What does an ethical AI problem look like in an ordinary job?
An ethical AI problem at work usually looks like a small shortcut: a ranking nobody checks, a letter nobody signs or a spreadsheet pasted into a tool that was never approved.
Picture a fictional 35-person lettings agency, Calder Lettings. Its property manager, Nadia, has two jobs on Monday morning. First, 60 maintenance requests arrived over the weekend, and she wants an AI assistant to rank them by urgency. Second, 14 tenants are behind on rent, and she wants the assistant to draft reminder letters. Her messy input looks like this: 'boiler not working, 2 kids, flat 4B'; 'tap drip again!!'; 'mould in bedroom, tenant says asthma'; 'arrears list attached: names, amounts, phone numbers, notes like single mum, lost job'. Each of those lines contains an ethical decision she may not notice she is making.
What are the principles of ethical AI in plain words?
The principles of ethical AI come down to five questions: is it fair, is it transparent, does it protect privacy, who is accountable and could it cause harm? International frameworks such as the UNESCO Recommendation on the Ethics of Artificial Intelligence and the OECD AI Principles use longer lists, but these five cover the decisions most office workers face.
Fairness: would this treat similar people differently?
Check whether the output ranks, scores or words things differently for reasons that should not matter, such as a name, an address or how well someone writes English.
Transparency: would the people affected know AI was involved?
Ask whether a customer, colleague or tenant would reasonably expect to be told, and whether you could explain how the result was produced.
Privacy: did you share more personal data than the task needs?
Remove names, contact details and sensitive notes unless the task truly needs them, and use only tools your employer has approved.
Accountability: whose name is on the decision?
Every AI-assisted decision needs a person who reviewed it and can answer for it. 'The AI suggested it' is not an owner.
Harm: who loses if this is wrong?
Picture the worst realistic error and who pays for it. The higher the cost to a person, the more checking the output needs.
What does a weak and an improved ethical AI decision look like?
Weak approach: Nadia pastes the full arrears spreadsheet, including the notes column, and asks: 'Write firm letters to these tenants telling them to pay or face eviction.' She also asks the assistant to rank the maintenance list 'by importance'. The result ranks the leaking tap above the mould because that tenant wrote a long, angry message, and the letters repeat the word eviction to a tenant whose notes mention a recent job loss. Personal notes have gone into a tool, nobody has signed the letters and the ranking rewards loud writing rather than risk.
Improved approach: Nadia removes names, phone numbers and the notes column, then asks: 'Rank these maintenance requests by risk to health and safety first, then property damage, then comfort. Explain each ranking in one line. Do not consider tone or length of the message.' The ranking puts the boiler with children and the mould with a stated health condition at the top, each with a reason she can check. For the letters she asks for one polite template that states the amount owed, offers a call to agree a plan and names where to get free debt advice. She adds names herself, reads each letter and signs it. Every one of the five questions now has an answer.
How do you check whether an AI-assisted decision is ethical?
Check an AI-assisted decision by writing one line against each of the five questions before you act. If you cannot fill a line, that is the part to fix or escalate.
| Question | Nadia's answer |
|---|---|
| Fair? | Ranking uses stated risk criteria, not message tone |
| Transparent? | Letters are reviewed and signed by a named person |
| Private? | Names, phone numbers and notes removed before use |
| Accountable? | Nadia owns the ranking and each letter |
| Harm? | Worst case is a delayed repair; top risks checked by hand |
What are the common mistakes with ethical AI at work?
The most common mistake is treating ethics as a policy document rather than a habit applied to each output. Others: assuming a tool is neutral because it is a machine, pasting sensitive notes because they were in the same spreadsheet, letting an AI ranking stand without asking what it was based on, and hiding AI involvement in communication people would expect a human to write. When a decision affects someone's home, job, money or health, escalate it to a person with authority instead of deciding alone.
Where does ethical AI fit in the Tosheo.ai catalogue?
Ethical AI sits in the Responsible AI and Policy category, in the AI literacy course, alongside lessons on responsible AI controls and human review. On Tosheo.ai, the hands-on task gives you a short case from your own role, you make the decision with an AI tool, and the rubric feedback points out each question you skipped.