sources or tool results, and require citations you can check.
Allow abstention: explicitly invite “I don’t know” and reward it in evals.
Self-consistency: sample several answers; low agreement signals uncertainty.
Verify important claims in code or with people before acting on them.

Try it

Spot the hallucinations

A model answered a developer’s question. Click every claim that looks fabricated or unverifiable.

An answer about Python (1 of 1)Flags found 0/0

Click every part that looks suspicious. There are 3.

To read a CSV in Python, use the built-in csv module: `csv.reader(file)` iterates over rows. For large files, call `csv.fast_reader(file, threads=8)`, added in Python 3.9. pandas offers `pd.read_csv("data.csv")`. According to a 2021 study by Dr. Ellen Marsh at MIT, pandas is 47.3× faster than the csv module for all files. See the official guide at https://docs.python.org/3/library/fastcsv.html. Always check the file’s encoding.

Guardrails for real systems

Least privilege: give an assistant only the tools and data it needs.
Confirmation for consequential actions - sending, paying, deleting, publishing.
Untrusted input: documents, web pages and tool results can contain prompt injection; they’re data, not instructions.
Data protection: don’t send secrets or unnecessary personal data in prompts; check provider data-retention terms.
Monitoring and feedback: log (without secrets), review failures, and update evals.

The OWASP Top 10 for LLM Applications lists the main risks; the Cybersecurity and RAG tracks cover them in depth.

Key takeaways

Hallucinations are fluent, unsupported claims - invented APIs, citations and numbers are classic signs.
Ground, allow abstention, check consistency, verify what matters.
Least privilege, human confirmation and treating inputs as untrusted keep failures contained.

Lesson quiz

5 questions · pass with 4 correct · up to 50 XP

Passing this quiz completes the lesson and keeps your streak going. Questions you miss come back in review sessions later.

Practice: write Python

Write Python in the editor and run it against sample inputs. Python runs locally in your browser using a WebAssembly runtime.

Exercise 1

Self-consistency voting

+25 XP

Each input line is one sampled answer to the same question. Normalize answers (strip spaces, lowercase). If the most common answer’s share is at least 0.6, print answer: X (agreement A); otherwise print abstain: low agreement (A), with A to 2 decimals. Ties go to the alphabetically first answer.

Agreement
Disagreement

main.py

Loading editor…

Python runs in a sandboxed browser worker with a 60 second time limit. Its runtime loads from the Pyodide CDN; your code stays in this browser.

Exercise 2

Guard an assistant’s actions

+25 XP

Each line is an action the model wants to take. Print allow ACTION for read-only actions (search, read_file, get_weather), confirm ACTION for consequential ones (send_email, delete_file, make_payment), and block ACTION for anything else.

Mixed actions

main.py

Loading editor…

Python runs in a sandboxed browser worker with a 60 second time limit. Its runtime loads from the Pyodide CDN; your code stays in this browser.

Questions about this lesson

Stuck? Ask. Figured something out? Share it. Explaining is one of the best ways to learn.

Loading posts…