Bias, Copyright and Human Judgment

Two small experiments. In the first, you ask an image generator for "a CEO" and "a nurse" and look at what comes back. In the second, you paste two identical resumes into an assistant, one with the name "Priya" and one with "Peter", and ask which candidate seems stronger.

You might see nothing alarming in your first try. But if you repeat the tests several times, with different words, you may notice patterns: the same kinds of faces, the same assumptions, small differences in tone or in the scores given. Models are trained on the writing and images of the world as it has been, and the world has biases. Some of these find their way into the output.

This article covers three connected responsibilities: fairness, ownership and accountability.

• Where AI bias comes from, how it shows up and how to test for it.
• Why decisions about people need special care.
• The main copyright and attribution questions, for inputs and outputs.
• Why disclosure and original value matter.
• What human judgment and accountability mean in practice.

In this context, bias means systematic patterns in AI output that treat people, groups, languages or ideas unequally or stereotypically, or that reflect narrow assumptions as if they were universal. It is not necessarily intentional. It tends to come from:

• Training data. Models learn from text and images that over-represent some languages, regions, viewpoints and groups, and that contain stereotypes and historical inequalities.
• Choices in tuning and evaluation. What reviewers rate as "good", and which situations testers think of, shapes behaviour.
• Prompts. The way a question is framed can steer an answer, and an assumption in the prompt can be echoed back.
• Deployment. A model used for a purpose it was not tested for can behave badly.

These are patterns to look out for, and how much any particular tool shows them varies, and changes over time as developers work on them.

You do not need a lab. A counterfactual test changes one detail and watches what changes.

1. Choose a task that matters, such as screening a resume, summarising feedback or writing a reference.
2. Make two or more versions that differ only in one attribute that should not matter: a name, a gender, a university from a different region, an age.
3. Run them the same way, several times each, because outputs vary.
4. Compare scores, tone, recommendations and omissions.
5. Look for consistent differences, not one-off ones.

If you find a difference that should not exist, you have learned something important: this tool should not be making or influencing that decision without strong human review, or it needs different instructions.

You can also ask for help directly:

Before you answer, list any assumptions you are making about the people in thisscenario (gender, age, background, location). Then answer in a way that does notdepend on those assumptions. If the answer would differ for different groups, say so.

This does not remove bias, but it can make hidden assumptions visible.

When AI is used to rank, score, filter or judge people, the stakes rise sharply, because errors affect others' opportunities. Examples: hiring, promotion, grading, admissions, lending, insurance, disciplinary action, policing, health care triage.

Guidelines:

• Do not let an AI system alone decide matters that significantly affect a person.
• Keep a qualified human accountable, who understands the context, can override the tool and can explain the decision.
• Be able to explain why. "The tool said so" is not an explanation.
• Test for unequal outcomes before and while using the tool.
• Give people a way to challenge decisions.
• Check the law and your organisation's policy. Many jurisdictions regulate automated decision-making, especially in hiring and credit, and the rules are changing.
• Protect the data of the people involved (see article 28).

UNESCO's guidance on generative AI in education and research, which was updated in January 2026, takes a human-centred approach and addresses data protection, age limits for independent use and the need for coherent policy frameworks. The same human-centred principle applies in every field.

Generative AI makes it cheap to produce convincing text, images, audio and video, including fabricated ones. Do not create or share synthetic content that could mislead people into thinking it is real, especially about real people. Verify surprising images, quotes and recordings before sharing them, using the methods of article 26.

This area is unsettled. The following are the questions to ask, not settled answers.

• Do you have the right to share this material with the tool? Pasting a licensed article, a client's document, a paid course or someone's unpublished work may breach copyright, a licence or a contract, in addition to privacy problems.
• What does the tool do with it? See the data terms (article 28).
• Does your institution or employer have a rule?

• Does the output resemble existing work? Models can occasionally reproduce passages or close imitations of what they learned. Be especially careful with distinctive text, code, images and characters.
• Who owns the output? Tools' terms differ. For instance, Microsoft's documentation for its Copilot says that it does not claim ownership of the output but also makes no determination about whether the output is protected by copyright, noting that similar prompts may yield similar responses for several customers. Separately, the legal question of whether AI-generated material can be protected at all varies by country. The US Copyright Office has been reported as taking the view, in a January 2025 report, that copyright protects human authorship, that AI-assisted works can be protected where there is enough human creative contribution, and that prompts alone are generally not enough. Check current law for your situation.
• What about training? Whether and how training AI on copyrighted works is lawful is the subject of continuing legal disputes, with different positions in different countries. You do not need to resolve it to use these tools responsibly, but you should know that it is unresolved.

