A faster assessment may benefit an organisation while putting the person being assessed at a disadvantage. Suppose an AI-generated summary of a job application omits relevant experience. You then want to know who would notice, whether the applicant can have the error corrected and whether someone returns to the original material before a decision is made.
These questions give an ethics discussion something concrete to address. You have to decide whose interests count, where the boundaries sit and what someone can do if the result is wrong. Working agreements help put those choices into practice. They do not resolve the underlying choices for us.
1. Define the purpose, the people affected and the limits
Describe what the application is intended to do and who may be affected. A shorter process is one interest; fair and careful treatment is another. Decide which judgements an employee must make and how the system may support them.
In the recruitment example, an organisation might require the original application to remain available and an employee to check relevant information before deciding. That is an example of a working rule, not a complete assessment of whether the application is appropriate.
2. Make results open to scrutiny and challenge
A plausible explanation is not enough. A reviewer needs to establish what information was used and where significant uncertainty remains. People also need a way to report an error or adverse effect, with someone responsible for considering a request for reassessment.
Explain the role of AI where people need that information to assess an outcome fairly or where disclosure is required. The appropriate detail depends on the application and the people involved.
3. Use and retain only the data needed
Establish what data the application requires and the terms under which a supplier processes it. Apply the same scrutiny to the records retained for oversight. Comprehensive logs can contain sensitive information too. Decide what to record, who can access it and when it will be deleted.
Assess personal data, confidential information and rights in the material used. A supplier contract does not remove the need for those decisions.
4. Give someone authority to intervene
Human oversight requires more than a name on an approval. The responsible person must be able to challenge a result, request further evidence and interrupt use. Make the route for raising problems accessible both to staff and to people affected by the application.
Agree on the warning signs that should trigger a pause and who can make that decision.
5. Review consequences you did not anticipate
Once the application is in use, look beyond time savings and error rates. Examine complaints, recurring exceptions and groups receiving different treatment from what was intended. Involve affected people where appropriate. Revise the application or its rules if the original judgement no longer holds.
There are resources to help organise this work. NIST offers a voluntary risk framework with a profile for generative AI. ISO/IEC 42001 sets requirements for an AI management system. The legal obligations that apply require a separate assessment of the specific application and the organisation’s role.
How this connects to Common Sense Management
Common Sense Management is about clarity, responsibility and simple agreements that work in practice. Thick manuals are not part of it. The same applies to AI. Keep it small:
- clear goals
- a clear owner
- clear agreements
- a regular rhythm
Seen this way, ethical AI is simply part of good management. It does not need to be an extra layer on top of the work.
How I would assess an ethical AI policy
A useful agreement explains what happens when interests conflict or someone challenges a result. That is how I would assess an ethical AI policy: can the people involved use it when the situation becomes difficult?
Want to work through these questions for an application in your organisation? Let’s talk.