07 · Ethical Considerations in AI-Assisted Work¶
Governance (Module 3) covers what an organization requires. Ethics covers what's right even where nothing requires it — the situations where the policy is silent or ambiguous and the decision comes down to individual judgment. This module works through the recurring ethical questions that come up in AI-assisted work.
Disclosure: when to say AI was involved¶
There's no single universal rule, but a useful test is: would the recipient's assessment of the work change if they knew AI drafted or assisted it? If yes, disclosure is the honest default — for a performance review, an academic submission, a piece of journalism, or anything presented as someone's independent judgment, silence misrepresents the work. For a routine internal email or a first-draft outline you substantially reworked, disclosure is usually unnecessary because it wouldn't change anything meaningful about how the work is assessed. When genuinely unsure, disclose — the cost of over-disclosing is small; the cost of a discovered non-disclosure is trust, which is expensive to rebuild.
Attribution and originality¶
AI-assisted output that closely reproduces a specific source (Level 2, Module 3's "verify anything specific" applies here too) raises the same attribution questions as any other unattributed source use. Passing off AI-assisted analysis as entirely your own original thinking, in a context where originality is explicitly the point (a thesis, a competitive proposal, an award submission), is a form of misrepresentation regardless of whether any policy explicitly bans it.
Fairness and bias in AI-assisted decisions¶
AI output reflects patterns in its training data, which can encode biases the user doesn't intend and may not notice. This matters most where AI assistance touches decisions about people — screening resumes, drafting performance language, summarizing feedback about someone. Treat any AI-assisted output that evaluates or characterizes a person as needing the highest level of human review, not the lowest, precisely because the failure mode (a subtly unfair characterization that reads as reasonable) is exactly the kind of error fluent AI output is prone to producing and humans are prone to missing.
Impact on others' work and livelihoods¶
Introducing AI assistance into a workflow can shift who does what work and how it's valued — this is not automatically wrong, but it's an ethical question, not just an efficiency one, when it affects real people's roles. Being honest about this impact, including with the people affected, is part of the change-management responsibility from Module 2, not separate from it.
A simple test for ambiguous cases¶
When a situation isn't clearly covered by policy: would you be comfortable if the person affected knew exactly how the output was produced and reviewed? If the honest answer is "I'd rather they didn't know," that discomfort is usually pointing at the actual ethical issue, even if you can't immediately name it.
How It Actually Works¶
The bias risk described above isn't a bug that a "less biased" model release fixes outright — it's a structural consequence of how these systems are trained. A language model learns to continue text by finding statistical patterns across an enormous training corpus, and that corpus is a snapshot of existing human writing, with all its existing skew: who gets described with which adjectives, whose credentials get mentioned and whose don't, which names cluster with which outcomes in historical text. The model doesn't "believe" any of this — it has no beliefs — but it reproduces the correlations it was shown, and those correlations tend to surface most exactly where they matter most: in output that characterizes a person. This is why "the model didn't intend to be unfair" is true and irrelevant at the same time; intent isn't the mechanism, pattern-matching against biased training data is, and no amount of the model "trying harder" changes what patterns it learned.
This also explains why human review is the actual mitigation rather than a better prompt. A biased pattern in the output is invisible from inside a single generation — the model can't flag its own blind spot, because a blind spot is precisely the thing it has no signal for. A prompt that says "be fair and unbiased" changes surface phrasing (softer words, more hedges) far more reliably than it changes the underlying judgment, because the instruction is itself just more text the model continues from, not a constraint on the statistical process underneath. Catching a subtly unfair characterization requires a reader who brings context the model doesn't have and doesn't need to have to sound fluent — which is exactly why fluency is the trap: confident, well-formed sentences read as considered judgment even when the underlying signal is just "this pattern was common in the training data," and the more fluent the output, the harder people find it to apply the same scrutiny they'd apply to a rougher first draft.
The disclosure question has a similar structural root. Nothing in how these models generate text distinguishes "this is a first-draft outline I'll heavily rework" from "this is a finished, publishable assessment of a person" — the model produces both with the same fluent confidence, because fluency is a property of the generation process, not a signal about how much downstream verification the content has or needs. That asymmetry — output confidence is constant, but appropriate trust varies enormously by context — is exactly why the "would their assessment change" test in this module has to live outside the model, as a judgment call only the human in the loop can make.
Exercise¶
Think of one AI-assisted piece of work you've produced where disclosure was genuinely ambiguous. Apply the "would their assessment change" test to it in hindsight, and decide what you'd do differently, if anything, next time a similar case comes up.