An ethical failure in a prompt usually arrives as a cost failure first. A fabricated citation gets caught in review, the work is redone, and the retry ships with a longer prompt and less trust. A leaked customer record is worse: a fraction of a cent becomes a notification, a legal review, and a lost account.
Hallucination is a verification cost
Models generate text that fits the shape of an answer. When the pattern is strong and the facts are thin, the output reads exactly like a correct one. Confidence is a property of the prose, not of the evidence. Ask for a source and a model will often supply a real-looking title, a plausible author, and a URL that resolves to nothing.
The cheap defence is structural: require the evidence, allow an explicit unknown, and route high-stakes claims to a person. The expensive defence is finding the error after publication.
Task: answer the question using only the supplied sources.
Question: {{question}}
Sources: {{sources}}
Rules:
- Cite the source line for every factual claim.
- If the sources do not answer it, write INSUFFICIENT EVIDENCE and name what is missing.
- Separate what the sources say from what you infer. Label inference.
- Do not invent titles, dates, names, or URLs.
Output: 5 bullets, then one line on confidence.An explicit escape hatch is cheaper than a confident guess.
Bias enters through the prompt
Bias is usually framed as a training problem. In practice the prompt decides which criteria are allowed to matter. Ask it to rank on “fit” and it fills the gap with patterns from training data. Proxies for age, gender, nationality, and class walk in through a word nobody defined.
- Name the criteria. Replace “best fit” with five job requirements and evidence for each.
- Remove identifiers from screening prompts: names, schools, addresses, photos, and dates.
- Test matched pairs. Change one attribute, hold everything else equal, and compare the rankings.
- Ask what is missing. A model will rank the candidates it was given, even when the strongest applicant is not in the pile.
Privacy and the three exits for data
Every prompt crosses three boundaries: the provider, your logs, and the model’s output. A customer email pasted into a chat has left your database. It may be retained for abuse monitoring, sit in a support transcript, or return inside a summary that gets pasted somewhere else.
- Provider side — check retention, training opt-out, and region before any personal data leaves your systems.
- Your logs — prompts and responses are often stored in full by default, in tools with wider access than the source system.
- Output — a model can repeat a name, key, or record it saw earlier in the conversation.
- Secrets — API keys and credentials are compromised the moment they are sent.
Task: rewrite the record below for analysis with no personal data. Remove: names, emails, phone numbers, addresses, account numbers, dates of birth, employer names. Replace each with a stable token: PERSON_1, EMAIL_1, ORG_1. Keep: role, country, tenure band, and the issue text. Flag anything you cannot classify instead of guessing. Output: the redacted record, then the token list you used.
Redaction is a preprocessing step, not a promise the model makes.
The pitfalls that waste tokens
Ethical shortcuts are expensive shortcuts. The waste lands in retries, reviews, and incidents, not on the invoice.
- 1Over-trusting the first answer. Skipping the check moves cost from tokens to rework. One wrong number in a board report can cost more than a month of prompts.
- 2No verification step. Fluent guesses get fact-checked line by line. Put sources, unknowns, and a confidence line in the prompt.
- 3Leaking PII to save time. Pasting a raw customer file is faster for one request and far more expensive after an incident.
- 4Ethics boilerplate. “Be fair, accurate, and unbiased” spends tokens and changes nothing. Replace it with a criterion, a failure example, and an output check.
- 5Re-pasting policy on every call. Stable rules belong in a cached system prompt, not in every user message.
What a responsible prompt contains
Responsible prompting is unglamorous: name the sources, state what is unknown, exclude what is not needed, and leave a trail a reviewer can follow.
Task: {{task}}
Audience: {{audience}}
Allowed sources: {{sources}}
Excluded data: no names, credentials, or customer identifiers.
Evidence: cite a source for each claim; label inference as INFERENCE.
Uncertainty: if evidence is missing, say so and stop.
Review: {{reviewer role}} checks every factual claim before use.
Output: {{format}} in {{limit}} words.Four lines of constraints cost less than one retraction.
The economics are simple once you price the downside. A verified answer costs a few more input tokens. An unverified one costs a reviewer, and sometimes a lawyer. The prompt is where that choice gets made.
Frequently asked questions
Can a prompt prevent hallucination entirely?
No. A prompt can require sources, permit an explicit unknown, and separate inference from evidence, which makes fabrication easier to catch. It cannot guarantee that every claim is true. High-stakes outputs still need a human check against the source.
Is removing names enough to make a prompt private?
Not usually. Combinations of role, location, tenure, and a rare condition can re-identify a person. Redact direct identifiers, coarsen quasi-identifiers, and check your provider’s retention and training terms before sending anything sensitive.
Does ethics guidance in a prompt improve model behavior?
Vague guidance such as “be unbiased” rarely changes output. Specific rules do: named criteria, matched-pair tests, required evidence, and an instruction to stop when information is missing. Those rules cost a few tokens and are auditable.
What is the most expensive prompting mistake?
Shipping an unverified answer into a decision someone acts on. The token cost of a verification pass is small. The cost of a wrong figure, a biased shortlist, or a leaked record is measured in rework, legal exposure, and trust.
All prompts and examples in this article are original. Cost figures are computed at build time from the TokenCalc AI catalog and will move when prices change.