Bias, Tone and Audience Fit

An output can be accurate, complete and still wrong for its purpose: too breezy for a bereavement message, too technical for a customer, or quietly assuming that every manager is a man. This lesson covers three related evaluation questions. Does the output carry assumptions you did not intend? Is the tone right for the reader? Is the framing fair?

Where bias comes from

Claude is trained on a vast body of human writing, and that writing carries the patterns of the people who produced it. Anthropic invests heavily in reducing harmful bias, and modern Claude models are careful about it, but subtle effects persist, and your own prompt can introduce bias too. Common sources:

• Defaults in examples. Ask for "a sample customer persona" and you may get a default nationality, age and gender unless you specify a range.
• Leading prompts. "Why is remote work less productive?" presupposes the conclusion; Claude will usually note the premise but may still lean into it.
• Imbalanced source material. If the three articles you upload all argue one side, a grounded summary will too, and correctly so; the bias is in the inputs.
• Cultural assumptions. Date formats, holidays, legal frameworks and idioms tend to default to a particular country unless told otherwise.
• Confirmation. Claude tends to be agreeable. If you present a plan positively, the critique you get back may be gentler than the plan deserves.

Checking for bias

Quick tests that take seconds:

• Swap a variable. If the output describes a candidate, re-run with a different name or gender and compare the adjectives.
• Ask for the counter-case. "What would someone who disagrees say?" exposes one-sided framing.
• Look at who is centred. Who is the implied reader, decision-maker or default person in the text?
• Check for loaded words. "Admitted", "claimed", "finally" and similar verbs smuggle in judgement.
• Audit your own prompt for a presupposed answer.

When you find a problem, fix it at the source: neutral prompt, balanced inputs, explicit instruction to present multiple perspectives.

Tone and register

Register is the level of formality and the choice of vocabulary. Claude can produce almost any register, but it has to be told which. Left unspecified it tends towards a helpful, moderately formal, slightly enthusiastic voice, which is wrong for many business contexts (legal notices, condolences, disciplinary letters, terse operational updates).

The most reliable ways to set register are, in order of effectiveness:

1. Show an example of the voice you want.
2. Name a reference ("the plain, unhurried tone of our incident post-mortems").
3. Describe it with concrete rules ("no exclamation marks, no rhetorical questions, short sentences, no adjectives of enthusiasm").
4. Use a Style (Domain 5) so the register is applied automatically.

Vague adjectives ("professional", "engaging", "friendly") are the least effective because they mean different things to different people.

Audience fit

Audience fit is about depth and assumed knowledge as well as tone. The same facts need different treatment for a board, a customer, a regulator and a new starter. When evaluating, ask:

• Would this reader understand every term without looking it up?
• Does it answer the question this reader actually has?
• Is the level of detail right: enough to act on, not so much that the point is buried?
• Does it respect the reader's time and status? A one-line reply to a detailed customer complaint reads as dismissive, however accurate.

Before and after

Write a message to staff about the redundancies.

This will produce something that is probably too warm in places, too corporate in others, and generically structured. Compare:

Write a message to all staff announcing that 12 roles in the Bristol
office will be made redundant following the site review. Reader: all
employees, many of whom know the people affected. Tone: sober, direct,
respectful; no euphemisms ("right-sizing", "letting go"), no upbeat
framing about the future in the first half. Say what support is
available and when more detail will follow. Here is the tone we used
for the site review announcement: [paste]. Under 250 words.

The constraints on euphemism and framing are doing the ethical work here; they stop Claude from softening a message that should not be softened.

Scenario: the job advert

A recruiter asks Claude to draft an advert and gets copy describing the ideal candidate as a "young, energetic digital native who will hit the ground running". Nothing is factually wrong, but the age-coded language is both exclusionary and, in many jurisdictions, a legal risk. The recruiter asks Claude to "review this advert for language that could discourage or exclude candidates by age, gender, disability or background, and suggest neutral alternatives". Claude flags each phrase and offers replacements. The lesson: Claude is a capable bias reviewer when asked, but it will not necessarily volunteer the review.

Exam focus

• Bias in output can originate from training patterns, from leading prompts, or from imbalanced source material; check all three.
• Quick bias tests: swap a variable, ask for the counter-case, look at who is centred, audit your own premise.
• Register is set most reliably by example, then by named reference, then by concrete rules; vague adjectives are least reliable.
• Audience fit covers depth and assumed knowledge, not just tone; the same facts need different treatment for different readers.
• Claude can act as a reviewer for exclusionary or loaded language when explicitly asked.