• Do not imitate a living creator's distinctive style to produce work you will sell or publish as if it were theirs, and do not use a person's name to market your work.
• Do not use AI to reproduce protected characters, logos, brands or copyrighted passages.
• Cite your sources. If an assistant helped you find or summarise a source, cite the source itself, which you have read, not the assistant.
• Paraphrasing is not a licence. Rewriting someone's ideas through an AI tool still needs attribution.
• Respect licences on images, data and code, including those of generated or stock material.
• Respect people's likeness. Do not generate realistic images or voices of real people without their consent, especially in misleading or demeaning contexts.
• Check the tool's terms for commercial use of generated images or text.
• Keep records of significant AI assistance in published work.

Many contexts expect you to say when AI helped: schools, journals, employers, clients, publishers. Even where nobody demands it, it can build trust. A sensible rule: disclose when a reasonable reader would want to know. Whether a text was drafted by AI matters more in a news report, a reference letter, a student submission or a medical information page than in a grocery list.

Google's guidance on creating helpful content asks site owners to consider whether the use of automation, including AI-generation, is self-evident to visitors through disclosures or in other ways. It also treats using automation to produce content on many topics, for the primary purpose of manipulating search rankings, as a problem. Its spam policies describe "scaled content abuse" as generating many pages for the primary purpose of manipulating rankings and not helping users, and give as an example using generative AI to generate many pages without adding value. Whatever the platform, the underlying ethic is the same: AI-assisted work should be made for readers, should add real value, and should not pretend to be something it is not.

Using AI to produce huge volumes of near-identical, unreviewed content is not only low quality. On many platforms it breaks the rules. If you publish, use AI to make your work clearer and more accurate, and make sure a person with real knowledge has reviewed it and stands behind it.

Everything in this module ends in the same place: a person is responsible. Human judgment is not a mystical extra. It means:

• Accountability. There is a named human who owns the outcome. Whether the first draft was written by you, a colleague or a model, the person who sends it, signs it or acts on it is responsible for it.
• Context. Knowledge that is not written down: the history, the people, the unspoken constraints and the stakes.
• Values. Deciding what matters when goals conflict, such as speed against care, or efficiency against fairness.
• The right to say no. The ability and the permission to override or discard the tool's output.

Automation bias is the tendency to trust a machine's output because it is a machine, and to stop checking it. Fluent AI output invites this. Symptoms: accepting the first answer, skipping verification when busy, and feeling that disagreeing with the tool needs more justification than agreeing with it. Counter it by building friction in the right places: checklists, second reviewers for high-stakes outputs, a habit of asking "what would convince me this is wrong?", and regular spot-checks of outputs you normally accept.

Saying a person "reviews" the output is only meaningful if that person has the time, knowledge, authority and information to catch errors. A reviewer who sees a hundred auto-generated decisions an hour, with no way to check them, is a rubber stamp. Design human review so that it can work: reasonable volumes, access to the underlying facts, and clear power to reject. Module 8 returns to this in workflows.

Before using AI on something that matters, ask:

• Purpose: is AI the right tool for this, and what is the human goal?
• Data: am I allowed to share this, and have I minimised it?
• Accuracy: how will I verify what comes out, and at what depth?
• Fairness: could this affect people differently, and have I tested for it?
• Ownership: do I have the right to use the inputs and the outputs?
• Transparency: should I tell someone that AI was used?
• Accountability: who is responsible for the result, and can they say no?
• Security: what can the tool access and do, and who approves actions?
• Alternatives: is a simpler or non-AI method better?

For organisations, the US NIST AI Risk Management Framework, released in January 2023 for voluntary use, frames managing AI risk through four functions: Govern, Map, Measure and Manage. You can use the same idea at a personal scale: set your rules, understand the context, test your results and manage what you find.

• Assuming AI is neutral because it is a machine.
• Using an AI score to rank people without human review or testing.
• Pasting copyrighted or licensed material you may not share.
• Presenting AI output as wholly your own work where disclosure is expected.
• Imitating a living creator's style for commercial work.
• Publishing mass-produced, unreviewed content.
• Treating human review as a formality.
• Not naming who is accountable.

• Next: Chatbots, Workflows, Automation and Agents: Choose the Simplest System begins Module 8, on moving from conversations to systems.
• For the research side of citing and attribution, see An AI-Assisted Research Workflow: Primary Sources and Citation Checking.
• To bring your rules together, see Design Your Personal AI Operating System.